DailyArxiv - Ed1sonChen

LLM agents that orchestrate frozen vision-language-action (VLA) policies improve across episodes through text memory, which records what the agent did but not how the task is done. A demonstration video shows it, but fits poorly into an agent's context. The full video slows every turn, fixed keyframes lose the contact detail that decides whether a grasp holds, and what the agent needs shifts from the task's structure while planning to the frames around each contact. We introduce Recursive Video In-Context Learning (RV-ICL), a training-free method that turns a demonstration into a hierarchy the agent navigates rather than a prompt it receives. The hierarchy is built from the sub-events of the demonstration, such as grasps and releases. Its levels grow finer, from keyframes of the whole task to phases, moments and short clips, and are exposed through read-only tools. The agent reads the coarse levels before planning. During execution it re-enters the hierarchy whenever a step needs more detail and loads only the clip of its current sub-goal. One demonstration per task is enough. Built on RPent, RV-ICL raises success from 92.6% to 96.5% on LIBERO-PRO and from 86.7% to 95.8% on LIBERO-Plus.

Action-conditioned robot world models must respond precisely to robot trajectories while preserving realistic visual dynamics, yet learning both from heterogeneous robot videos remains challenging. Simulation offers structured motion supervision, but appearance differences hinder direct transfer, and inaccurate simulation predictions can misguide real-video generation. We present SimForcing, a simulation-guided framework that uses simulation both as a source of transferable motion knowledge and as a controllable reference for prediction. First, we transfer motion knowledge from a simulation teacher through latent-motion distillation, aligning temporal changes in latent space to internalize motion priors while mitigating the influence of appearance differences. Second, we introduce multi-block simulation conditioning with condition dropout to exploit predicted simulation trajectories without relying excessively on their accuracy. Our simulation-conditioning classifier-free guidance scheme unifies these two ideas by balancing predictions based on internalized motion knowledge with those additionally guided by simulation latents. The jointly trained student generates both simulation conditions and real-domain videos, requiring no additional world model at inference. On Bridge, SimForcing achieves the best PSNR, SSIM, LPIPS, and FVD among the compared methods without external embodied pretraining. Evaluation on InternData-A1 further supports its applicability across robot datasets. Moreover, using our trained world model to initialize a vision-language-action model improves LIBERO success, suggesting its utility for downstream policy learning. \url{https://github.com/Wang-Xiaodong1899/SimForcing}

Code: https://github.com/Wang-Xiaodong1899/SimForcing

Closed-loop evaluation of end-to-end driving requires continuous rollouts that reveal how earlier decisions affect subsequent driving. However, existing benchmarks evaluate only short segments and fail to capture later consequences. Ambiguous directional commands also obscure the intended navigation objective. We introduce Odyssey, a closed-loop benchmark for long-horizon driving comprising 100 scenarios, each reconstructed from a 100-second nuPlan driving log to preserve the context of navigation maneuvers and traffic interactions. To provide a consistent navigation objective, Odyssey replaces directional commands with explicit standard-definition (SD) map routes that specify which roads to follow, while sensor-based planning determines local driving actions. Throughout these rollouts, diffusion-based refinement of 3DGS-rendered images reduces rendering artifacts along the ego trajectory. To assess how effectively planners follow these routes and prepare for upcoming maneuvers, we introduce SD Route Compliance and Pre-Lane Change Score. These assessments are complemented by RouteDS, which extends the Driving Score with penalties for SD-route deviations and failed lane preparation. We adapt state-of-the-art planners, including vision-language-action (VLA) models, and evaluate their navigation performance using these metrics. Odyssey highlights open questions in route representation and integration for E2E driving. Benchmark code and adapted baselines will be released publicly.

Dense affordance supervision is an appealing auxiliary signal for vision-language-action (VLA) policies, yet naively co-training an affordance head can severely damage instruction following. We present a controlled study of how to wire such a head into a modern VLA on the LIBERO benchmark. Our recipe reads the backbone through a stop-gradient and re-injects an intermediate head feature into the action expert via a learned bridge. The stop-gradient is a precondition: letting affordance gradients reach the backbone drops the policy below the headless base (85.5% vs. 93.1%). With the backbone protected, a same-budget 2*2 ablation over injection topology (concatenation vs. residual) and bridge initialization (zero vs. random) shows initialization is the dominant lever. The best wiring, an actively initialized residual bridge, reaches 96.2%, matching the far more elaborate three-expert AffordanceVLA (95.8%) with under 1% extra parameters. Two probes explain the mechanism: ground-truth affordances fed as an input hurt, and inference-time zeroing shows a lazy bridge acts only as a training-time regularizer while an active bridge becomes load-bearing.

8 pages, 4 figures, 2 tables

Adapting vision-language-action (VLA) models to deployment-time distribution shifts is important for reliable robotic operation, but conventional first-order adaptation can exceed the memory budget of inference-oriented deployment platforms. Zeroth-order (ZO) optimization offers a forward-only alternative with inference-level memory, but accurate gradient estimation requires many perturbation queries, making naive ZO prohibitively slow for large VLA models. We present VLA-ZO, a framework for fast ZO adaptation that exploits the structure of VLA computation. By confining adaptation to the action side, VLA-ZO keeps the expensive vision-language prefix frozen and reuses its conditioning states across perturbation queries and optimizer steps, while schedule-aware prefetching hides state-transfer overhead. On LIBERO camera-viewpoint shifts, VLA-ZO reduces end-to-end adaptation time by 25.59$\times$ at $q=16$ and 32.54$\times$ at $q=64$ relative to baseline ZO, while improving average task success from 48.27% without adaptation to 58.17% and 63.58%, respectively. These results show that making ZO faster can make larger query budgets practical, providing a promising path toward resource-efficient VLA adaptation on deployment platforms.

Instruction following is central to language-conditioned robot policies: language should determine what to do when the same scene permits multiple valid actions. Yet successful execution alone cannot establish whether a policy follows the instruction or infers the task from the scene. We study this ambiguity through scene-preserving instruction interventions, using valid target substitutions, arbitrary nouns, and unrelated sentences while holding the scene fixed. We evaluate vision-language-action (VLA) policies and world-action models (WAMs) in simulation and in real-world experiments. Our analysis addresses three questions: (a) Does task success imply instruction following? When instructions request a different visible object, all evaluated policies predominantly approach and pick up the incorrect original target associated with the scene. (b) Is this failure caused by language insensitivity? Instruction perturbations affect task performance. A layerwise action lens shows intermediate action predictions respond to these perturbations. Linear probes accurately recover instructed targets, indicating modified instructions are encoded despite rarely determining target selection. (c) Why does encoded language fail to control action? Attention analysis indicates weak instruction-token contributions to action generation. Target-token attention can remain focused on the original object, revealing a mismatch between target encoding and visual grounding. UMAP and shared non-negative matrix factorization show target information remains accessible within representations increasingly organized by scene identity. Our findings expose a grounding gap concealed by nominal success and provide a diagnostic framework. They further establish a concrete criterion for progress: policies should reliably follow valid changes in user intent, even when they conflict with scene-favored behavior.

Generalization in multi-arm collaboration can be studied as composing familiar atomic skills in new ways across arms. However, existing evaluations offer limited insight into which training and architectural choices support this ability under different coordination requirements. We introduce \textbf{ACG-Bench}, a benchmark for \emph{Arm-wise Compositional Generalization} that provides a common testbed for studying skill recomposition in dual-arm policies. It contains 23 task--condition pairs across 8 task families, with 6 in-domain conditions and 17 unseen compositions covering reordering, synchronization, their combination, and cross-task composition. All methods receive the same per-arm atomic prompts, and success requires achieving the task goal while satisfying physical milestones and specified order or timing constraints. Using $π_{0.5}$ as a common vision-language-action backbone, we compare representative data-augmentation and architectural strategies with shared source data and a common evaluation protocol. Our architectural study examines arm-token grouping, skill-specific LoRA adapters (SkillLoRA), and arm-wise attention (AWA), highlighting the complementarity of skill-conditioned parameters and attention structure. Combining these choices yields \textbf{AE-VLA}, which achieves 21.53% generalization success in simulation, compared with 2.94% for Single $π_{0.5}$, 3.06% for MA-VLA, and 5.53% for two independently controlled $π_{0.5}$ policies. On physical SO101 robots, AE-VLA reaches 39.00% mean success across five unseen conditions, compared with 10.00% for the strongest baseline. These findings provide empirical guidance for designing dual-arm policies that generalize beyond fixed training routines.

Vision-Language-Action (VLA) policies are typically evaluated under the assumption that the robot starts acting only after the user has finished typing or speaking. In real interactions, however, entering an instruction can take several seconds, leaving the policy idle for a substantial fraction of the interaction. Partial instructions may already contain sufficient information to begin acting before the full instruction arrives. We introduce Premover, a parameter-lightweight module that reduces interaction latency by overlapping instruction delivery with robot execution while keeping the VLA backbone frozen. Premover uses a learned vision-language focus map to identify where the current partial instruction refers in the visual scene, and an action readiness gate to determine when execution can begin. On the SO-101 robot with speech and online transcription, Premover reduces mean wall-clock time by 17.9%, from 60.5s to 49.7s, while maintaining task success rates. In simulation on LIBERO with pi0.5 at the speaking rate, Premover reduces mean wall-clock time by 8.3% while maintaining comparable success to waiting for the complete instruction before execution.

Vision-Language-Action (VLA) models acquire broad embodied capabilities through large-scale pretraining, yet their generalization remains far more fragile than that of LLMs and VLMs. The prevailing remedy, post-training via supervised fine-tuning or reinforcement learning, improves task-specific performance but narrows the generalist capability that makes pretraining valuable. We identify a key bottleneck: VLA failures stem not only from action generation but also from action evaluation. A diagnostic pass@k study confirms that frozen VLAs already contain competent behaviors in their output distribution, with overall success rates rising from 33% at pass@1 to 92% at pass@32. Inspired by this, we propose SVA (Search, Value, and Act), a simple framework that equips frozen VLA policies with long-term consequence awareness. SVA first uses Monte-Carlo tree search in simulation to sample and explore the VLA's output distribution under a finite search budget, collecting diverse trajectories annotated with empirical returns; this knowledge is then distilled into a lightweight Q-value model that predicts the expected consequence of candidate actions; at deployment, the frozen VLA proposes multiple candidates and the evaluator selects the one with the highest uncertainty-regularized Q-value, requiring no simulator access. By decoupling action proposal from consequence evaluation, SVA preserves the generalization capacity of the VLA backbone while improving task success. Across embodied benchmarks, SVA consistently improves generalization on unseen tasks and exhibits strong test-time scaling; across real-robot manipulation tasks, it improves average success from 42.8% to 55.6% (+12.8 points). Strikingly, SVA enables a 9B VLA to outperform a 27B VLA by 7 points at 27% lower inference latency, suggesting that scaling test-time evaluation is more cost-effective than scaling model size.

Vision-language-action (VLA) models currently demonstrate strong performance in a wide range of real-world robotics tasks. However, they often still lack the generalization ability to handle large visual out-of-distribution shifts. Post-training of VLAs with reinforcement learning (RL) has been shown to benefit robustness, but significant room for improvement remains. In this work, we systematically study the effect of image augmentation on VLA post-training. We find that it is crucial to augment only the critic module during RL updates, while leaving the actor's input clean during both rollouts and updates. For $π_{0.5}$ and GR00T N1.5 this raises out-of-distribution success on LIBERO-Plus by $7.8$ and $10.0$ points respectively, while augmenting the actor collapses training entirely. We investigate a range of augmentation types and strengths, and provide practical recommendations for improving generalization in VLA post-training.

Accepted at the NeurIPS 2026 workshop RoboPAD. Code and videos at https://bramgrooten.nl/vla-augm/

Benchmarks expose vision-language-action (VLA) policies to few canonical instructions, while exhaustive deployment testing is impossible. We introduce Object-Grounded Attention Monitoring (OGAM), connecting systematic testing to runtime assurance: testing reveals attention divergence between successful and failed executions, and OGAM uses this signal to stop failures beyond the finite suite. We generate scene-grounded instructions through pairwise combinations of action templates and objects, and separately test meaning-preserving paraphrases. All 87 out-of-benchmark cases reveal problematic behavior across OpenVLA, OpenVLA-OFT, UniVLA, and $π_{0.5}$: none completes any of the 24 feasible instructions, while infeasible or hazardous requests also trigger behavior substitution. At each action query, we project gradient-weighted visual attention through object masks and group it by instruction role for comparison across tasks and policies. Dynamic time warping aligns this course with a successful reference despite speed differences; conformal calibration on successful episodes sets the early-stopping threshold for sustained deviations, with a nominal false-stop target of $α=0.05$. Across four policies, OGAM stops 87-100% of failed episodes at median times of 5-12s within a 20s budget, with observed false-stop rates of 3-5%, without failure-labeled training. Finite testing thus identifies attention patterns that support online intervention before failure fully unfolds.

Vision-language-action models (VLAs) act on instructions without being able to refuse, so screening harmful requests falls to monitors. We test whether these monitors catch ordinary robot tasks requested for harmful reasons, holding the task fixed while varying only how explicitly the intent is stated. $π_{0.5}$ completes the task at every level of explicitness, as often as for harmless controls. Text guards flag nearly every blunt request but few implied ones: up to 95% of implied-harm runs end with the task done and no flag raised, and up to 90% even after recalibrating on robot instructions. Monitoring the model's activations does not close this gap. Linear probes separate harmful from harmless instructions almost perfectly in the base language and vision-language models, but this separation weakens after robot training in two model families, most for implied harm.

Under review at the SPAIS 2026 workshop (CoRL 2026)

A common approach to world-model simulation for vision-language-action (VLA) systems is to predict future RGB observations and then re-encode them into policy inputs, introducing an indirect interface between simulation and downstream policy execution. We instead investigate whether world dynamics can be modeled in a compact, policy-oriented state derived from VLM visual tokens. A key challenge is that raw VLM visual tokens are high-dimensional, making efficient and accurate autoregressive dynamics modeling challenging. To address this, we introduce Token-World, an action-conditioned world model that compresses VLM features into a compact token state, learns future dynamics in this reduced space, and maps predicted states back to the original policy-facing representation for downstream use. Across manipulation benchmarks, Token-World improves open-loop feature fidelity and policy-action consistency over recent world-model simulators, with slower degradation over long rollout horizons. In closed-loop evaluation, its simulated policy performance correlates more strongly with reference policy performance than Ctrl-World ($r=0.794$ vs.\ $0.583$), while requiring lower simulation latency. Ablations further show that compact-representation design and dimensionality substantially affect future-state prediction. Code will be available at https://chuyaofu.github.io/Token-World/.

Submitted to IEEE International Conference on Robotics and Automation (ICRA) 2027

Vision-Language-Action (VLA) models serve as unified policies for robotic manipulation, yet their expensive inference forces robots to pause between policy calls, resulting in stop-and-go execution that interrupts smooth motion and prolongs task completion. Extending the action chunk reduces policy calls and hence these pauses, but predicting farther into the future makes long-chunk execution unreliable. To understand where this unreliability arises, we analyze action errors within long chunks and find that they concentrate around transitions between manipulation subskills, growing sharply with chunk length. This suggests the importance of transition timing, i.e., when to switch subskills within a chunk. Motivated by this observation, we introduce RACE (Reliable Action-Chunk Extension), a framework that predicts the transition timing from an auxiliary one-step denoising pass and conditions action generation on it. By learning and conditioning on transition timing, RACE reduces errors at subskill transitions and enables reliable execution of longer action chunks. Across simulation benchmarks, RACE outperforms fine-tuning at the same chunk length; with 2x longer chunks, it surpasses recent state-of-the-art and efficient VLAs in success rate, and with 4x longer chunks, it remains competitive. On a real robot, RACE uses 4x longer chunks, which reduces the idle time caused by stop-and-go execution by about 5x, while achieving a higher success rate than fine-tuning with the same chunk length. Code and a real-robot demo are available at https://github.com/Seonghoon-Yu/RACE-VLA

Vision-language-action (VLA) models have shown strong potential as generalist robot policies, but adapting them to unseen tasks often requires costly parameter updates. Recent work such as RICL introduces in-context adaptability by retrieving expert demonstrations based on the current VLA observation and providing them as additional context at test time. The effectiveness of this adaptation therefore depends critically on the retrieval mechanism. In this work, we systematically study how different retrieval methods affect both retrieval quality and task performance within the RICL framework. Specifically, we compare four different methods: image-based retrieval, retrieval augmented with VLA's state, retrieval using features from the VLA backbone, and random retrieval. Our experiments yield three main findings. First, no retrieval method consistently dominates the others in task success, while surprisingly, random retrieval achieves a non-trivial success rate. Second, standard retrieval-quality diagnostics do not reliably reflect downstream VLA performance. Third, demonstrations from different but related tasks can provide useful transferable information. Together, these results provide an initial step toward understanding how retrieval mechanisms shape the in-context learning capability of VLA models and their downstream task performance, while highlighting the need for more careful design and evaluation of retrieval mechanisms for reliable test-time adaptation.

NeurIPS 2026 Workshop PTA

Most vision-language-action (VLA) models rely on current observations and lose task-relevant evidence once it leaves view, limiting performance on long-horizon, memory-dependent tasks. Existing efforts incorporate compressed historical features or sparse visual keyframes. However, isolated snapshots can leave the policy uncertain about what changed during past interactions and which action should follow. To overcome this limitation, we propose EvoMem-VLA, which constructs state-evolution memory by explicitly encoding and retaining observed changes between historical states. These change representations preserve evidence of interaction outcomes, allowing the policy to track task progress beyond isolated snapshots. Specifically, we introduce conditional delta tokenization to encode ordered frame pairs into directional, source-conditioned delta tokens, each associated with its corresponding state evidence. A shared VLM backbone supports task-adaptive routing: normal long-horizon tasks follow a direct action route, whereas multi-stage tasks use a subtask route that generates an executable subtask as an additional input for action generation. With a single jointly trained policy for each simulation benchmark, EvoMem-VLA achieves success rates of 80.7% on RMBench, 82.0% on RoboMME and 83.8% across four real-world tasks spanning two robot embodiments. These results represent substantial improvements over the previous state of the art in all three evaluation settings.

Vision-language-action (VLA) models achieve favorable task performance, yet runtime errors in closed-loop execution evolve with alternating updates of actions and observations. Existing VLA quantization methods mainly trade off inference speed and model performance while rarely investigating how quantization alters runtime errors. By comparing error distributions between full-precision and quantized models, we identify distinct evolutionary patterns for the two types of errors: quantization enlarges the variance of translation error distributions and increases their dispersion, whereas the distribution center of rotation errors gradually shifts across execution steps, demonstrating cumulative drift. Motivated by this observation, we rethink the optimal quantization strategy for VLA models and propose \textit{DyQ-VLA}, a runtime-error-aware quantization framework, which dynamically selects activation precision according to execution steps and tracks as well as compensates rotation errors via accumulated quantization residuals. Corresponding operators and runtime adaptation strategies are devised within the framework to enable dynamic-precision execution. Experiments show that \textit{DyQ-VLA} achieves a 1.88 to 1.93 inference speedups while maintaining comparable or higher average task success rates. Moreover, it reduces mean execution steps by 17.7% to 19.8%. Our code is here: https://anonymous.4open.science/r/DyQ-VLA-7F51/.

Visual token pruning is an effective way to accelerate vision-language models and is especially useful for vision-language-action (VLA) inference, where many visual tokens must be processed before predicting robot actions. Existing pruning methods usually estimate which tokens can be pruned based on attention scores or feature diversity, retaining tokens that are either highly attended or visually different from others. However, most of them use fixed pruning schedules, such as pruning once at a preset layer or pruning at uniformly spaced layers. Such schedules can be risky for VLA models, because the model may not know which visual regions matter for the action in early layers. Tokens that look unimportant at first may become useful after the model combines visual observations with the language instruction. In this work, we propose SAPrune, a training-free visual token pruning framework for efficient VLA inference. Instead of pruning at fixed layers, SAPrune uses a small calibration set to observe how action-to-visual attention changes across layers, and chooses pruning layers only after the attention pattern becomes more reliable. At each selected layer, SAPrune applies a dual-path pruning rule: one path protects strongly attended visual tokens from pruning, while the other prevents useful surrounding context from being discarded. Experiments on LIBERO, SIMPLER, and real-world robotic tasks show that SAPrune prunes 87.5% of visual tokens and achieves up to 1.718x inference speedup while maintaining competitive task success rates.

Accepted at NeurIPS 2026

Most vision-language-action models represent future motion as fixed-rate action chunks, tying temporal resolution and prediction horizon to a fixed output budget. This pointwise representation wastes capacity on highly correlated neighboring actions, leaves temporal continuity and smoothness to be learned implicitly, and forces a tradeoff between long-horizon coverage and the local precision required for contact-rich manipulation. To address these limitations, we introduce Vela, a vision-language-action foundation model that represents future robot behavior as continuous trajectories. Vela combines a compact spline-based action representation with motion-dependent temporal support and a shared action interface for heterogeneous embodiments, allowing a fixed output budget to adapt its temporal resolution across motions. We pretrain Vela on large-scale multi-embodiment robot data and evaluate it on LIBERO-X, EBench, and two real-world long-horizon tasks, egg-cake cooking and potato shredding, obtaining promising results across simulation and physical manipulation. These results highlight the potential of continuous action representations as a foundation for future embodied foundation models. Project page and more results: https://clementine24.github.io/Vela/ .

A safe action is not necessarily a viable one. Under a frozen vision-language-action (VLA) policy, an action can be likely and locally admissible yet leave no policy-supported route to safe task completion. We call this the feasibility-likelihood gap: likelihood ranks the current action, whereas feasibility depends on the futures that remain after it. We derive the exact next-block marginal of the history-conditioned policy-environment trajectory law restricted to safe task completion. The derivation exposes a candidate-dependent feasible-future mass with two roles: its support records whether safe completion remains possible under the frozen continuation process, and its magnitude measures how much weighted safe-completion mass is preserved. Exact evaluation is impractical online, so we develop a selective finite-candidate approximation, derive conditions for recovering the best retained viable candidate, and instantiate it as an alarm-triggered, training-free reranker. On Safety-CHORES, VICS-G lowers mean cumulative safety cost by 1.9%-57.5% across six settings while remaining within 2.5 percentage points of policy sampling in success and 0.82 steps in mean episode length. The resulting decoder is tied to an exact policy-relative safe-completion target, yet requires neither retraining of the base policy nor online trajectory rollouts.

Vision-language-action (VLA) models translate multimodal observations into low-level robot actions. During robot operation, each control period sets an inference deadline, and overruns leave the robot acting on stale observations, reducing task success. Meeting this deadline motivates on-device or nearby edge execution, where a single robot requires batch-1 inference outside the design point of LLM serving systems. Although VLA architectures combine familiar vision-language, autoregressive, and diffusion-style components, their runtime behavior in this batch-1 control setting remains uncharacterized. We characterize four representative VLA models on an edge GPU server and two onboard SoCs, using single-inference profiling and 43,200 closed-loop episodes. Action tensor dimensionality determines whether a stage is memory- or compute-bound, platform balance can shift that bottleneck, and GPU frequency scaling yields a platform-dependent energy-latency sweet spot. In closed-loop operation, overlapping inference with action execution creates an accuracy-speed-energy tradeoff, and no configuration is Pareto-dominant across deployment SLOs. These results guide joint design of VLA model architectures, hardware, and runtime policies.

To appear in the Proceedings of the 32nd ACM International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS '27). 17 pages, 14 figures

Vision-language-action (VLA) models encode semantic information from vision-language pretraining, but manipulation also requires precise spatial reasoning. We present GeoBridge-VLA, a two-stage method for learning geometric features from a pretrained VLA's frozen visual encoder and using them for action prediction. Stage I trains a feature bridge and geometry decoder with depth supervision. Stage II freezes these modules and trains a gated residual interface together with the action-side projections and action expert. The residual augments the existing visual tokens without adding a second image encoder or increasing the token count. Deployment requires RGB, robot state, and language, but no depth observations. Under matched evaluation conditions, GeoBridge-VLA achieves 70.9% success on LIBERO, compared with 60.0% for SmolVLA. Disabling the residual in the same trained checkpoint reduces success from 70.90% to 69.85%, with mixed effects across suites. On a physical ROBOTIS OMY robot, GeoBridge-VLA succeeds in 148 of 200 trials (74.0%) across four tasks, compared with 108 of 200 (54.0%) for SmolVLA.

Dual-system Vision-Language-Action (VLA) models improve real-time robotic control by pairing a slow, reasoning-capable generalist with a fast specialist action expert. However, existing methods invoke the generalist at a fixed frequency, ignoring the fact that decision-making complexity varies throughout a rollout. This static strategy wastes computation in easy phases and can delay renewed reasoning when the scene changes unexpectedly. We propose TUD (Triggering generalist reasoning via predictive Uncertainty for Dual-system VLA), an adaptive inference framework that selectively skips unnecessary generalist calls. TUD measures the cross-step dispersion of action re-predictions at the upcoming chunk slot under the cached generalist context, as a predictive uncertainty signal. This signal captures how much the future action plan shifts as new observations arrive and is computed from forwards the architecture already runs, requiring neither manual phase labels nor an auxiliary uncertainty model. On VLA-Arena, it achieves a higher success rate at matched call budgets than alternative uncertainty baselines while maintaining low wall-clock overhead, and more consistently separates successful from failed rollouts. Also, TUD finds a more favorable cost-success trade-off than non-adaptive baselines, tracing an entire operating curve as a single threshold is varied, and substantially reduces VLM calls at matched success rate. The same trade-off appears in our real-robot experiments, where TUD cuts generalist calls by 75% relative to the strongest fixed-interval baseline while achieving an even higher success rate. Our results suggest that predictive uncertainty provides a practical criterion for adaptive reasoning in efficient VLA control.

Accepted by NeurIPS 2026

Scaling robot data and model capacity has improved Vision-Language-Action (VLA) policies, but further progress is constrained by the high cost of robotic data. Verifier-guided test-time scaling offers an efficient alternative by sampling multiple action candidates and selecting the one most likely to lead to task success at inference time. Existing classification-based verifiers learn from trajectory-level outcomes but treat all visited states equally, even though their value for candidate discrimination can vary across a trajectory. At many states, plausible actions are similar and provide limited discrimination signal, while only a sparse set of decision-critical states admits meaningfully different actions that can substantially affect downstream outcomes. To address this, we propose DiVeR, which estimates decision criticality from the dispersion of sampled action representations. DiVeR then uses this signal to reweight verifier learning toward states where action selection is most consequential, without requiring step-level annotations or additional environment interaction. Across LIBERO, RoboCasa, and real-world experiments on a Franka Research 3 robot, DiVeR consistently improves task success through more effective verifier-guided action selection, while adding negligible verifier inference overhead.

The linear representation hypothesis (LRH) has become a standard lens for measuring and intervening on semantic information through the internal representations of large language models (LLMs). A growing body of work has begun extending this perspective to vision-language-action (VLA) models, but the dynamical nature of embodied interaction introduces an additional challenge. Unlike semantic attributes commonly studied in LLMs, such as gender or language, a physical quantity of interest (QoI) in a VLA evolves jointly with the system dynamics: the representation influences the actions selected by the policy, which alter the physical state and, in turn, the next representation. In this paper, we develop a theoretical, signature-based formulation of the LRH for VLA that unifies representations and policies. On the representation side, we establish the existence of representations from which the future evolution of a QoI under a candidate action trajectory can be recovered via linear probing. On the policy side, we introduce a signature generalized linear model for stochastic action chunks. This structure yields a monotonic change in the expected future QoI along linear paths in natural parameter space, enabling linear steering. We first validate this structure in a controlled oracle setting, then examine whether the same probing and steering mechanisms emerge in a pretrained VLA.

Three-dimensional perception is critical for robotic manipulation, particularly for high-precision tasks, as recovering metric depth and precise 3D object positions from monocular RGB observations is inherently ill-posed. However, many Vision-Language-Action (VLA) models rely solely on RGB observations for perception. Leveraging recent advances in foundation models for stereo matching, we introduce ExStereo, a stereo module that augments pre-trained 2D VLAs with 3D perception. ExStereo reconstructs scene geometry from stereo image pairs and renders multi-view observations as an explicit stereo representation for stereo feature extraction. The action tokens from the action expert selectively attend to the resulting stereo tokens through our proposed action-stereo cross-attention mechanism, enabling the policy to generate robot actions conditioned on 3D scene information. To learn robust 3D representations, we introduce a mid-training stage before task-specific post-training, using a self-supervised learning objective on large-scale stereo data. We validate our approach by fine-tuning two publicly available VLAs, $π_{0.5}$ and SmolVLA, and evaluate them in simulation and on a real-world bimanual PiPER platform. Across both settings, VLAs fine-tuned with ExStereo consistently outperform baselines, demonstrating the effectiveness of stereo perception for robotic manipulation. Our project website is at: https://exstereo-vla.github.io/ExStereo/.

Learning-based models (e.g., visuomotor and Vision-Language-Action (VLA)) are increasingly explored for industrial robotic manipulation, where model predictions are directly translated into physical actions. This tight coupling between model behavior and physical execution makes hidden security vulnerabilities particularly consequential. While backdoor attacks have been widely studied in conventional AI models, their effects on deployed learning-based robotic arm manipulation systems remain less understood: a backdoored robot can behave normally during benign operation while inducing attacker-specified behaviors only when specific triggers are present, posing potentially serious risks in physical environments. In this work, we present a preliminary empirical security study of backdoor attacks and defenses in learning-based robotic manipulation on two real commercial industrial robotic arms (FANUC and xArm). We investigate whether a backdoor can reliably induce semantically incorrect manipulation behaviors while remaining stealthy under nominal task execution. We then develop an online defense pipeline that detects and neutralizes triggers at runtime, and compare its effectiveness against an offline fine-tuning defense. Beyond defense effectiveness, we further evaluate the computational latency and execution overhead introduced by the defense pipeline to assess its suitability for high-throughput industrial operation.

Accepted at the IROS 2026 Workshop on Industrial Applications of Robot Learning. Added the required Distribution Statement A for public release and distribution

Vision-language-action (VLA) and world-action models (WAMs) often degrade under out-of-distribution task variations despite retaining partial task capability. To recover such capability, we propose RoboIRS, an inference-time internal representation steering method that uses successful and failed rollouts to train linear classifiers, select outcome-relevant intervention locations, and derive task-specific steering directions without updating policy parameters. On 15 simulation tasks with a frozen $π0.5$ policy, RoboIRS improves the average success rate from 44.4% to 66.2%, outperforming alternative inference-time intervention baselines while adding little inference time. We further validate RoboIRS on real-robot manipulation using the same $π0.5$ policy and demonstrate its applicability to a world-action model Cosmos Policy, where the average success rate improves from 35.4% to 55.4%. These results show that directly steering internal robot-policy representations can improve the performance of robot policies at inference time. Project website is available at https://rollingoat.github.io/roboirs/.

Vision-language-action (VLA) policies commonly learn action chunks through a fixed left-to-right (LTR) factorization, although the same expert trajectory distribution admits many valid chain-rule factorizations. We identify factorization order as an overlooked regularization choice and introduce causally anchored permutation (CAP), which samples action reveal orders with a tunable chronological prefix. Its auxiliary objective trains one shared policy to predict actions from different known subsets of the same expert chunk, while deployment retains deterministic LTR control. We call this conditional-set augmentation: it creates multiple conditional prediction problems from one expert chunk without adding demonstrations. This discourages reliance on the single chronological prefix used by ordinary teacher forcing. Controlled experiments show that CAP consistently outperforms standard LTR training on LIBERO and LIBERO-Plus, with the same advantage appearing in cross-dataset CALVIN evaluation. A diagnostic that measures the expected squared difference between a chunk's joint log likelihood under two reveal orders verifies that CAP training internalizes agreement across reveal orders. These findings position sampled subset-conditioned auxiliary objectives as a general recipe for constructing VLA regularizers, illustrated by an extension to diffusion action generators.

Vision-Language-Action (VLA) models generalize across static manipulation but fail when objects move during task execution. They map the current observation to an action and assume the scene is stationary between observation and execution, so at any non-trivial object speed the resulting latency exceeds the time available to grasp. We close this gap with AHEAD (Anticipatory Horizon Extrapolation with Adaptive Dynamics), a predict-then-act wrapper that augments a frozen VLA with a motion-aware latent world model. A small world model trained on manipulation video forecasts future patch tokens in the VLA's feature space, conditioned on per-token velocity and acceleration from optical flow. A language-and-motion saliency mask concentrates prediction on task-relevant patches, and the model rolls forward for an adaptive horizon, halting when prediction uncertainty crosses a threshold. The frozen action decoder then receives the predicted future tokens in place of the current ones. AHEAD adds 4.9M parameters to a frozen 7B OpenVLA and reaches 79 to 97% success across 20 dynamic simulation scenarios where the strongest baseline reaches 31 to 58%. On a physical UFactory xArm 7, AHEAD succeeds on 29/30 to 30/30 on three conveyor and rolling-ball tasks, 23/30 on paddle interception, and 19/30 on projectile catching where every baseline scores 0/30.

Accepted at the 10th Conference on Robot Learning (CoRL 2026). 31 pages, 7 figures, 19 tables

A vision--language--action (VLA) policy can complete complex tasks while ignoring the evidence that should determine its actions. An object held near the wrist camera can displace the instructed target. Language and action show the same pattern: a familiar noun can trigger the operation it was paired with in training even after the verb changes, and a gripper that closed on nothing may lift anyway. We call these dependencies modality shortcuts: regularities in successful demonstrations make visual, lexical, or motor cues sufficient to predict expert actions without the task evidence needed for the underlying decision. More demonstrations of the same kind can raise task success while leaving these shortcuts intact. We propose Perturbot which makes task-relevant evidence easier to use and shortcuts insufficient on their own: it applies task-preserving wrist-view perturbations, enriches instructions with decision-relevant captions, and adds random and failed trajectory segments relabeled with the behavior they contain. It complements scaling by changing what is scaled, and leaves inference unchanged. Moreover, we propose GroundingFscore, an offline score that diagnoses how severely a policy relies on modality shortcuts. Task success rate shows whether a policy improves, while GroundingFscore reveals whether the policy scales healthily, relying on task evidence rather than shortcuts. Together, Perturbot and GroundingFscore provide a training-and-evaluation framework for disentangling VLA decisions from shortcut priors while preserving responsiveness to task-relevant evidence.

Robots need to anticipate how their actions will change the world, since manipulation success hinges on the resulting contacts and object motions. However, existing Vision-Language-Action (VLA) policies that predict future observations from shared features leave the forecast decoupled from the actions the policy will actually execute, and impose no physical constraints on how the scene may evolve. We introduce ForeAct3D, a framework for policy-grounded future world modeling within VLA policies. Learnable geometric queries decode depth, semantic segmentation, and camera pose from the policy representation into current and future semantic 3D scene states, and the future queries are conditioned on the policy-generated action chunk to ground the forecast in the planned interaction. A physical-consistency closure relates the two states through background staticity and instance-level rigidity, and anchors the wrist-camera pose to end-effector kinematics. These objectives shape the shared representation used for action generation during training, and no future prediction is required at inference. Without robot pretraining, ForeAct3D achieves 98.3% average success on LIBERO and an average task length of 3.73 on CALVIN, outperforming its base policy on every suite. Ablations show that semantic 3D supervision, physical consistency, and action conditioning each improve manipulation performance, and that action conditioning substantially improves future object localization. Real-world experiments on spatial placement, object insertion, and sequential manipulation further raise average success from 6.7% to 37.8% over the base policy. The project page and code are available at https://github.com/anthonytao80-crypto/ForeAct3D.

Repetitive tasks are common in real-world manipulation, yet current vision-language-action (VLA) policies remain unreliable for repeated actions. We present LoopVLA, a cross-subtask event-memory interface for looped task execution. Its central idea is to compress a just-completed subtask into memory for the next subtask's action expert, making completed execution available for subsequent repetition and stopping decisions. Learnable event latents query the closed segment's tokens, while action features query history and joint history--event context; the two action-side readouts are residually fused and injected directly into the action head. On a memory benchmark, LoopVLA achieves $80.22%$ success on the repetition task group with online subgoal prediction, bringing evaluation closer to realistic deployment while improving overall success over strong memory baselines.

Vision-language-action (VLA) models face a timing gap between low-rate inference and high-rate robot execution. We characterize this gap through end-to-end latency measurements of model inference and the robot execution chain. Repeated Flow Matching denoising contributes substantially to inference cost, while robot-side delays mainly arise from perception acquisition, communication scheduling, and physical response. Analysis of the velocity field shows relatively stable magnitude and direction in early integration, followed by stronger directional correction near the terminal steps. Based on this stage heterogeneity, we propose two-stage non-uniform denoising, reducing the number of steps from 10 to 2 and model-inference time from 61.557 ms to 21.956 ms. We also develop a distributed real-time VLA framework with independent inference, action-publication, and robot-control rates, modular observation acquisition, and action-provenance logging. Using π0.5 as the baseline, we evaluate six real-time execution methods on a long-horizon physical garment-folding task. Legato performs best overall among training-based methods, while Temporal Smoothing leads among training-free methods; both perform strongly in task success, completion time, action continuity, and acceleration smoothness. Combining two-step denoising with representative execution methods substantially reduces inference cost with a small reduction in task performance. These results motivate joint optimization of model-inference efficiency and robot-system timing.

31 pages, 21 figures (including 10 supplementary figures), and 10 tables (including 3 supplementary tables). Project page: https://embodied.magiclab.top/works/inference/index.html. Code: https://github.com/MagiclabRobotics/Inference

Vision-language-action policies may predict a transferable manipulation strategy yet fail to realize it reliably on the encountered object: objects compatible with the same grasp differ in geometry and compliance, and visual feedback degrades under closure occlusion. AgenticTactileVLA is presented as an execution-time supervisor that shifts part of object-specific adaptation from prediction to physical interaction. A fixed VLA provides the approach and hand targets; the supervisor decides whether to remain transparent, refine finger flexion, retain or release the corrected configuration, return control to the VLA for retry, or select a compliant hand-control regime. It uses finger-position and motor-effort feedback as proprioceptive contact evidence and requires neither tactile sensors nor VLA retraining. On a Unitree G1 with a BrainCo Revo2 hand, a randomized matched-block evaluation on five objects held out from VLA training yields 61.3% completion for the base VLA, 72.0% for unconditional close-to-stall control, and 84.0% for the supervisor under a shared budget; the gain is positive on every object and persists under moderate pose perturbations. Ablations show the gain is not explained by extended closure alone, and that selective triggering reduces correction episodes by 65.7% with no detected change in completion. A retention audit shows acceptance predicts retention in 88.9% of held-out cases, while compliant objects expose conservative false rejection. A thin-walled-cup study demonstrates contextual routing to compliant control, matching an always-compliant reference. These results suggest that contact-guided execution-time adaptation can improve the object-level generalization of a fixed VLA to held-out objects by adapting physical realization without object-specific retraining.

Vision-language-action models benefit from the understanding and reasoning capabilities of pretrained vision-language models, but action-only supervision provides limited grounding in world dynamics. Conversely, world-action models inherit spatiotemporal priors from video generation models, yet remain limited in semantic understanding and reasoning under distribution shifts. We introduce UniWAM, a unified architecture that integrates a physical reasoner, a world generator, and an action predictor to jointly learn semantic understanding of the physical world, visual generation, and action prediction. To ensure the quality of the training data, we developed a rigorous data cleaning and annotation pipeline for both human egocentric data and robot data. To adapt the vision-language component to embodied tasks while preserving its inherited language capabilities, we represent low-level actions in natural language and introduce a pre-training recipe that assigns complementary supervision from visual question answering (VQA) data, human egocentric data, and robot demonstrations to the appropriate model components. During post-training, future visual noise augmentation reduces reliance on precise future predictions, while history-conditioned flow matching uses encoded action history to initialize action generation. Together, these designs significantly reduce denoising steps while maintaining performance. UniWAM achieves state-of-the-art (SOTA) performance across multiple evaluations, including in-distribution performance, robustness, generalization, instruction following, and long-horizon task execution. Furthermore, we uncover a log-linear scaling law of unified human-robot co-training, demonstrating the effectiveness of large-scale pre-training on a mixture of human and robot data.

Recursive computation repeatedly compresses or reuses intermediate states, creating a simple tension: information that must remain useful across longer recursive paths is also exposed to more opportunities for loss before reaching the final prediction. Existing reconstruction or local-prediction objectives provide tractable supervision, but do not ensure that the retained information remains sufficient for subsequent recursive computation. We identify local predictive sufficiency with recursive predictive closure: controlling local predictive deficiencies at individual interfaces controls the resulting discrepancy at the root. We then turn this principle into a tractable training procedure. Starting from a variational characterization, we derive finite predictive tests and an empirical predictive deficiency that measures predictive value retained across compression. Its predictive sensitivities define margin-relaxed half-space constraints on parameter updates, and we project the host optimizer's proposed update onto their intersection only when predictive preservation would otherwise be violated. Across temporal graphs, language memory, vision-language-action control, and recursive self-improvement, the method matches or improves the corresponding host models under matched compression budgets, with larger gains under heavier recursive or memory demands, while better preserving predictive information across successive transformations. Crucially, the same task-agnostic predictive-preservation principle is instantiated across all four settings through host-compatible interventions while keeping the endpoint task, backbone, and evaluation protocol fixed. These results establish predictive preservation at recursive interfaces as a general training principle for recursive compression.

Recent Vision-Language-Action (VLA) models equipped with Flow Matching (FM) action heads achieve state-of-the-art performance in complex robot manipulation. However, the multi-step iterative ODE solving required by FM introduces inference latency that precludes responsive physical control. While current acceleration efforts optimize the Vision-Language Model (VLM) backbone, the action head bottleneck remains overlooked. To address this, we propose ProbeFlow, a training-free adaptive inference framework tai- lored for continuous robotic control. By evaluating geometric trajectory complexity via the cosine similarity between initial and lookahead velocity vectors, ProbeFlow dynamically sched- ules integration steps to prune redundant network evaluations. On the MetaWorld benchmark, it accelerates action decoding by 14.8x (reducing average steps from N = 50 to 2.6) and cuts end-to-end system latency by 2.8x without compromising the manipulation success rate. On the long-horizon LIBERO benchmark, the probe automatically allocates a denser schedule to navigate semantic bottlenecks, effectively resolving the flow solver delay. Real-world physical deployments confirm that ProbeFlow successfully mitigates action decoding latency while ensuring execution stability, offering a highly practical solution for low-latency continuous generative policies.

Robotic manipulation often requires inferring task-relevant states from past interactions when the current observation alone is insufficient to determine the appropriate action. Despite progress in benchmarking memory-augmented vision-language-action (VLA) models, application-oriented tasks requiring history-dependent semantic inference remain underrepresented. We introduce GiT (Grounded in Time), a dataset and benchmark for grounding manipulation decisions in past events across biolaboratory, household, and industrial scenarios. It includes real-robot and Universal Manipulation Interface (UMI) style demonstrations covering 18 bimanual tasks, together with simulation data and a ManiSkill-based evaluation suite covering nine tasks. Fine-grained subtask annotations and annotated counterfactual task pairs, in which similar current observations require different actions depending on prior events, support policy learning and targeted evaluation of history use. Evaluations of representative end-to-end VLA models in simulation and on selected real-world tasks reveal substantial room for improvement in history-dependent manipulation. The dataset and benchmark are available at the project page.

Vision-language-action models (VLAs) inherit strong semantic grounding from pretrained vision-language backbones but are typically optimized for predicting actions rather than future observations. They can see and act, but they do not imagine the future before acting. World action models (WAMs) built on video diffusion backbones can imagine but treat language as frozen conditioning on a continuous latent space. Unified models bring these modalities into one architecture, but they either decode autoregressively, one token at a time, or keep video continuous with an auxiliary action head. In this work, we present SUAVE, a Single vocabulary Unified Action-Video modEl in which a masked diffusion transformer generates video and actions conditioned on language, with all three modalities represented as discrete tokens in a shared sequence. Choosing which tokens to mask at inference turns the same network into a world model, a robot policy, or a video-action model. For action-free co-training, the action positions of unlabeled video are filled with mask tokens and excluded from the loss. Simulation and real-world experiments demonstrate two findings. First, a single SUAVE model predicts long-horizon video and acts as a policy, competitive with dedicated world models and specialized action policies on static and dynamic manipulation tasks. On a real robot, our model generates subgoal images and an action chunk spanning one second of motion in 1,030 ms on an RTX 5090 GPU, sustaining closed-loop control at 2.5 actions per second. Second, pretraining on robot video and co-training on human video substantially improves policy performance and zero-shot robustness to distribution shift. Together, these results show that masked diffusion is a practical and versatile foundation for unified video-action modeling.

Mastering robot manipulation skills via reinforcement learning (RL) remains largely sample-inefficient. The most common RL algorithms rely on random action sampling to discover new strategies, resulting in agents that allocate most of their training budget to motions in free space, away from the contacts from which manipulation skills emerge. Existing intrinsic motivation methods based on model disagreement or epistemic uncertainty improve on isotropic noise, but they can also reward uncertainty in functionally irrelevant transitions, such as erratic motions in free space. In this work, we argue that tactile feedback provides a natural signal for exploration, and introduce TacEx, a framework that incorporates touch into epistemic uncertainty-driven exploration by decomposing model uncertainty across sensory modalities and directing curiosity toward the tactile channel. By anchoring curiosity to the sense of touch, TacEx drives the robot to discover complex contact dynamics, learning to manipulate and grasp objects without task rewards or expert demonstrations during exploration. The interaction-dense dataset collected through this tactile-driven curiosity supports offline learning of downstream pick-and-place policies without additional environment interaction. We further use tactile-driven exploration to post-train vision-language-action (VLA) models. Although the VLAs are initially pre-trained without tactile feedback, post-training with TacEx substantially improves downstream performance while remaining highly sample-efficient.

16 pages, 6 figures, 1 table. Preprint, under review

Adversarial patches can disrupt Vision-Language-Action (VLA) models by manipulating visual observations, leading to failures in robot control. However, it remains poorly understood which internal mechanisms underlie these failures and how targeted interventions can mitigate them. In this work, we mechanistically analyze VLA representations using a sparse autoencoder (SAE) and identify a feature whose activation strongly correlates with the presence of an adversarial patch. Based on this analysis, we suppress the identified feature at inference time only when a linear probe detects an attack. This intervention improves robustness without the cost of fine-tuning the VLA. We evaluate our method against VLA adversarial patch attacks on LIBERO-10. Conditional intervention improves success rate under intermittent attacks, whereas continuously applying the same intervention substantially degrades policy performance. These results show that attack-related internal representations can provide useful targets for VLA adversarial defense and that controlling when to intervene is important for limiting disruption to nominal policy behavior.

Long-horizon mobile manipulation presents significant challenges due to compounding execution errors and capacity interference between locomotion and arm control. While recent Vision-Language-Action models excel at short-horizon tasks, they lack the hierarchical reasoning required for multi-stage objectives. Furthermore, existing hierarchical agents suffer from rigid sub-task mapping, inflexible replanning, and a lack of continuous learning. To address these limitations, we introduce MobiAgent, a dual-loop agentic framework that bridges robust deployment execution and recursive policy self-improvement. During deployment, the Inner Loop decouples high-level reasoning from low-level control through highly composable atomic skills. It employs Vision-Language models for receding-horizon planning and visual reflection, dynamically composing skills to ensure robust error recovery. These skills are executed by specialized flow-matching experts that share a unified VLM backbone, maximizing reusability while mitigating capacity interference. Concurrently, the Outer Loop drives automated lifelong learning by autonomously segmenting and verifying deployment rollouts, clustering them to discover atomic skills, and continuously fine-tuning the skill library without human annotations. Evaluations on RoboCasa, BEHAVIOR-1K, and real-world tasks demonstrate the effectiveness of MobiAgent. It outperforms $π_{0.5}$-TA by 22.5 percentage points on BEHAVIOR-1K and enables robust recovery from execution failures. Through autonomous data recycling, success improves from 7.50% to 27.50% on RoboCasa and from 32.5% to 57.5% on Astribot S1.

Accepted at the Conference on Robot Learning (CoRL) 2026. Project page: https://kaiknower.github.io/mobiagent

Vision-language-action (VLA) benchmarks measure whether a policy completes a requested manipulation task, but binary success can hide safety violations along the trajectory: a policy may reach the goal while applying excessive contact, disturbing bystander objects, destabilizing a held object, or entering robot self-contact. We present SafeVLA-Bench, a post-hoc safety-evaluation framework for existing simulator-based VLA benchmarks that reveals violations missed by success-only evaluation. It encodes task-aware safety requirements as Signal Temporal Logic (STL) invariants with quantitative robustness semantics. Alongside native success, it reports the safety rate and the success-but-unsafe rate (SBU) used in prior safety evaluations, and introduces the Violation Severity Index (VSI), a bounded worst-violation depth score. We instantiate SafeVLA-Bench on LIBERO and RoboCasa-365, evaluating twenty-seven policy-benchmark entries across tabletop and kitchen manipulation tasks. High task success does not imply safe execution: the fifteen tabletop policies above 90% mean success still have 18-28% unsafe-episode rates, and 38-56% of successful RoboCasa-365 rollouts violate at least one active safety clause. A post-training case study further shows that SafeVLA-Bench can be used to improve policy safety. Project page: https://safevla.org

46 pages (10 main + references + appendix), 7 figures, 21 tables. Project page: https://safevla.org

Vision-Language-Action (VLA) models that couple pretrained Vision-Language Models (VLMs) with continuous action experts have achieved strong manipulation performance, yet generalization to out-of-distribution (OOD) language instructions remains poor. A known challenge is the structural imbalance in VLA data, where language is far less diverse than visual and action content, making policies prone to visual shortcuts. While discrete-action methods mitigate this through vision-language co-training, continuous action experts lack such protection: they start from random initialization and learn entirely from imbalanced data, producing noisy gradients that corrupt the VLM and fail to exploit its language capability. We address this from a Bayesian perspective, factorizing the policy into a language-agnostic Vision-Action (VA) prior and a language-conditioned VLA likelihood, and propose APT, a two-stage training method emphasizing Action expert PreTraining. In Stage 1, the action expert is pretrained as a VA prior on vision-action pairs from a frozen VLM, bypassing the language imbalance. In Stage 2, language tokens are injected through a gated fusion mechanism that integrates VLM features while preserving the learned visuomotor prior. APT applies to mainstream VLA architectures, including the $π$ and GR00T-style architectures. Comprehensive experiments validate that APT achieves consistent gains on unseen instructions and compositional tasks. Project Page: https://xukechun.github.io/papers/APT/

Vision-Language-Action (VLA) driving augments end-to-end (E2E) planning with language-enabled visual backbones, but how VLM representations differ from vision-only encoders after policy learning, and whether such differences matter for planning, remains unclear. Under a unified VLM-hidden + diffusion-policy paradigm, we compare multiple VLM families/scales (InternVL3 and Qwen3VL) with standard vision-only encoders (ResNet, ViT, and EVA-CLIP) while keeping the downstream planner fixed. We study representation, behavior, and system design. CKA/CCA and Shared--Unique SAE show that policy learning enlarges a common decision subspace, but both branches retain non-transferable residual factors. We further replicate this shared-plus-unique representation pattern on nuPlan using the AsyncDrive planning stack. Latent-intervention policies and scenario-level analysis further show that these residuals are behaviorally meaningful: vision-only encoders are stronger in simple geometry-dominant scenes, whereas VLMs are more effective in semantically complex and interaction-heavy long-tail cases. The two branches also exhibit distinct progress--braking and path-choice tendencies, and an oracle best-of-two VLM+ViT selector reaches 93.58 PDMS on NAVSIM. We convert this complementarity into two lightweight systems: HybridDriveVLA, which selects from a compact cross-model candidate set using a learned trajectory scorer and improves PDMS from 90.80 to 92.10, and DualDriveVLA, a fast--slow variant that invokes the VLM in only 15% of scenarios, achieving 91.00 PDMS with about $1.9\times$ lower latency than the VLM baseline. Code will be available at https://github.com/WilliamXuanYu/HybridDriveVLA.

Accepted at NeurIPS 2026

Human interventions provide crucial corrective signals for post-training Vision-Language-Action (VLA) models. However, enabling seamless humanoid interventions is a formidable systems challenge due to complex whole-body kinematics and dexterous-hand control. Consequently, the collected intervention trajectories are often suboptimal, and methods that rely on human interventions as expert supervision can absorb hesitant, inefficient, or even erroneous behaviors. To address both the system and algorithmic challenges, we propose ROVE, a reinforcement learning framework for humanoid VLA post-training with imperfect human interventions. First, ROVE introduces a human-in-the-loop pipeline capable of collecting deployment and intervention data for humanoid manipulation. Second, it utilizes Optimistic Value Estimation (OVE) to prioritize high-value behaviors from mixed-quality trajectories. To further robustify value estimation, we incorporate cross-embodiment human experience videos to provide rich supervision for long-tailed failure and recovery modes. The resulting critic yields informative advantage signals, steering the VLA actor to focus on high-value behaviors rather than indiscriminately imitating all actions. On challenging real-world contact-rich and fine-grained humanoid manipulation tasks, ROVE outperforms experience-learning baselines and consistently improves across multiple rollout-intervention iterations.

Generalist robot policies, such as vision-language-action models (VLAs) and world-action models (WAMs), combine powerful pretrained backbones with expressive generative action heads trained via flow matching on large-scale robotic datasets. Despite their strong empirical performance in robotic manipulation, these policies lack mechanisms to quantify confidence in their predictions and to detect when their actions may be unreliable. This presents a critical limitation for real-world deployment in non-stationary environments, where models inevitably encounter scenarios outside their pretraining distribution and may fail without warning. To address this, we derive an efficient method to quantify epistemic uncertainty in flow-matching models by leveraging velocity-field disagreement (VFD) across a small ensemble. We successfully use this uncertainty estimate for detecting failures during deployment and active fine-tuning of flow-based generalist policies. For the latter, we propose SAVE, a simple yet effective method for uncertainty-guided active multitask fine-tuning that reduces the number of costly expert demonstrations required to adapt generalist policies to new tasks. We conduct experiments in simulation and the real world, across VLAs and a WAM. VFD yields better-calibrated uncertainty estimates predictive of downstream performance and detects failures with 8 pp higher overall accuracy than existing methods. Across three real-world tasks, SAVE improves final average success from 39 % to 47 % with a fixed demonstration budget. Our results show that measuring epistemic uncertainty with VFD enhances both failure awareness and adaptation of generalist robot policies. Project website: tum-lsy.github.io/uq_generalist_policies.

Project page: tum-lsy.github.io/uq_generalist_policies/. 41 pages, 18 figures

DAgger improves robot policies by aggregating expert supervision from states visited during policy execution. Robot-gated DAgger automates expert queries, allowing the robot to decide when to request expert takeover. While existing gates emphasize detecting the need for assistance, takeover timing also shapes the content of these demonstrations and their value for policy learning. We propose TimelyDAgger, combining Bridge-PCA monitoring of internal vision-language-action (VLA) features with Feedback-guided Threshold Adaptation based on expert behavior to improve takeover timing. We introduce an evaluation framework linking failure detection, takeover timing, and policy improvement, including Target-Aligned Supervision Ratio (TASR) for assessing supervision quality without retraining. Experiments show that takeover timing affects policy learning, with TimelyDAgger achieving competitive failure detection and higher post-training success in most evaluated settings under matched expert-action budgets. Project website: https://seen-e.github.io/TimelyDagger/.

8 pages, 10 figures, 1 table

We present Move-Then-Operate, a Vision language action framework that explicitly decouples robotic manipulation into two distinct behavioral phases: coarse relocation (move) and contact-critical interaction (operate). Unlike monolithic policies that conflate these heterogeneous regimes, our architecture employs a dual-expert policy routed by a learnable phase selector, introducing a structural inductive bias that isolates phase-specific dynamics. Phase labels are automatically generated via an MLLM-based pipeline conditioned on lightweight contextual cues such as end-effector velocity and subtask decomposition to ensure alignment with human motor patterns. Evaluated on the RoboTwin2 benchmark, our method achieves an average success rate of $68.9%$, outperforming the monolithic $π_0$ baseline by $24%$. It matches or exceeds models trained on $10\times$ more data and reaches peak performance in $40%$ fewer training steps, demonstrating that architectural disentanglement of move and operate phases is a highly effective and efficient strategy for mastering high-precision manipulation.

World models learn to predict how their environment will evolve, making them an important foundation for general-purpose robotic control. Yet world action models depend on camera inputs whose manipulation can corrupt the visual representations used across tasks and action policies. Existing attacks on these models optimize against the victim's actions or predicted futures and therefore require access to target-model outputs. In this paper, we propose an attack, TAPDreamer, against world action models that instead uses a public encoder alone to construct a fixed local perturbation that transfers across tasks and action architectures. TAPDreamer requires no target-policy queries. Our key insight is that interactions between patch-induced changes in attention weights and value vectors broadcast a nearly identical representation shift far beyond the patch footprint, and this shift remains stable across task observations. Guided by this insight, TAPDreamer uses six frames from one source task to maximize the global L1 distance between clean and patched encoder representations. In closed-loop evaluation, one frozen patch per benchmark, covering about 6.5% of the input, reduces FastWAM's success rate from 97.7% to 0.0% across 40 LIBERO tasks and from 90.8% to 0.0% across 50 RoboTwin tasks; matched random patches retain 81.5% and 79.2% success. The same patches reduce success to 2.1% and 0.8% on two DreamWAM configurations and to 10.0% on Motus. These results show that protecting downstream action generation alone is insufficient: defenses for world action models must also secure shared visual encoders against persistent local perturbations.

Project Page: https://tapdreamer.github.io

Long-horizon planning with latent world models requires reasoning across timescales and levels of abstraction. Existing task-agnostic JEPA world models predict and plan at a single timescale or with multiple horizons in one shared latent space. We introduce H-JEPA, an end-to-end recipe for training a hierarchy of action-conditioned JEPAs in which each level predicts farther ahead in its own learned latent space. Planning proceeds top-down: the top level optimizes progress toward the goal, and each level's predictions become subgoals for the planner below it. When factors in the data evolve at separated timescales, higher levels discard fast, unpredictable detail and retain slower task-relevant state. Across four simulated navigation and manipulation environments, hierarchical planning improves over a flat JEPA; on Visual AntMaze, a three-level hierarchy raises success from 18% to 73% using less planner compute. Ablations attribute these gains to both temporal decomposition and higher-level goal representations. With inverse-dynamics supervision, the approach extends to diverse real-robot videos from DROID, where hierarchy improves offline planning fidelity at lower planner compute.

Occupancy world models use historical occupancy states to forecast future 3D scenes, but their robustness under corrupted temporal inputs remains poorly understood. Existing evaluations primarily emphasize clean forecasting accuracy and provide limited evidence about how errors enter, persist, and propagate through the occupancy perception-forecasting chain. This paper introduces OccStress, a robustness stress-testing benchmark for the occupancy forecasting chain. OccStress contains 21 corruption families with 61 severity configurations and 10,827 strict temporal anchors across 3 datasets. OccStress covers both 3D occupancy perception and 4D occupancy forecasting through two complementary tracks. This design separates model-mediated pipeline errors under standardized sensor stressors from the intrinsic sensitivity of 4D forecasting models to corrupted occupancy states. OccStress further defines temporal injection protocols to test whether errors in the current state, recent history, or earlier history affect future forecasts differently. OccStress provides aggregate metrics for evaluating robustness along the occupancy forecasting chain. Experiments with five 4D forecasters reveal that current occupancy models are substantially affected by both upstream prediction errors and direct state corruptions, and that clean performance alone is an insufficient description of source- and position-specific robustness. Code, data, and evaluation tools are available at https://insailab.org/OccStress.

Accepted by NeruIPS 2026

Methods for generalization in model-based reinforcement learning typically assume that an agent cannot recover the latent context governing the environment dynamics from its own experience, and therefore supplies it externally. We formalize and test this assumption with \emph{predictive sufficiency}, which quantifies what access to the context adds to next-step prediction under the visitation distribution an agent induces, and separates that quantity into a history-recoverable part, a residual requiring the true context, and the deficit added by a finite model. We classify context-aware algorithms by the predictive risk their conditioning set can target and demonstrate across environments of increasing identification difficulty that the headroom does not follow the MDP class. The same task under different priors leaves predictive headroom in one setting and nothing distinguishable from zero in another, where the agent's behavior implicitly identifies the context and any benefit of such a mechanism cannot be attributed to missing information. Where headroom persists, the learned state exposes it only partially, and adding the true context still lowers the risk. Our contribution is a practical criterion for matching contextual mechanisms to the information available to them, estimated from the ordinary trained agent without a reference policy.

World models must predict how an environment evolves under sequences of actions, enabling agents to compare possible futures and reason about counterfactual actions before acting. Long-horizon prediction is commonly obtained by recursively applying a one-step transition model, but intermediate errors can compound over time. Multi-step dynamics models instead condition on a sequence of future actions and predict their consequences directly, but become harder to learn as horizon grows: the model must track interacting state changes across the trajectory, endpoint supervision provides weak credit assignment, and intermediate predictions can remain plausible while losing information needed for later states. We show that these challenges can be addressed by casting the internal evolution of a multi-step transition as structured reasoning over textual world states: reasoning over sparse state changes reduces the burden of state tracking, a predictive-gain objective rewards the learned state for improving over a matched predictor that conditions on raw history instead, and intermediate predictive rewards supervise each state along the trajectory. Because these intermediate states are explicit textual representations of the world, they provide semantically meaningful targets that can be inspected, scored, and corrected during training. Across ScienceWorld, Jericho, and CEO-Bench, our approach achieves the strongest average long-horizon performance against recursive and non-recursive baselines that condition directly on raw history, with gains increasing at longer horizons. In a controlled counterfactual study, our model is also the only one with statistically significant sensitivity to future actions.

Action-conditioned robot world models must respond precisely to robot trajectories while preserving realistic visual dynamics, yet learning both from heterogeneous robot videos remains challenging. Simulation offers structured motion supervision, but appearance differences hinder direct transfer, and inaccurate simulation predictions can misguide real-video generation. We present SimForcing, a simulation-guided framework that uses simulation both as a source of transferable motion knowledge and as a controllable reference for prediction. First, we transfer motion knowledge from a simulation teacher through latent-motion distillation, aligning temporal changes in latent space to internalize motion priors while mitigating the influence of appearance differences. Second, we introduce multi-block simulation conditioning with condition dropout to exploit predicted simulation trajectories without relying excessively on their accuracy. Our simulation-conditioning classifier-free guidance scheme unifies these two ideas by balancing predictions based on internalized motion knowledge with those additionally guided by simulation latents. The jointly trained student generates both simulation conditions and real-domain videos, requiring no additional world model at inference. On Bridge, SimForcing achieves the best PSNR, SSIM, LPIPS, and FVD among the compared methods without external embodied pretraining. Evaluation on InternData-A1 further supports its applicability across robot datasets. Moreover, using our trained world model to initialize a vision-language-action model improves LIBERO success, suggesting its utility for downstream policy learning. \url{https://github.com/Wang-Xiaodong1899/SimForcing}

Code: https://github.com/Wang-Xiaodong1899/SimForcing

World-model controllers rely on action-conditioned dynamics for prediction and planning, yet real control systems often execute commands asynchronously due to communication delay, packet loss, reordering, and actuator buffering. We study how asynchronous execution changes the action semantics assumed within world-model controllers, rather than treating it only as an external control disturbance. Through controlled interventions, we identify two architecture-dependent failure modes: planning-based controllers such as TD-MPC2 suffer from a future-action timeline mismatch between imagined and executed action sequences, while recurrent world models such as DreamerV3 can attribute observed transitions to commands that were not actually applied. Our analysis shows that TD-MPC2 requires the correct future action sequence during latent dynamics rollout, whereas DreamerV3 requires timely attribution of each transition to the action that generated it. Based on these findings, we introduce two lightweight execution-consistent interfaces, Future-Sequence for TD-MPC2 and Applied-Action Feedback for DreamerV3, that correct these mismatches without modifying the pretrained world models. Experiments across delays, packet loss, reordering, multiple control domains, measured network traces, and a process-separated asynchronous stack consistently support both diagnoses and the corresponding architecture-specific corrections.

Embodied world models predict the visual consequences of candidate actions before execution. However, existing action-conditioned world models often adopt uniformly weighted visual generation objectives that can be misaligned with embodied prediction needs. Even with explicit motion conditioning, these objectives can underemphasize spatially sparse changes that are critical to interaction. We propose KineWorld, a transport-aware world-modeling framework that extends robot kinematics from motion conditioning to the spatial allocation of generative supervision. Kinematic Transport Lifting (KTL) constructs renderer-derived, camera-aligned transport fields from commanded robot motion. Transport-Aware World Diffusion (TAWD) calibrates their motion support on the video-latent grid and reweights future-RGB flow matching through a normalized mixture of uniform and transport-focused distributions. We train KineWorld using ALOHA-AgileX bimanual manipulation data from RoboTwin 2.0. KineWorld achieves an EWMScore-P of 68.95 in single-view evaluation and a TWB-Score of 54.82 in multi-view evaluation. These results support a shift from appearance fitting toward action-consequence modeling for embodied decision-making.

36 pages. Project page and code: https://modaxiansheng.github.io/KineWorld/

As embodied AI and world models increasingly operate in dynamic 3D environments, visual perception must move beyond passively interpreting given observations toward actively deciding what to observe. We study this problem through camera planning in dynamic 3D story worlds, where the camera must not only generate smooth motion, but also decide what visual evidence should be acquired before it moves. We formulate this capability as Narrative-Grounded World Visual Attention, where the camera acts as an embodied observer that determines what to observe, how to compose the observation, and how to shift attention over time under narrative intent and physical 3D constraints. To realize this capability, we propose Look-Before-Move, a camera planning framework that separates observation specification from motion execution. It first builds a Semantic Observation Contract to convert directorial intent into executable visual constraints, then performs Monte Carlo Viewpoint Search to find narrative-compliant and geometrically feasible viewpoints, and finally applies Semantic Trajectory Grounding to connect selected viewpoints into continuous, collision-aware, and temporally coherent camera motion. We further construct a dynamic 3D Story World Benchmark based on StoryBlender, covering 50 stories, 457 scenes, and 1585 shots with animated characters, semantic scene configurations, and executable 3D environments. Experiments show that our framework improves subject perception, intent consistency, and trajectory quality over representative baselines, demonstrating the importance of organizing visual attention before generating camera motion.

Accepted at NeurIPS 2026 (Main Track, Poster). 30 pages (including references and appendices), 19 figures

World models, which learn how environments respond to actions, are emerging as a powerful paradigm for planning through imagined futures, transforming decision-making across games, robotics and autonomous driving. Bringing this capability to manufacturing could enable process decisions on timescales inaccessible to high-fidelity simulation. Here we introduce a generative world model for localized highly dynamic laser melt pool that predicts evolution from histories of temperature and phase morphology under candidate actions. Its generative latent dynamics capture the effects of unresolved melt flow, enabling more accurate recursive rollouts than deterministic regressors under transient laser inputs. Because the learned dynamics are differentiable, the model can serve directly as a predictive control plant. Gradients through imagined futures optimize laser schedules that regulate melt-pool depth over previously unseen geometry, path, initialization. We further distil this optimization into an amortized policy that produces control actions in a single forward pass, providing a proof of concept for real deployment on machines.

Programmable world models separate executable dynamics from visual generation, offering a promising foundation for next-generation game engines. However, their visual adherence to explicit rules and interactions remains insufficiently evaluated. Existing benchmarks assess visual quality, controllability, and instruction or physical adherence, but rarely test fidelity to fine-grained, program-specified world events. We introduce PROWBench, comprising 170 programmatically constructed episodes and 600 proxy videos covering diverse scenes and interactions. PROWBench logs entity states and timestamped events, including those outside the camera's field of view, as replayable world records, from which it renders synchronized views and proxy representations. This enables generated videos to be checked against the observable consequences of program execution. An extensible framework constructs scenes, controls behaviors, and can render each camera view in different representations, such as coarse 3D, and bounding boxes. The benchmark covers first- and third-person perspectives, with synchronized multi-view observations available for a subset of episodes. Grounded in these records, PROWBench evaluates entity control, long-horizon memory, and, with two VLM-based metrics, Logic-Render Alignment and Interaction Success Rate, adherence to the prescribed timeline and the visual realization of timestamped engine-recorded events.

Project page: https://alaya-lab.github.io/PROWBench

The enduring vision of general-purpose robots serving humanity hinges fundamentally on policy steerability. However, prevailing paradigms of learning from expert demonstrations demand massive real-world data even on simplified grippers, rendering them prohibitively expensive for high-dimensional, data-scarce dexterous hands. To overcome this bottleneck, we present a full-stack system that scales dexterous VLA pre-training from egocentric human videos and enables data-efficient real-robot post-training. It integrates EgoSmith, a data pipeline that curates in-the-wild egocentric videos into 9,606 hours of pre-training data with 8.3x higher throughput and better accuracy than prior SOTA; a unified Robot Stack for teleoperation and human-in-the-loop correction tailored for dexterous hands; and EgoSteer, a world-model-enhanced VLA operating on a morphology-aligned action space. Human data pre-training equips EgoSteer with language-guided manipulation priors, which are grounded through robot post-training and further refined via DAgger. Empirically, EgoSteer robustly executes free-form instructions across 45 diverse tasks, demonstrating adherence to user intent amid multiple candidate tasks and generalization. The pre-trained model also few-shot adapts to five complex long-horizon tasks, including box folding, on two embodiments with 79% average progress. All system code, datasets, model checkpoints, and an evaluation gallery are publicly available at https://egosteer.github.io/.

Realistic environment replicas are increasingly valuable for training and evaluating LLM agents, yet the original systems may be inaccessible or impractical to reproduce. We explore agentic language world modeling: rather than rebuilding an executable environment, a world model agent serves as the environment for a task agent and supports faithful and stateful simulation. We instantiate this paradigm with Trace2Env, a learning-free framework for settings where the original system is unavailable but historical interaction traces remain accessible. Trace2Env reconstructs these traces into a reusable environment worldbook containing environment schemas, grounded evidence, and induced behavioral knowledge. At runtime, the world model agent actively consults the worldbook together with persistent episodic state to infer each action's observation and lasting state effects. Across nine environments, Trace2Env improves both next-observation fidelity and long-horizon interaction consistency over conventional prompt-based LWMs. In multi-turn interaction, task agent actions generated against Trace2Env remain valid more often when replayed in the real environment, indicating that its simulated dynamics better preserve the consequences of earlier actions across successive turns. These results establish agentic language world modeling as an alternative direction for building realistic environment replicas without reconstructing the original executable system.

Existing game world models typically adopt role-specific interactions, where player and NPC roles are bound to fixed characters. This limits their flexibility in multi-character games, where different characters may receive external control while NPCs must react to interactions triggered by players. This setting raises two key challenges: how to flexibly assign control roles to individual characters, and how to support direct player control and reactive NPC behavior within a unified model. These challenges are particularly pronounced in shared-view 2D games, where multiple, potentially visually identical characters share the same viewpoint, making camera cues insufficient to distinguish their roles. To address these challenges, we introduce ReactiveGWM, a reactive game world model that flexibly assigns control modes at initialization and jointly simulates externally controlled players and reactive NPCs. Specifically, ReactiveGWM introduces Spatial Role Binding, which grounds learned character handles to their corresponding regions in the initial frame using instance masks. Building on these handles, Unified Agency Conditioning unifies heterogeneous control signals across characters by encoding player actions and conditional NPC rules into character-specific control-token groups. Each group is then bound to its corresponding character handle, enabling the model to apply each control signal to its designated character. Meanwhile, causal self-attention restricts temporal context to the current and preceding latent frames when generating player actions and NPC responses. Experiments on two multi-character 2D games demonstrate that ReactiveGWM supports flexible character control across different player/NPC role assignments while jointly generating accurate player-controlled behaviors and reactive NPC responses, enabling more configurable and richer multi-character interactions.

The code is available at https://inv-wzq.github.io/ReactiveGWM/

What does it mean to understand the world? Contemporary world models often operationalize understanding as accurate future prediction in latent or observation space. Developmental cognitive science, however, suggests a different view: human understanding emerges through the construction of internal theories of how the world works, even before mature language is acquired. Inspired by this theory-building view of cognition, we introduce Learning-to-Theorize, a learning paradigm for inferring explicit explanatory theories of the world from raw, non-textual observations. We instantiate this paradigm with the Neural Theorizer (NEO), a World Theory Model, that induces latent programs as a learned Language of Thought and executes them through a shared transition model. In NEO, a theory is represented as an executable, compositional program whose learned primitives can be systematically recombined to explain novel phenomena. Experiments show that this formulation enables explanation-driven generalization, allowing observations to be understood in terms of the programs that generate them.

Ultrasound is widely used, but acquisition quality is heavily dependent on the operator's knowledge and expertise. With demand for examinations outpacing the supply of trained sonographers, operator-guidance systems aim to close this gap by instructing a less trained user how to move the probe toward a target view. In this paper, we propose a retrieval-induced latent transition model for ultrasound acquisition dynamics, formulating ultrasound operator guidance as multi-step planning and retrieval in a world model. Using a V-JEPA 2.1 backbone, observations are first encoded into a latent space where anatomically related views lie close together. We then retrieve similar views from a reference database containing encoded latent states and corresponding probe positions and orientations. Rather than learning a parametric transition function, we directly use physically executed transitions from the database to establish our nonparametric, retrieval-induced transition model that supports receding-horizon planning. At deployment, guidance is generated from the live ultrasound image feed alone, without any probe tracking hardware. Applied to carotid ultrasound, the proposed planner reaches the target view in 86% of retrospective closed-loop episodes, versus 52% and 43% for representative baselines, outperforming both on every target view, including the challenging longitudinal internal and external carotid artery views. A prospective feasibility study on unseen volunteers, run in real time on a CPU using distillation, reaches 83% target-view reachability. Because planning is driven by proximity to any encodable goal latent, the same world model can navigate back to any previously acquired, patient-specific frame, supporting reproducible longitudinal imaging for e.g. perioperative or follow-up monitoring.

11 pages, 8 figures, 5 tables

Latent world models based on Joint-Embedding Predictive Architecture (JEPA) are deterministic by design. While successful in fully observable scenarios, this paradigm breaks down when past observations and actions lead to multiple plausible future possibilities, e.g., due to occlusion. We introduce EpicWorldModel, a framework to train stochastic JEPAs for environments and tasks with inherent uncertainty under partially observability. We jointly train the EpicWorldModel predictor with its latent representation space to directly predict multiple potential future states using a flow-matching objective, when the goal-relevant scene content is absent from the conditioning history. We show that flow predictive variance, motivated by its relation to an upper bound on predictive entropy, serves as a useful exploration guidance for planning. By incorporating this uncertainty signal into Cross-Entropy Method (CEM)-based planning, our approach balances goal-reaching with exploration of uncertain regions where occluded goals are most likely to be located. We demonstrate the effectiveness of EpicWorldModel through a series of latent planning experiments with the best or on-par performance across tasks, showing up to 22% empirical improvement in success rate over LeWorldModel.

Accepted at NeurIPS 2026. 18 pages, 8 figures, 5 tables (11 pages main text, 4 pages appendix)

The last mile toward enterprise AGI is a company that runs itself. Training and adapting such agents require longitudinal enterprise data, which remain scarce, costly to acquire, and often restricted by privacy constraints. Historical archives are also frequently incomplete and record only what actually happened. They cannot show the outcomes of alternative decisions. We introduce MiniCorp, an office simulator for studying how agents can collectively run a company while generating enterprise data at scale. Using an e-commerce company as a demonstration, MiniCorp connects two interacting worlds. The external world models customers, dynamic competitors, and market mechanisms. The internal world consists of agents that observe events, discuss their options, and make strategic decisions. These decisions have lasting effects on the market, and the resulting feedback informs the firm's later decisions. As the firm and market interact, MiniCorp continuously records the agents' communications and decisions. These records preserve the information available at the time and the business results that followed. Checkpointing allows the same situation to be replayed under different decisions, providing comparisons unavailable in static archives. We evaluate end-to-end fidelity against patterns reported in empirical studies of real markets. These evaluations provide agents with realistic market feedback and reduce the risk that they learn to exploit flaws in the simulator. Our experiments show agents coordinating across roles and adapting their decisions to market feedback. With explicit long-term strategic guidance, they also sustain advertising exploration despite weak early returns. MiniCorp thus provides an environment for studying AI-run companies and a scalable source of longitudinal and counterfactual enterprise data for agent training and evaluation.

Graphical User Interface (GUI) agents have emerged as a promising paradigm for automating complex digital workflows across diverse applications. However, training highly capable and generalizable agents fundamentally relies on massive, high-fidelity visual-action trajectories, which are notoriously difficult to acquire. While human demonstrations are unscalable, existing GUI world models rely on text descriptions or HTML rendering, discarding crucial pixel-level visual details like icons and layout styles. To address this issue, we introduce Infinite-Dreamer, a simulation-free data synthesis method powered by a pixel-level Image Editing World Model. By conceptualizing GUI transitions as image editing tasks, we leverage Vision-Language Models (VLMs) to describe action-induced UI changes as structured delta-text. We then fine-tune an image editing backbone to controllably synthesize realistic screenshot transitions. We utilize this model to generate both single-frame visual robustness data and multi-step imaginary trajectories. To validate the effectiveness of our approach, we fine-tune the Qwen3-VL baseline solely on the synthesized data to obtain Infinite-Actor, and evaluate it on AndroidWorld, MobileWorld, and AndroidControl-Curated benchmarks. Infinite-Actor consistently outperforms the Qwen3-VL baselines across scales: Infinite-Actor-8B improves AndroidWorld Pass@1 by +4.45 and nearly doubles the MobileWorld Pass@3 success rate, while Infinite-Actor-2B improves Pass@1 by +9.05. Code is available at https://github.com/swaydy-n/Infinite-Dreamer.

A common approach to world-model simulation for vision-language-action (VLA) systems is to predict future RGB observations and then re-encode them into policy inputs, introducing an indirect interface between simulation and downstream policy execution. We instead investigate whether world dynamics can be modeled in a compact, policy-oriented state derived from VLM visual tokens. A key challenge is that raw VLM visual tokens are high-dimensional, making efficient and accurate autoregressive dynamics modeling challenging. To address this, we introduce Token-World, an action-conditioned world model that compresses VLM features into a compact token state, learns future dynamics in this reduced space, and maps predicted states back to the original policy-facing representation for downstream use. Across manipulation benchmarks, Token-World improves open-loop feature fidelity and policy-action consistency over recent world-model simulators, with slower degradation over long rollout horizons. In closed-loop evaluation, its simulated policy performance correlates more strongly with reference policy performance than Ctrl-World ($r=0.794$ vs.\ $0.583$), while requiring lower simulation latency. Ablations further show that compact-representation design and dimensionality substantially affect future-state prediction. Code will be available at https://chuyaofu.github.io/Token-World/.

Submitted to IEEE International Conference on Robotics and Automation (ICRA) 2027

Long-horizon video world models require persistent memory to preserve scene consistency over extended rollouts. Softmax attention retains the full generation history through a growing KV cache, whereas recurrent linear attention compresses history into fixed-size states with substantially lower memory cost. However, we identify severe long-range forgetting in Gated DeltaNet (GDN), where information from distant but relevant scenes is progressively attenuated by subsequent state updates. To address this limitation, we propose HLA-WM, a training-free hybrid linear-attention framework that combines coarse-grained geometry-guided retrieval with fine-grained recurrent linear-state computation. HLA-WM exploits the affine structure of GDN to cache compact chunk-wise transition summaries, retrieve scene-relevant historical chunks using camera geometry, and recompose them into query-specific recurrent states. On the $60$-second SANA-WM-Bench, HLA-WM improves all six aggregate revisit-consistency and camera-control metrics of the base autoregressive generator without additional training, including a $0.74$ dB PSNR gain and a $28.5%$ reduction in rotation error. The improvements persist after downstream refinement and generalize to MBench-A, where HLA-WM consistently improves all three revisit-consistency metrics across all four subsets and all evaluated inference modes over $547$ samples. At a $60$-second context, HLA-WM reduces historical-state memory by $12\times$ relative to full KV caching while incurring at most a $1.6%$ reduction in inference throughput. These results demonstrate that selectively addressable recurrent memory can improve long-range scene recall while preserving the efficiency advantages of GDN. Project page: https://caesarhhh.github.io/hla-wm/

This paper asks whether a bounded neural architecture can exhibit a meaningful division of labor between intuition and deliberation on a classic 64-item syllogistic reasoning benchmark. More broadly, the benchmark is relevant to ongoing debates about world models and multi-stage reasoning in AI. It provides a controlled setting for testing whether a learned system can develop structured internal computation rather than only one-shot associative prediction. Experiment 1 evaluates a direct neural baseline for predicting full 9-way human response distributions under 5-fold cross-validation. Experiment 2 introduces a bounded dual-path architecture with separate intuition and deliberation pathways, motivated by computational mental-model theory (Khemlani & Johnson-Laird, 2022). Under cross-validation, bounded intuition reaches an aggregate correlation of r = 0.7272, whereas bounded deliberation reaches r = 0.8152, and the deliberation advantage is significant across folds (p = 0.0101). The largest held-out gains occur for NVC, Eca, and Oca, suggesting improved handling of rejection responses and c-a conclusions. A canonical 80:20 interpretability run and a five-seed stability sweep further indicate that the deliberation pathway develops sparse, differentiated internal structure, including an Oac-leaning state, a dominant workhorse state, and several weakly used or unused states whose exact indices vary across runs. These findings are consistent with reasoning-like internal organization under bounded conditions, while stopping short of any claim that the model reproduces full sequential processes of model construction, counterexample search, and conclusion revision.

Corrected Figure 1 to show candidate-logit aggregation before the final softmax and corrected KL-divergence notation. Model implementation and numerical results are unchanged

The component that dominates a latent world model's prediction error need not be the one whose repair most improves action selection. We show this by comparing action sequences from identical physical starts and separating endpoint error into a candidate-pool center and action-relative responses. Across four model families and four tasks, a confirmation pool of 256 new starts per task and 300 shared candidates per start shows that center error dominates MSE in 14/16 model-task cells. Yet in six of these cells, an oracle that corrects only the action-relative responses yields better physical rank correlation and top-30 elite quality than one that corrects only the center, while leaving more latent MSE (family-wise corrected intervals). The preference differs across the evaluated settings: a separate LeWorldModel (LeWM) study that executes oracle-selected actions favors center repair on PushT and on Reacher with a render-matched goal. Matched-candidate tests localize ordering loss: for LeWM, encoding realized endpoints raises physical Spearman from 0.464 to 0.975 on that Reacher setting and from 0.193 to 0.631 on PushT (64 starts per task), while Cube's encoded-goal cost remains uninformative. A 72-run objective study improves selected response diagnostics, while incremental closed-loop planning gains remain unconfirmed. These results separate error magnitude from the decision effects of oracle correction and motivate evaluating representation, prediction, and planning as separate stages.

World--action models (WAMs) promise a unified model that predicts action-conditioned futures, generates feasible actions, and supports planning in imagination. However, existing joint video--action models often use computationally heavy, fixed-horizon backbones ill-suited to streaming inference and stable long-horizon open-loop rollouts. We introduce FLEX-WAM, a Flexible and Efficient Block-Causal World--Action Model for unified simulation and policy inference. FLEX-WAM supports variable-length contexts and non-causal prediction horizons, as well as infinite autoregressive generation frame by frame or block by block. Its block-causal, KV-cacheable architecture combines axial attention and blockwise diffusion forcing to enable efficient real-time rollout and deployment-time latency--throughput tradeoffs without retraining. Joint training can nevertheless produce plausible futures that weakly respond to commanded actions. We address this failure mode by balancing state and action flow-matching gradient contributions across the state--action diffusion-noise grid and regulating world-model sampling using Forward-Dynamics (FD) elasticity, an efficient training-time proxy for action responsiveness. Across simulated and real-world datasets, FLEX-WAM achieves superior multi-step prediction quality and latency while producing stable joint state--action rollouts for thousands of steps. As a joint action proposer and simulator within MCTS, it solves long-horizon PushT and all five OGBench Puzzle-4x4 tasks entirely in imagination. On a bimanual OpenArm-based robot, a single checkpoint jointly serves as a play policy and expected-outcome predictor, enabling real-time identification and collection of model--reality mismatches for future self-improvement.

Action-conditioned time-series forecasting requires accounting for how future actions and exogenous forcings influence multiple targets through partially observed effects with different delays and persistence. Direct conditioning leaves the evolution and target-specific influence of these effects implicit in the predictor, while static relational graphs specify connections without tracking evolving effects. This motivates representing future-driver influence through structured latent states that evolve over the forecast horizon and route information to individual targets. We introduce BeliefGraph-JEPA, a structured latent world model that factorizes driver influence into typed latent-effect states. These states are rolled forward under future drivers and routed through a graph to target-specific nodes, forming the predictive base of a joint-embedding predictive architecture. A capacity-controlled residual supplements this base with direct driver information. On four multi-target clinical, agricultural, environmental, and industrial systems, the framework outperforms a range of pretrained and supervised known-future-covariate baselines. Matched controls isolate latent dynamics, future rollout, graph routing, and residual capacity; future rollout and graph-first residual routing improve forecasting across all four systems.

Learning action-conditioned world models for dexterous manipulation that are genuinely controllable requires capturing complex, high-dimensional hand kinematics from limited real-world data. We present Mask2Real-WM, a two-stage world model that improves controllability for dexterous hands under a limited real-data budget by decoupling pixel prediction into a dynamics model and a rendering model. The dynamics model predicts future segmentation masks from past masks and a high-dimensional action sequence. The rendering model converts the predicted masks into photorealistic RGB images. This design allows us to train the dynamics model on a large synthetic dataset spanning the full range of hand motions and interactions. We then fine-tune the dynamics model and train the rendering model on only 2.5 h of real demonstrations to obtain a controllable world model. We compare five models with matched training recipes on a 23-DoF robotic system, measuring per-DoF controllability both with random target commands and with sinusoidal per-DoF actuation, judged blind by human raters. The decoupled model achieves the strongest controllability, with further gains from simulation data. Beyond improved controllability, Mask2Real-WM produces sharper frames while maintaining comparable perceptual quality.

8 pages, 7 figures, 3 tables. Preprint. Project page: https://srl-ethz.github.io/Mask2Real-WM/

Video world models are largely regarded as predictive models of the physical world and are therefore expected to anticipate the consequences of observed events. However, evaluation has mainly focused on reference similarity, physical-law consistency, or judgment plausibility, estimating anticipation only indirectly. We address this directly: when a release or impact has just occurred but its consequence is withheld, can a world model anticipate what should happen next? We introduce an event-anchored evaluation based on 62 controlled real-world free-fall recordings and 124 clips spanning three object types, with fine-grained release and impact annotations and ground-truth trajectories. The protocol separates consequence production, temporal placement, and physical realization. Across six contemporary video generation and world models, Runway and Veo produce release and subsequent impact events at rates above 93% but often initiate them substantially late, whereas Cosmos-Predict-2.5 and MAGI-1 frequently preserve the pre-event state and produce little or no measurable consequence. Among measurable falls, plausible timing does not necessarily imply physically consistent motion. We further conduct a 15-participant, 20-condition human study in which participants describe the expected consequence from a single event-anchored frame and draw its trajectory. Human predictions favor the recorded future in aggregate while revealing genuine ambiguity among plausible continuations. Overall, physical foresight emerges as a sequence of distinct challenges: initiating a consequence, anchoring it in time, and realizing its motion.

8 pages, 2 figures, 6 tables

Time-series foundation models offer a unified approach to forecasting across heterogeneous domains. Textual context and auxiliary observations provide complementary information about temporal dynamics, yet reusable multimodal predictive representations remain underexplored. We introduce Pythia, a foundation world model that learns context-conditioned latent dynamics across datasets through a joint-embedding predictive architecture. A stop-gradient numerical reference guides contextual corrections to predicted future states. A separate probabilistic decoder then adapts to the frozen predictive representation and observed history, decoupling world-model pretraining from observation-space forecasting. On MUSE, Pythia-Tiny's normalized mean absolute scaled error (MASE) and weighted sum quantile loss (WSQL) are 0.6879 and 0.4269, reducing errors by 6.26% and 5.00% relative to the strongest model evaluated in the published MUSE leaderboard. Through a series of controlled experiments, we investigate how to design a time-series world model through shared pretraining and how joint-embedding predictive learning can incorporate multimodal information. The results support separating predictive representation learning from probabilistic readout and show complementary contributions from entity descriptions, events, and covariates.

Patient world models used for clinical trial simulation can agree on transition kernels and arm-specific risks, yet disagree on the fraction of patients harmed by switching treatment---the counterfactual quantity that matters for intervention-aware reasoning. We audit this reliability gap in a two-stage shared-response SCM: a categorical intermediate health state is followed by common terminal care. Under independent stages, the sharp harm interval has closed-form endpoints for at most three intermediate states, with an exactness boundary at four states. Declared dependence and response-mismatch budgets yield calibrated outer bounds when stage independence or complete mediation is relaxed; in a symmetric three-state model the entire sensitivity frontier is sharp, $[0,\min{1/2,1/3+(ρ+δ)/2}]$, and shows exactly how budgets erase the gain over endpoint-only bounds. Two eight-variable response LPs propagate interventional uncertainty for finite-sample audits. Exact witnesses verify attainability. On public clinical simulators (EpiCare; sepsis), native configurations show little resolved stage dependence and no additional joint-compatibility gain over pairwise transport---honest negative results for reliability claims. All experiments are locally reproducible; guarantees remain conditional on the stated causal model. The results provide a concrete protocol for deciding when a patient world model is safe to trust for counterfactual harm.

35 pages, 4 figures. Includes appendices

Real-time interactive video generation requires low-latency, streaming, and controllable rollout. Existing autoregressive (AR) diffusion distillation methods have achieved strong results in the chunk-wise 4-step regime by distilling bidirectional base models into few-step AR students, but they remain limited by coarse response granularity and non-negligible sampling latency. In this paper, we study a more aggressive setting: frame-wise autoregression with only 1--2 sampling steps. In this regime, we identify the initialization of a few-step AR student as the key bottleneck: existing strategies are either target-misaligned, incapable of few-step generation, or too costly to scale. We propose \textbf{Causal Forcing++}, a principled and scalable pipeline that uses \emph{causal consistency distillation} (causal CD) for few-step AR initialization. The core idea is that causal CD learns the same AR-conditional flow map as causal ODE distillation, but obtains supervision from a single online teacher ODE step between adjacent timesteps, avoiding the need to precompute and store full PF-ODE trajectories. This makes the initialization both more efficient and easier to optimize. The resulting pipeline, \ours, surpasses the SOTA 4-step chunk-wise Causal Forcing under the \textit{\textbf{frame-wise 2-step setting}} by 0.1 in VBench Total, 0.3 in VBench Quality, and 0.335 in VisionReward, while reducing first-frame latency by 50% and Stage 2 training cost by $\sim$$4\times$. We further extend the pipeline to action-conditioned world model generation in the spirit of Genie3. Project Page: https://github.com/thu-ml/Causal-Forcing and https://github.com/shengshu-ai/minWM .

Language-guided multi-camera tracking must preserve a target identity across unobserved gaps, where similar candidates and uncertain returns can make early associations unreliable. A wrong match can corrupt the history used to predict later observations and propagate identity errors across subsequent camera handoffs. We propose ReWorld-Track, a recursive event world model that carries association uncertainty into future predictions. Candidate matches and continued waiting define alternative target states, whose posterior probabilities are used to update a persistent recurrent belief. This representation preserves uncertainty about alternative trajectories through successive observations. This belief predicts the next camera, arrival time, and entry region, while appearance and language evidence guide association. By training across successive handoffs, the model learns to retain uncertainty that remains useful for later predictions and identity decisions. ReWorld-Track achieves HOTA scores of 65.19 on CityFlowV2 and 45.36 on MTMMC, with improved identity continuity across repeated handoffs. On MTMMC, its structured posterior update gains 0.50 HOTA points over a similarly sized generic updater and 0.94 points over fixed-moment soft association, raising next-camera accuracy from 86.03% to 87.41% and reducing median arrival-time error from 0.78 s to 0.71 s for subsequent target returns.

41 pages, 8 figures. Corrected an author name; scientific content unchanged

Latent world models plan by rolling a frozen predictor forward under candidate action sequences and ranking the candidates by the latent distance between their imagined end state and the goal. However, this ranking breaks down when the goal lies several plans away, because the latent distance measures how closely an end state resembles the goal rather than how far it remains from reaching it. To address this, we propose TEMPO, a temporal-distance planning objective that leaves the world model untouched, learns only from the demonstrations already used to train it, and adds negligible cost to the planner's search. TEMPO learns a small map of the frozen latent in which the distance between two states of an episode reflects the number of environment steps between them, and blends this distance into the planner's cost. It requires no rewards, policies or success labels and, being a cost rather than a model, applies to frozen world models with one latent vector per state that plan by a latent distance. We evaluate TEMPO on eleven simulated environments (e.g., maze navigation, tabletop pushing, robotic arm control and three-dimensional manipulation) with the LeWM and PLDM planners. With a small MLP that adds at most 0.3% to a plan's arithmetic, TEMPO improves both planners at every goal distance, including the one-plan setting of their evaluations, raises LeWM from 36% to 99% on TwoRoom three plans from the goal, and remains competitive on a broad range of 2D and 3D navigation, reaching and manipulation tasks.

High-fidelity atomistic evolution over long timescales requires more than observing the current crystal configuration. Instantaneous atomistic snapshots are often incomplete: locally similar configurations can correspond to different hidden dynamical contexts, future event preferences, and waiting-time scales. We argue that this snapshot ambiguity makes long-horizon atomistic evolution fundamentally a memory-based world-state restoration problem. To address this, we introduce AtomWorld-Mem, a memory-restored atomistic world model that recovers the latent world state missing from instantaneous crystal snapshots. AtomWorld-Mem treats the evolving alloy as an AtomWorld: spatial encoders write multi-scale atomistic keyframes from dense local topology and sparse long-range defect context, while short-term event memory and long-term structural memory integrate these keyframes across time to restore a future-predictive evolutionary state. The restored state is used to prioritize legal vacancy-mediated events under single-event Kinetic Monte Carlo (KMC) constraints, while event legality, physical execution, and residence-time updates remain governed by the underlying simulator. Empirically, AtomWorld-Mem improves long-horizon atomistic progress under fixed microscopic event budgets while maintaining high-fidelity evolution across energetic, structural, and vacancy-transport observables. It further transfers zero-shot across diverse unseen alloy-temperature AtomWorlds, suggesting that the learned memory-restoration mechanism captures reusable principles of hidden-state inference rather than a system-specific local energy heuristic. These results position memory-restored world-state modeling as a promising route toward efficient, physically grounded, and transferable atomistic evolution.

A coding agent may safely modify one repository while silently breaking downstream services, libraries, or datastores that depend on it. Exhaustively running integration tests after every agent action is impractical, so the agent must predict these failures before executing them. Existing software world models predict the agent's own observations, while static change-impact analysis only identifies where a change may propagate. We instead introduce the Software World Model (SWM), which models the broader system affected by a code change and predicts its blast set: the components that the change will break. SWM follows three stages: explore, learn, and act. Explore executes candidate changes from restored system states, prioritizing regions where observed failures contradict the dependency graph. Learn fine-tunes a language model on these execution outcomes to predict downstream breakage. Act converts sampled predictions into per-consumer break probabilities for change ranking, proactive migration, and deciding when another execution is worth its cost. On held-out synthetic systems, SWM improves blast-set F1 from 0.431 for static reachability to $0.571\pm0.037$, more than halves ranking regret, and improves migration return at all nine evaluation checkpoints. The two methods are complementary: reachability is stronger on dependencies represented in the graph, while SWM recovers failures caused by couplings the graph misses. Experiments on held-out real libraries further show that predicting structured failure outcomes, rather than only scalar risk, is important for downstream decision quality.

29 pages, 9 figures, 13 tables

Pretrained world models can generate diverse environments, yet users often want to explore a particular scene specified by their own video. This requires learning the scene's visual identity while retaining the quality of action-conditioned generation. We introduce Personalized World Models (PWM), a framework for customizing interactive world models from short scene videos through online reinforcement learning. In PWM, the support trajectory and its associated controls provide reward feedback on continuations sampled from the current policy. In the GRPO instantiation, group-relative optimization updates a compact LoRA adapter using a unified reward for scene appearance, visual continuity, and motion, while base-policy anchoring regularizes changes to the pretrained generation prior of a frozen Yume-5B backbone. The same adaptation procedure is applied across real and rendered environments. We also instantiate PWM with DiffusionNFT as an alternative reward-guided optimization method for learning the scene-specific adapter. We also introduce PWM-Bench, comprising 150 customization tasks across Indoor, Outdoor, and Gaming, with paired evaluation on held-out continuations. The GRPO and DiffusionNFT instantiations of PWM improve customization over native Yume in 71.3% and 65.3% of the evaluated scenes, respectively, with positive mean gains across all three domains. For the GRPO instantiation, matched SFT comparisons further demonstrate higher mean customization gains and better mean image-quality scores in every domain, while retaining frame-level visual quality close to the pretrained model.

Code: https://github.com/AIGeeksGroup/PWM, Website: https://aigeeksgroup.github.io/PWM

Urban navigation requires embodied agents to pursue long-horizon goals through local decisions based on egocentric observations. However, existing agentic navigation methods often struggle to translate distant goals into coherent local decisions in large-scale physical environments. Their reliance on linguistic reasoning over transient observations or limited history constrains anticipation of the consequences of actions and future conditions, despite the importance of such foresight for navigating long and complex urban routes. To bridge this gap, we propose PreAct-Nav, an agentic navigation framework that equips frozen policies with anticipatory reasoning for robust urban navigation. Our central idea is to anchor local decisions in persistent medium-horizon subgoals, assess the consequences of predicted actions before execution, and continually update the reasoning context using actual outcomes. At its core, a navigation memory module maintains the active subgoal and relevant experience across decisions, translating distant goals into actionable intermediate objectives. We further introduce a predictive world sandbox that uses an action-conditioned world model (AC-WM) to forecast world dynamics conditioned on candidate movements. A vision-language model (VLM) reasoner interprets these predictions under the current subgoal to retain or revise actions. After execution, real observations are used to assess outcomes, correct inconsistent assumptions, and update memory to continue or reformulate the subgoal. Extensive evaluations demonstrate that the proposed PreAct-Nav improves action selection through memory updates and visual prediction, with more pronounced gains on longer routes and routes with more turns.

Robotic controllers increasingly rely on analytical models, simulators, cost-query interfaces, and learned world models. However, physical deployment can deviate from nominal assumptions, and additional disturbances may arise even when the model itself is accurate. In control systems, disturbance observers (DOB) are widely used to estimate such unmeasured effects from nominal models and measured feedback. Classical DOB formulations are generally built around explicit plant models. This paper develops the sampling-based disturbance observer (SDOB), extending the DOB principle to a broader range of models, including simulators and learned world models, through state-rollout or cost-query interfaces. SDOB separates two observable channels: state-effect disturbances, for which the feedback state differs from its prediction, and cost disturbances, for which the same query state receives different costs as the perceived environment changes. Diverse simulation and real-robot experiments across traditional and learned models demonstrate the effectiveness of SDOB in compensating for sim-to-real mismatch and improving control performance.

15 pages, 14 figures, 7 tables. Project website: https://sampling-based-dob.github.io/

Open-ended discovery requires learning from repeated attempts while continuing to explore directions whose value is not yet apparent. Search with a frozen large language model (LLM) can reuse previous solutions in context, but cannot update the model from its successes and failures on the test problem. Reinforcement learning (RL) enables such adaptation; however, strongly favoring high-reward trajectories may suppress low-reward yet potentially promising directions too early. We introduce CURIO, a curiosity-driven test-time learning framework that complements task feedback with an Intrinsic Curiosity World Model (ICWM). The ICWM learns transitions in the policy's hidden-state representation and supplies prediction-error bonuses at sampled tokens outside the policy's top-k choices. Epoch normalization and an annealed weight regulate their contribution to the policy update. On six mathematical discovery tasks and single-cell denoising with Qwen3 backbones from 8B to 235B, three-run means improve over a matched task-only RL control on five mathematical objectives, match the best reported performance on Circle Packing, and improve denoising Score and mean squared error (MSE) on both held-out corpora at every tested scale. Relative gains reach 18.3% on Hadamard and 10.8% on denoising Score. Code-diversity measurements show greater structural variation among generated programs, supporting curiosity as a complementary exploration signal for learning in open-ended discovery.

Patient world models are increasingly proposed for longitudinal prediction, intervention-aware reasoning, and clinical-trial simulation. Causal or clinical intervention validity is distinct from predictive generalization and reliability; before making stronger claims, the underlying predictive state should generalize across patients, survive realistic shifts and missing observations, and expose failure through meaningful reliability signals. We evaluate these prerequisites in a deliberately narrow setting: short-horizon digital-biomarker forecasting from PhysioNet GaitPDB, comprising 165 participants, 306 recordings, and 51,129 context-future pairs. Using persistence, ridge, MLP, GRU, Transformer, and a compact JEPA-style predictor, we build an evaluation ladder that progressively removes raw temporal overlap, same-recording familiarity, and same-patient familiarity before testing unseen-patient generalization. For GRU, NMSE rises from 0.1227 under random-window splitting to 0.1393 after eliminating raw train-test overlap and to 0.1961 under patient holdout. Among 54 participants with repeated recordings, exposure to a different recording from the same patient improves GRU NMSE from 0.2177 to 0.1556, while a recording-excluded identity hypothesis is not supported at the participant level. Under participant-held-out evaluation, MLP and Transformer are statistically indistinguishable. Study shift, a four-times-longer prediction gap, and partial observation further degrade performance; under 50% temporal masking, Transformer NMSE rises to 0.611 while MC-dropout predictive variance falls. We do not claim a longitudinal or intervention-aware simulator. Instead, the results support a prerequisite evaluation stack of patient separation, repeated-measure controls, shift, missingness, and uncertainty validation before stronger patient-world-model claims are trusted.

13 pages, 7 figures. Selected for oral presentation at the NeurIPS 2026 Workshop on World Models for High-Stakes Health (WMHS) Replication Package: https://doi.org/10.17605/OSF.IO/AUZVR

Latent world models enable robots to plan by predicting the consequences of actions. Planning long tasks with control-rate actions requires many prediction steps, which enlarges the search space and accumulates error. Skill-level actions shorten these sequences, but a symbolic skill vocabulary requires domain knowledge and labeled demonstrations. We construct skill-level actions from the inputs of a flow-matching policy trained on demonstrations segmented into complete skills. The policy maps a noise seed and an observation, optionally with a code or label, to a complete skill execution, so one execution is one world-model transition. On this mechanism we propose four action abstractions with increasing task knowledge: a compressed seed, two discrete codes learned from the demonstrations, and a symbolic label. We evaluate them with a common world-model training procedure and planning framework on simulated block rearrangement tasks that require up to 14 sequential skills. The symbolic label succeeds in over 90% of the tasks that require up to six skills and degrades beyond. Without any label, an object-centric learned code matches it on single-skill tasks and retains half to three quarters of its success on tasks of two to five skills. Ablations attribute much of the label's advantage to its planner knowing which actions are applicable, rather than to the label itself. Beyond six skills the search, not the world model, limits success. Project page: https://andreumatoses.github.io/research/flow-skill-wm

9 pages, 7 figures, 1 table. Submitted to IEEE ICRA 2027. Project page: https://andreumatoses.github.io/research/flow-skill-wm

Electrocardiogram (ECG)-based models have achieved strong performance in diagnostic tasks, yet they remain limited in modeling how cardiac dynamics evolve under external interventions. In particular, existing approaches focus primarily on static prediction and lack mechanisms to capture ECG variations under different pharmacological conditions. In this work, we propose an ECG World Model for action-conditioned predictive simulation of cardiac electrophysiology. Moving beyond disjoint pipelines, our framework features a principled integration of physiological ordinary differential equation (ODE) priors into latent diffusion dynamics via energy regularization. This structural constraint enables the synthesis of physiologically plausible post-intervention ECG trajectories while effectively mitigating generative hallucinations. Building on this simulation process, we introduce an uncertainty-aware evaluation strategy that leverages the stochasticity of diffusion sampling to characterize both the expected clinical risk and its variability, allowing a more reliable comparative assessment of candidate interventions. We evaluate our method across diverse settings, including controlled drug-response scenarios and real-world clinical records. Beyond standard waveform metrics, experimental results demonstrate improved risk calibration and strong alignment with expert-informed treatment preferences. These results establish our approach as a robust foundation for safe and intervention-aware clinical decision support.

As video diffusion models (VDMs) advance toward world models, a key question arises: do they truly understand causality, or merely overfit to statistical temporal patterns? Existing benchmarks mostly rely on synthetic data, limiting real-world generalization due to the sim-to-real gap. We present YoCausal, a two-level benchmark inspired by the Violation of Expectation (VoE) paradigm from cognitive science. By temporally reversing real-world videos at zero cost as natural counterfactual samples, YoCausal establishes an arbitrarily extensible evaluation protocol. Level 1 introduces the Reverse Surprise Index (RSI), quantifying arrow-of-time perception via denoising loss. Level 2 introduces the Causality Cognition Index (CCI), which leverages a VLM to stratify datasets into causal and non-causal subsets, disentangling genuine causal reasoning from temporal bias. Evaluation of 13 state-of-the-art VDMs reveals that perceiving the arrow of time does not imply understanding causality, and a significant gap persists relative to human-level causal cognition. Project page: https://www.youzhexie.me/papers/YoCausal/

Project page: https://www.youzhexie.me/papers/YoCausal/

Vision-Language-Action (VLA) models generalize across static manipulation but fail when objects move during task execution. They map the current observation to an action and assume the scene is stationary between observation and execution, so at any non-trivial object speed the resulting latency exceeds the time available to grasp. We close this gap with AHEAD (Anticipatory Horizon Extrapolation with Adaptive Dynamics), a predict-then-act wrapper that augments a frozen VLA with a motion-aware latent world model. A small world model trained on manipulation video forecasts future patch tokens in the VLA's feature space, conditioned on per-token velocity and acceleration from optical flow. A language-and-motion saliency mask concentrates prediction on task-relevant patches, and the model rolls forward for an adaptive horizon, halting when prediction uncertainty crosses a threshold. The frozen action decoder then receives the predicted future tokens in place of the current ones. AHEAD adds 4.9M parameters to a frozen 7B OpenVLA and reaches 79 to 97% success across 20 dynamic simulation scenarios where the strongest baseline reaches 31 to 58%. On a physical UFactory xArm 7, AHEAD succeeds on 29/30 to 30/30 on three conveyor and rolling-ball tasks, 23/30 on paddle interception, and 19/30 on projectile catching where every baseline scores 0/30.

Accepted at the 10th Conference on Robot Learning (CoRL 2026). 31 pages, 7 figures, 19 tables

Robots need to anticipate how their actions will change the world, since manipulation success hinges on the resulting contacts and object motions. However, existing Vision-Language-Action (VLA) policies that predict future observations from shared features leave the forecast decoupled from the actions the policy will actually execute, and impose no physical constraints on how the scene may evolve. We introduce ForeAct3D, a framework for policy-grounded future world modeling within VLA policies. Learnable geometric queries decode depth, semantic segmentation, and camera pose from the policy representation into current and future semantic 3D scene states, and the future queries are conditioned on the policy-generated action chunk to ground the forecast in the planned interaction. A physical-consistency closure relates the two states through background staticity and instance-level rigidity, and anchors the wrist-camera pose to end-effector kinematics. These objectives shape the shared representation used for action generation during training, and no future prediction is required at inference. Without robot pretraining, ForeAct3D achieves 98.3% average success on LIBERO and an average task length of 3.73 on CALVIN, outperforming its base policy on every suite. Ablations show that semantic 3D supervision, physical consistency, and action conditioning each improve manipulation performance, and that action conditioning substantially improves future object localization. Real-world experiments on spatial placement, object insertion, and sequential manipulation further raise average success from 6.7% to 37.8% over the base policy. The project page and code are available at https://github.com/anthonytao80-crypto/ForeAct3D.

Joint-embedding predictive architectures (JEPAs) for world modeling train an encoder so a predictor maps a current embedding and action to the next frame's embedding, always from a single rendered frame. This has a structural blind spot: a renderer without motion blur draws a scene from configuration alone, so a single-frame embedding carries no velocity information, for any encoder, including the official released LeWM weights. We confirm this on official checkpoints across four real benchmarks (PushT, Reacher, Cube, TwoRoom): every linear velocity probe sits at or below chance while position probes reach R^2 about 0.95. We introduce RateIdent, a three-stage diagnostic protocol, and TI-JEPA, a lightweight fix splitting the latent into a pose code and an explicit finite-difference motion code, predicted jointly. Across three physically grounded environments, TI-JEPA gives a significant, seed-robust gain on a stop-at-goal planning task over a matched-memory baseline, e.g. 55% lower final distance on Pendulum (p=3.2x10^-10) and 64% on CartPole (p=5.1x10^-15). We reproduce this at official ViT-Tiny plus AdaLN-transformer scale, then push the same recipe onto real dm_control Reacher photographs trained from scratch, where TI-JEPA's branch separation exceeds the memory-having baseline's by roughly 38x, the paper's largest margin. Against a same-footprint recurrent RSSM-style predictor, TI-JEPA matches or beats its rollout accuracy on two of three environments, stays separately probeable for pose and motion, and wins outright on the most coupled one. A checkable formal argument and six evaluated environments show single-frame targets are the wrong object to predict when velocity matters, and a small, interpretable structural change fixes it with no privileged supervision. Code, checkpoints, and the project page are linked below the title.

32 pages, 17 figures, 12 tables. Code, model checkpoints, and project page are available via links in the paper

We introduce DreamFormer, a model-based agent that acquires language-conditioned, multi-task skills by imitating expert demonstrations within the latent imagination of a learned world model. DreamFormer first learns a task-agnostic Transformer world model from unstructured play data, then acquires task-specific behaviors by optimizing an intrinsic reward that aligns agent-generated rollouts with expert demonstrations in latent space. Since the policy is trained on-policy inside imagination, it is exposed to its own errors during training, mitigating the covariate shift inherent to offline behavioral cloning. To make long-horizon imagination affordable, DreamFormer encodes a high-resolution multi-view robotic observation into a single input token, avoiding both spatial downsampling and the multi-token representations used by prior Transformer world models. On the long-horizon CALVIN benchmark, DreamFormer outperforms LUMOS, the comparable model-based agent, on single-environment evaluation (2.52 vs 2.34 average tasks completed per chain of five) while keeping imagination rollouts tractable. Against HULC, the behavior cloning baseline, it nearly doubles performance on zero-shot transfer to an unseen environment (1.30 vs 0.67), indicating that dynamics learned by the world model transfer more readily than a directly cloned policy. This is consistent with the role attributed to internal models in biological agents, where a model of environment dynamics supports behavior in situations not previously encountered.

submitted to IEEE Transactions on Cognitive and Developmental Systems (TCDS)

Latent world models learn to predict observed transitions, yet low prediction error alone does not guarantee reliable planning. Inspired by self tickling experiments in neuroscience showing that disrupting motor sensory correspondence increases prediction mismatch, we examine whether learned world models preserve an analogous action consequence correspondence.The results show nearby alternatives can receive lower prediction errors despite producing physical outcomes farther from the recorded target. With that future treated as a goal, this reveals a concrete prediction planning mismatch: the model assigns a lower cost to an action that achieves the target less accurately. To mitigate this gap, we introduce Action Consequence Alignment (ACA), a training objective that complements forward prediction by penalizing the prediction error advantage of locally searched alternatives over factual actions without additional model components or environment interactions during training. The same principle can also guide additional data collection for self improvement. We demonstrate that across diverse environments and evaluation settings, ACA improves planning performance and reduces real goal error, while ACA guided data collection outperforms random local sampling. These results support action consequence alignment as a practical principle for bridging predictive learning and reliable planning.

Testing deep reinforcement learning (DRL) agents in cyber-physical systems aims to uncover diverse failures before deployment, but each execution can be expensive. Surrogate-assisted testing reduces this cost by learning to predict which test configurations are likely to fail. Prior surrogates treat the system as a black box and predict pass or fail outcomes directly; we instead model how a test unfolds and estimate failure from an imagined episode. We introduce DreamTest, a world-model surrogate for testing DRL agents. DreamTest adapts a recurrent state-space model to learn agent behaviour and environment dynamics from the agent's training log. Given a candidate configuration, imagined rollouts produce a failure score that guides search without executing every candidate in a simulator or real system. We evaluate DreamTest for failure prediction, test generation, and failure diversity on Parking, Humanoid, and DonkeyCar. Mean area under the precision-recall curve (AUPRC) exceeds the strongest baseline by 97%, 12%, and 39%, respectively, and gains on five out-of-distribution test sets reach 145%, 29%, and 44%. Under the same simulator-validation budget, the best "DreamTest + search" combinations find 29%, 22%, and 79% more novel failures on average. Across clusterings with k = 2-40, failures generated with DreamTest cover the most behavioural clusters for almost all k, indicating that DreamTest consistently discovers behaviourally diverse failures.

Predicting cellular responses to genetic perturbations is a central capability for virtual cells and a key step toward computational modeling of biological interventions. Most existing models directly map an unperturbed molecular profile and perturba- tion to the resulting observation without explicitly representing the latent cellular transition induced by the intervention. We introduce a mechanism-centric virtual cell world model that represents cellular state as a set of Latent Mechanism Units (LMUs) and treats genetic perturbations as actions on these latent states. Each LMU combines a reusable identity grounded in multimodal biological evidence with an observation-specific state, allowing a perturbation to induce mechanism- specific stochastic transitions before decoding the resulting transcriptional response. We train VCLMU through two-stage pretraining, first on around 200K pseudo-bulk perturbation profiles and then on gene-aligned single-cell perturbation data. Across six perturbation-disjoint benchmarks, VCLMU consistently improves perturbation- specific response recovery over strong baselines while maintaining competitive global response accuracy. We further analyze learned LMUs through enrichment between perturbation responses and LMU gene sets and show that they capture structured biological response programs. These results support mechanism-level latent state transition as a useful formulation for virtual cell models that aim to predict and interpret cellular responses to biological interventions.

This paper was accepted to the NeurIPS 2026 workshop on AI for drug discovery

World models learn to predict how their environment will evolve, making them an important foundation for general-purpose robotic control. Yet world action models depend on camera inputs whose manipulation can corrupt the visual representations used across tasks and action policies. Existing attacks on these models optimize against the victim's actions or predicted futures and therefore require access to target-model outputs. In this paper, we propose an attack, TAPDreamer, against world action models that instead uses a public encoder alone to construct a fixed local perturbation that transfers across tasks and action architectures. TAPDreamer requires no target-policy queries. Our key insight is that interactions between patch-induced changes in attention weights and value vectors broadcast a nearly identical representation shift far beyond the patch footprint, and this shift remains stable across task observations. Guided by this insight, TAPDreamer uses six frames from one source task to maximize the global L1 distance between clean and patched encoder representations. In closed-loop evaluation, one frozen patch per benchmark, covering about 6.5% of the input, reduces FastWAM's success rate from 97.7% to 0.0% across 40 LIBERO tasks and from 90.8% to 0.0% across 50 RoboTwin tasks; matched random patches retain 81.5% and 79.2% success. The same patches reduce success to 2.1% and 0.8% on two DreamWAM configurations and to 10.0% on Motus. These results show that protecting downstream action generation alone is insufficient: defenses for world action models must also secure shared visual encoders against persistent local perturbations.

Project Page: https://tapdreamer.github.io

World Action Models (WAMs) couple action generation with future visual prediction for robotic manipulation. However, completing the joint video-action denoising process at each replanning cycle incurs substantial latency, delaying action updates and limiting closed-loop responsiveness. We present Rolling-WAM, a formulation that distributes joint denoising across successive replanning cycles. Our method maintains a sliding window of video-action chunks at staggered noise levels. At each step, a rolling noise schedule fully denoises the imminent action chunk for execution, while partially refining farther-future chunks. As the window advances with new camera observations, the retained future chunks continue their denoising process. This distributes the computational cost over time while carrying an evolving visual-action context across chunk boundaries. Evaluations on LIBERO, RoboTwin, and a real-world Unitree G1 humanoid show that Rolling-WAM achieves competitive manipulation performance. By removing the need to denoise the entire prediction horizon from scratch, it delivers a 4.5x steady-state replanning speedup over standard joint WAMs.

10 pages, 7 figures, 5 tables. Under review. Project page: https://rolling-wam.github.io/

World Action Models (WAMs) incorporate visual representations from video generation backbones to guide action prediction. Recent efficient WAMs adopt Mixture-of-Transformers (MoT) architectures and compute video representations once for reuse by the action expert. However, intra-expert iteration (\ie, multi-step action denoising) and inter-expert waiting (\ie, sequential execution of the video and action experts) still limit inference efficiency. To this end, we present RealtimeWAM, an extremely efficient WAM variant with one-step action generation and asynchronous inference, addressing these two bottlenecks. To reduce intra-expert iteration, we propose Teacher-Anchored Consistency Distillation (TACD) to address a local-global error gap: low local consistency error alone does not guarantee accurate final actions. TACD supplements local consistency with explicit supervision from the frozen teacher's multi-step rollout endpoint, enabling accurate one-step action generation. Additionally, we propose Cross-Expert Wavefront Pipelining (CEWP) to eliminate unnecessary expert-level waiting. It overlaps the two experts through block-wise sharing of the video KV cache, synchronizing only immediately before the corresponding action attention consumes it. Extensive experiments across diverse benchmarks (\eg, LIBERO, LIBERO-Plus and RoboTwin) and model variants (\eg, Fast-WAM and Faster-WAM) demonstrate the superiority of RealtimeWAM. Notably, RealtimeWAM maintains near-lossless performance (\ie, $<1%$ drop) across these benchmarks while delivering significant end-to-end speedup (\eg, $\sim25\times$ on H100). Our code and checkpoints are available via this \href{https://github.com/ModelTC/LightX2V/tree/main/examples/realtimewam}{link}.

The code and checkpoints are available at $\href{https://github.com/ModelTC/LightX2V/tree/main/examples/realtimewam}{\text{this https URL}}$

World action models (WAMs) have emerged as a promising paradigm for robotic manipulation. They act by first predicting how a task should be performed and then decoding the actions from that future. However, the remaining actions are invalid once execution drifts from the prediction. Simply replanning from the already out of distribution state rarely restores what the task still requires; existing execution monitors decide when to stop, but not what to restore. We observe that the answer is already in hand: the future the WAM predicted before acting depicts exactly the states it intended to pass through. We introduce FAVOR (Future Anchored Verification and Online Recovery), a lightweight framework that keeps these predicted frames as anchors and uses them for verification and recovery. An Anchor Verifier compares each observation with its anchor, together with the executed actions, to flag deviations that break the task. Anchor-Guided Recovery uses a vision-language model to turn the flagged anchor into a short corrective instruction. Under strengthened instruction guidance, the WAM executes this instruction to return to the intended future. It then resumes the task. FAVOR raises the task success of the base WAM from 97.85% to 98.10% on LIBERO and from 72.60% to 72.98% on LIBERO-Plus without modifying the policy.

18 pages, 5 figures, 3 tables

Instruction following is central to language-conditioned robot policies: language should determine what to do when the same scene permits multiple valid actions. Yet successful execution alone cannot establish whether a policy follows the instruction or infers the task from the scene. We study this ambiguity through scene-preserving instruction interventions, using valid target substitutions, arbitrary nouns, and unrelated sentences while holding the scene fixed. We evaluate vision-language-action (VLA) policies and world-action models (WAMs) in simulation and in real-world experiments. Our analysis addresses three questions: (a) Does task success imply instruction following? When instructions request a different visible object, all evaluated policies predominantly approach and pick up the incorrect original target associated with the scene. (b) Is this failure caused by language insensitivity? Instruction perturbations affect task performance. A layerwise action lens shows intermediate action predictions respond to these perturbations. Linear probes accurately recover instructed targets, indicating modified instructions are encoded despite rarely determining target selection. (c) Why does encoded language fail to control action? Attention analysis indicates weak instruction-token contributions to action generation. Target-token attention can remain focused on the original object, revealing a mismatch between target encoding and visual grounding. UMAP and shared non-negative matrix factorization show target information remains accessible within representations increasingly organized by scene identity. Our findings expose a grounding gap concealed by nominal success and provide a diagnostic framework. They further establish a concrete criterion for progress: policies should reliably follow valid changes in user intent, even when they conflict with scene-favored behavior.

World--action models (WAMs) promise a unified model that predicts action-conditioned futures, generates feasible actions, and supports planning in imagination. However, existing joint video--action models often use computationally heavy, fixed-horizon backbones ill-suited to streaming inference and stable long-horizon open-loop rollouts. We introduce FLEX-WAM, a Flexible and Efficient Block-Causal World--Action Model for unified simulation and policy inference. FLEX-WAM supports variable-length contexts and non-causal prediction horizons, as well as infinite autoregressive generation frame by frame or block by block. Its block-causal, KV-cacheable architecture combines axial attention and blockwise diffusion forcing to enable efficient real-time rollout and deployment-time latency--throughput tradeoffs without retraining. Joint training can nevertheless produce plausible futures that weakly respond to commanded actions. We address this failure mode by balancing state and action flow-matching gradient contributions across the state--action diffusion-noise grid and regulating world-model sampling using Forward-Dynamics (FD) elasticity, an efficient training-time proxy for action responsiveness. Across simulated and real-world datasets, FLEX-WAM achieves superior multi-step prediction quality and latency while producing stable joint state--action rollouts for thousands of steps. As a joint action proposer and simulator within MCTS, it solves long-horizon PushT and all five OGBench Puzzle-4x4 tasks entirely in imagination. On a bimanual OpenArm-based robot, a single checkpoint jointly serves as a play policy and expected-outcome predictor, enabling real-time identification and collection of model--reality mismatches for future self-improvement.

FastWAM-style world action models enable efficient action-only inference, but generalize poorly under visual distribution shifts. Their reconstruction-oriented representations emphasize appearance-specific details, limiting generalization to unseen scenes and objects. Without observation history, the model also lacks temporal evidence for robustly identifying task-relevant state changes and motion in unfamiliar visual conditions. To address these limitations, we present the Causal Semantic World Action Model (CSWAM), which augments FastWAM with a causal semantic expert built on V-JEPA 2.1. V-JEPA provides temporally grounded representations of semantic state changes and motion with less dependence on appearance-specific details. The expert learns their future evolution from a sparse history of current and past observations and shares the history-derived context with both the video and action streams through causal attention. At inference, CSWAM conditions action denoising on the current video state and observed semantic history, retaining efficient action-only inference. We conduct simulation and real-robot experiments to evaluate generalization under distribution shifts. With embodied pretraining, CSWAM raises Randomized success on RoboTwin 2.0 Clean-to-Randomized transfer from 10.16% to 45.18%, a gain of 35.02 percentage points over FastWAM. Across two real-robot tasks and three OOD difficulty levels, CSWAM improves average success over FastWAM by 42.5 percentage points, from 27.5% to 70.0%.

World Action Models (WAMs) emerge as a promising foundation for policy refinement by predicting the consequences of sampled actions. However, visually plausible predictions can mislead policy refinement if they fail to reflect the input actions. To address this mismatch, we introduce $R^2$-WAM, a two-stage repair-and-reject post-training framework that first improves the consistency of predicted futures with input actions, then uses these futures to select inferior action samples for negative fine-tuning. The repair stage grounds imagination in observed robot behavior through a kinematic alignment score that measures agreement between predicted and demonstrated motion, enabling the predicted video to faithfully reflect its input actions. Using the repaired video model, the rejection stage compares imagined outcomes of sampled and demonstrated actions, selectively applying negative fine-tuning to samples whose predicted task progress falls below the demonstrated reference by a prescribed margin. Together, the two stages extend video prediction from representation learning to consequence-based policy refinement without additional environment interaction or changes to the inference procedure. $R^2$-WAM achieves 93.8% average success on RoboTwin 2.0 across clean and randomized settings. On the long-horizon real-world Fold Shirt task, it achieves 87.5% average success, compared with 0% for Fast-WAM.

Vision-language-action (VLA) and world-action models (WAMs) often degrade under out-of-distribution task variations despite retaining partial task capability. To recover such capability, we propose RoboIRS, an inference-time internal representation steering method that uses successful and failed rollouts to train linear classifiers, select outcome-relevant intervention locations, and derive task-specific steering directions without updating policy parameters. On 15 simulation tasks with a frozen $π0.5$ policy, RoboIRS improves the average success rate from 44.4% to 66.2%, outperforming alternative inference-time intervention baselines while adding little inference time. We further validate RoboIRS on real-robot manipulation using the same $π0.5$ policy and demonstrate its applicability to a world-action model Cosmos Policy, where the average success rate improves from 35.4% to 55.4%. These results show that directly steering internal robot-policy representations can improve the performance of robot policies at inference time. Project website is available at https://rollingoat.github.io/roboirs/.

World-action models (WAMs) jointly generate future video and robot actions through iterative diffusion, achieving strong performance on manipulation benchmarks but requiring tens of denoising steps, a cost that precludes real-time control. Step distillation has emerged as the natural remedy, but off-the-shelf methods break down in the joint video-action setting because video and action streams use different SNR-shifted noise schedules and reach training with substantially different marginal noise distributions, an asymmetry that single-modality distillation methods cannot accommodate. We introduce \textbf{Flash-WAM}, a modality-aware step-distillation framework inspired by consistency distillation that selects the consistency function for each modality to match its noise regime: a linear-gradient-scaling parametrization for the action stream's low-noise regime, paired with a variance-preserving parametrization for the video stream's high-noise regime, grounded in a structural analysis of the consistency-function family that characterizes the achievable gradient scaling under the consistency boundary condition. Instantiated on LingBot-VA, Flash-WAM compresses inference to a single step in each modality. On RoboTwin 2.0, this reduces per-chunk latency from $8.1$ seconds to $348$ ms on NVIDIA L40S, a $23{\times}$ speedup that enables real-time inference. Flash-WAM preserves task success on simulation benchmarks ($85.5%$ RoboTwin 2.0, $95.7%$ LIBERO) and substantially recovers real-world performance ($60%$ average on a Unitree G1 humanoid robot), while naive consistency distillation drops to $24%$ at the same step budget.

18 pages, 4 figures, 9 tables. Project page: https://flashwam.github.io

World-action models (WAMs) augment robot policies with action-conditioned environment dynamics, yet existing approaches largely rely on future observation reconstruction or generic latent prediction and lack structured, control-oriented world representations tightly coupled with action generation. We introduce Magic-W0, a world-action foundation model that jointly models structured physical state evolution and continuous actions. Magic-W0 represents interaction as a Structured World Transition consisting of Current State, Transition, and Future State. Current State combines vision-language context with Current 3D Geometry; Transition is represented by 3D Motion capturing action-induced three-dimensional changes; and Future State is represented by Future Semantics describing task-relevant outcomes. To couple prediction and control, we propose a layer-aligned world-action interaction architecture in which evolving action hypotheses condition world-transition prediction, while predicted world representations continuously inform action generation. Magic-W0 is pre-trained on large-scale egocentric human manipulation, UMI, real-robot, and simulation data, with latent supervision for geometry, 3D motion, and future semantics from pre-trained visual models. Inference-time interventions show that structured world representations respond systematically to changes in candidate actions and that action-related information propagates through shared 3D representations into future semantic predictions. On RoboDojo-Sim, Magic-W0 achieves an average Score of 27.10, the highest among the compared WAMs. Across multiple real-robot tasks, it also demonstrates strong downstream performance after fine-tuning with limited downstream data, supporting generalization and rapid adaptation.

29 pages, 15 figures, 7 tables. Project page: https://embodied.magiclab.top/works/wam/magic-w0/index.html; Code: https://github.com/MagiclabRobotics/Magic-W0

Vision-language-action models benefit from the understanding and reasoning capabilities of pretrained vision-language models, but action-only supervision provides limited grounding in world dynamics. Conversely, world-action models inherit spatiotemporal priors from video generation models, yet remain limited in semantic understanding and reasoning under distribution shifts. We introduce UniWAM, a unified architecture that integrates a physical reasoner, a world generator, and an action predictor to jointly learn semantic understanding of the physical world, visual generation, and action prediction. To ensure the quality of the training data, we developed a rigorous data cleaning and annotation pipeline for both human egocentric data and robot data. To adapt the vision-language component to embodied tasks while preserving its inherited language capabilities, we represent low-level actions in natural language and introduce a pre-training recipe that assigns complementary supervision from visual question answering (VQA) data, human egocentric data, and robot demonstrations to the appropriate model components. During post-training, future visual noise augmentation reduces reliance on precise future predictions, while history-conditioned flow matching uses encoded action history to initialize action generation. Together, these designs significantly reduce denoising steps while maintaining performance. UniWAM achieves state-of-the-art (SOTA) performance across multiple evaluations, including in-distribution performance, robustness, generalization, instruction following, and long-horizon task execution. Furthermore, we uncover a log-linear scaling law of unified human-robot co-training, demonstrating the effectiveness of large-scale pre-training on a mixture of human and robot data.

Vision-language-action models (VLAs) inherit strong semantic grounding from pretrained vision-language backbones but are typically optimized for predicting actions rather than future observations. They can see and act, but they do not imagine the future before acting. World action models (WAMs) built on video diffusion backbones can imagine but treat language as frozen conditioning on a continuous latent space. Unified models bring these modalities into one architecture, but they either decode autoregressively, one token at a time, or keep video continuous with an auxiliary action head. In this work, we present SUAVE, a Single vocabulary Unified Action-Video modEl in which a masked diffusion transformer generates video and actions conditioned on language, with all three modalities represented as discrete tokens in a shared sequence. Choosing which tokens to mask at inference turns the same network into a world model, a robot policy, or a video-action model. For action-free co-training, the action positions of unlabeled video are filled with mask tokens and excluded from the loss. Simulation and real-world experiments demonstrate two findings. First, a single SUAVE model predicts long-horizon video and acts as a policy, competitive with dedicated world models and specialized action policies on static and dynamic manipulation tasks. On a real robot, our model generates subgoal images and an action chunk spanning one second of motion in 1,030 ms on an RTX 5090 GPU, sustaining closed-loop control at 2.5 actions per second. Second, pretraining on robot video and co-training on human video substantially improves policy performance and zero-shot robustness to distribution shift. Together, these results show that masked diffusion is a practical and versatile foundation for unified video-action modeling.

Recent world action models (WAMs) reuse pretrained video VAEs whose encoder latents directly condition downstream action policies. Quantization must therefore preserve not only reconstruction fidelity but also the policy-facing latent contract expected by the frozen policy. Direct NVFP4 leaves W4A4 quantization error uncompensated, whereas joint quantization-aware training (QAT) can recover reconstruction by moving this representation. On Wan2.1, joint QAT nearly matches FP32 VBench-7 (0.7403 versus 0.7409), yet LIBERO success collapses from 95.5% to 10.5%. Controlled decoder-only experiments show that activation quantize-dequantize operations alter the reconstruction signal and decoder Jacobian, redirecting the gradient returned to the encoder and inducing persistent latent drift. Based on this mechanism, we introduce LatentQuant, a two-stage NVFP4 QAT framework that first aligns the quantized encoder with its high-precision counterpart, then freezes it while adapting the decoder. Across Wan2.1 and Wan2.2, LatentQuant preserves near-baseline control and high reconstruction quality, achieving 95.75% success on LIBERO and 68.8% on RoboTwin. On NVIDIA B300 GPUs, NVFP4 execution achieves 1.17x-1.26x end-to-end VAE speedups over BF16 cuDNN.

Humanoid loco-manipulation demands coordinated body and hand behavior, while conventional robot pre-training data provide limited coverage of such whole-body motion. We present WB-WAM, a World Action Model that incorporates explicit whole-body action supervision into generative video pre-training. A shared physical action space integrates body, root, and dexterous hand annotations from heterogeneous sources, enabling joint video and action learning from 1880.2 hours of partially annotated video and motion data. The resulting priors are refined through PICO mid-training and adapted to robot tasks with auxiliary forward kinematics supervision. We construct WB-Datasets to support these stages with retargeted egocentric human demonstrations and robot trajectories, allowing task-aligned human motion to supplement limited robot data. Evaluations in simulation demonstrate strong whole-body task performance with 81.9% in HumanoidArena, while real-world experiments further validate WB-WAM with 84.0% mean success across five tasks. Moreover, task-aligned PICO mid-training improves downstream task performance while reducing the need for real-robot demonstrations. These results support heterogeneous whole-body pre-training and human motion transfer as a practical route to data-efficient humanoid loco-manipulation.

Project Page: https://wb-wam.github.io

World action models (WAMs) have advanced robot control by predicting how observations and actions evolve over time. Despite this progress, RGB and action based future prediction does not explicitly address the spatial understanding needed for robot manipulation. Existing efforts often add a limited set of spatial prediction tasks through specialized heads or branches, leaving both the range of spatial supervision and the model architecture fragmented. We introduce XGenAct, a world action model that represents RGB observations, robot actions, metric depth, surface normals, and functional role segmentation as RGB videos through deterministic codecs. By sampling perception and action tasks during training, XGenAct uses one video diffusion transformer and one objective to learn temporal prediction across these spaces without modality specific learned heads. On held out RLBench tasks, structured perception training improves average closed loop success over RGB only training, and XGenAct achieves 52% success in the five task external comparison, versus 26% for the strongest evaluated baselines. It also predicts future depth and segmentation more accurately than the evaluated pipelines that generate RGB first and then apply a frozen perception expert.

27 pages, including appendix

World action models integrate future visual dynamics with robot action prediction, but their scalability remains limited by the need for action-annotated robot trajectories. Observation-only videos contain rich evidence about interaction dynamics, but existing approaches typically use them either to pretrain visual representations that must later be adapted for control, or to infer latent actions that are subsequently grounded to robot commands. We present NAVA-WAM, which introduces native action-prior learning by directly pretraining the action policy from observation-only videos, avoiding indirect representation-to-control transfer or a separate latent-action model. Our training consists of two stages. First, we pretrain on observation-only videos, where future-video flow-matching supervision over visual transitions is propagated through transition-structured joint attention to optimize the Action-DiT and learn action-relevant priors. Second, we use action-labeled demonstrations to post-train the Action-DiT for robot control through joint video--action flow matching, while asymmetric attention decouples the visual branch from iterative action denoising and enables efficient action-only inference. Extensive experiments show that NAVA-WAM consistently outperforms prior approaches under both in-distribution and out-of-distribution settings, while demonstrating strong action-label efficiency and effective real-robot generalization. These results establish native action-prior learning as an effective approach to directly pretrain action policies from observation-only videos, providing a scalable path beyond action-labeled robot data.

Project Page: https://zhaochongan.github.io/projects/NAVA-WAM

Generalist robot policies, such as vision-language-action models (VLAs) and world-action models (WAMs), combine powerful pretrained backbones with expressive generative action heads trained via flow matching on large-scale robotic datasets. Despite their strong empirical performance in robotic manipulation, these policies lack mechanisms to quantify confidence in their predictions and to detect when their actions may be unreliable. This presents a critical limitation for real-world deployment in non-stationary environments, where models inevitably encounter scenarios outside their pretraining distribution and may fail without warning. To address this, we derive an efficient method to quantify epistemic uncertainty in flow-matching models by leveraging velocity-field disagreement (VFD) across a small ensemble. We successfully use this uncertainty estimate for detecting failures during deployment and active fine-tuning of flow-based generalist policies. For the latter, we propose SAVE, a simple yet effective method for uncertainty-guided active multitask fine-tuning that reduces the number of costly expert demonstrations required to adapt generalist policies to new tasks. We conduct experiments in simulation and the real world, across VLAs and a WAM. VFD yields better-calibrated uncertainty estimates predictive of downstream performance and detects failures with 8 pp higher overall accuracy than existing methods. Across three real-world tasks, SAVE improves final average success from 39 % to 47 % with a fixed demonstration budget. Our results show that measuring epistemic uncertainty with VFD enhances both failure awareness and adaptation of generalist robot policies. Project website: tum-lsy.github.io/uq_generalist_policies.

Project page: tum-lsy.github.io/uq_generalist_policies/. 41 pages, 18 figures

Genuine embodied agency requires robots to turn continuous real-world experience into lasting, transferable skills. This demands continual learning that integrates new capabilities without eroding prior knowledge as tasks and environments evolve. Experience replay mitigates forgetting, but storing complete demonstrations becomes costly as tasks accumulate. World-action models offer a generative alternative, reconstructing past experience through joint predictions of actions and future observations. However, visually coherent rollouts may contain actions that cannot realize the predicted transitions, while new-task adaptation can disrupt previously learned behavior. RIFAR therefore combines reliability screening with drift-aware replay selection. It reconstructs trajectories from compact demonstration prefixes and uses a frozen inverse-dynamics model to assess action-visual consistency. Training first combines current demonstrations with the highest-quality screened trajectories. RIFAR then compares action predictions before and after this adaptation on identical historical inputs, reselecting trajectories with larger normalized drift from the same screened pool for continued training. Across three LIBERO suites and real-world experiments, RIFAR surpasses the previous state of the art in WAM-based generative replay. On LIBERO-Goal, it achieves 90.97 AUC while retaining only 320 historical time steps per task, approximately 4.9% of the steps retained using 50-demonstration replay.

15 pages, 6 figures, 9 tables, including appendices

As AI agents become increasingly capable, agent-driven robotic control is emerging as a compelling paradigm. However, prevailing vision-language-action (VLA) models and world action models (WAMs) still rely on natural-language instructions to specify manipulation tasks, an ill-suited interface for agent-driven control: referentially ambiguous, spatially imprecise, redundant with the agent's inherent language understanding, and entangling intent with execution. We present AR-WAM, a visual-conditioned, agent-ready world action model that replaces language with two complementary conditions: a visual grounding prompt (a bounding box of the target) denoting the interaction object and location, and a learnable operation token dictating the atomic skill to execute. Our compact 0.5B-parameter model, with a frozen pretrained visual encoder and no language encoder, predicts scene evolution within compact latent states while decoding actions, exposing the policy's intent through explicit, supervisable reasoning signals. A model-agnostic compatibility layer provides three primitives (detect, execute, and query) so that local VLMs or online agent APIs can drive the policy directly, with long-horizon memory and closed-loop error recovery delegated to the agent side. On RoboTwin 2.0, RMBench, and a real Astribot S1 dual-arm platform, AR-WAM attains the highest average success on standard manipulation (85.7% over the clean and randomized settings) and outperforms all baselines on memory-dependent and real-robot long-horizon tasks, improving success rates by 5.9% and 36.7%, respectively, while maintaining the lowest inference latency (14.1 ms). Project page is at https://ar-wam.github.io/.

World action models jointly learn to forecast world dynamics and predict robot actions, such that the learned internal world dynamics guide accurate actions. Existing approaches typically represent the world as RGB frames or latent counterparts while predicting actions as end-effector poses or joint angles, but they often struggle to capture the 3D spatial structure and contact geometry central to dexterous manipulation. We introduce Point World Action Model (PointWAM), a 3D world action model that decomposes the world into a scene (i.e., environment) and hands (i.e., actor), and jointly forecasts both as 3D point trajectories within a shared space-time coordinate frame. This explicit, disentangled representation enables effective pre-training on large-scale human demonstration videos without requiring any task-specific object or keypoint selection. Given a colored point cloud and a language instruction, PointWAM predicts how the scene and hands co-evolve in 3D space over time, then retargets the forecast hand motion to robot actions. Pre-training on human videos improves average DexJoCo success by 56.9 percentage points, and scene-trajectory supervision adds 10.9 points over forecasting the hands alone. With both, PointWAM surpasses the prior state of the art on ten DexJoCo tasks by 11.7 points and outperforms strong VLAs on a real robot.

Preprint. Project page: https://chrockey.github.io/PointWAM

Vision-Language-Action (VLA) foundation models have scaled rapidly to enhance manipulation performance and generalizability, but this scaling incurs high computational costs that render real-world deployment increasingly challenging. Existing approaches typically mitigate this issue by designing smaller architectures or reducing the iterative denoising steps in flow-based policies. In this work, we propose FastOPD, a foundation-to-lightweight VLA framework that enables the practical deployment of large-scale VLAs through efficient on-policy distillation. Specifically, FastOPD adapts a flow map for single-state teacher supervision and combines it with a self-consistency objective to construct a compact student that learns the teacher dynamics. Furthermore, we theoretically demonstrate that minimizing this objective allows the distilled student to recover a distribution on par with that induced by an ideal few-step teacher model. We evaluate FastOPD across diverse foundation policies in simulation and real-world experiments. On LIBERO, FastOPD retains 84% of the performance of $π_{0.5}$ with only two inference steps, reducing inference latency by 78.1% while outperforming existing few-step distillation baselines in average success rate. With LingBot-VLA as the teacher, FastOPD improves the single-step success rate over the base student by 15.9 percentage points on RoboTwin 2.0. We further demonstrate its applicability to a World Action Model (WAM) and deploy a compact student distilled from MolmoAct2 on a real robot.

Project page: https://fastopd.github.io/

World Action Models (WAMs) couple visual dynamics prediction with action generation, yet they do not explicitly support the reuse of action experience across manipulation tasks. Furthermore, existing WAMs struggle to capture underlying cross-task semantic relationships that could guide target action prediction, as redundant background elements interfere with the extraction of key visual information. To address these challenges, we develop a novel Action Experience Dictionary (AED) that encodes historical physical action trajectories into shared action embeddings to support skill reuse and model cross-task relationships. Specifically, we first aggregate historical actions to align with visual observations and retrieve action embeddings from the AED using a pretrained action tokenizer. Subsequently, we visually condition the pooled embeddings through cross-attention and prepend them to noisy action tokens, providing interaction context and action intent for prediction. To model action-related motion and reduce reliance on irrelevant background cues, we introduce a motion-aware transition loss that supervises visual feature change prediction over random temporal intervals. Experiments on simulation benchmarks and in real-world cross-embodiment settings verify the effectiveness of our AED. The project code is available at https://github.com/JiahuaDong/AED .

World action models (WAMs) have emerged as a promising paradigm for robotic control by jointly predicting future visual dynamics and actions from an initial observation and instruction. However, existing WAMs struggle with long-horizon prediction, as generating dense video rollouts is highly inefficient. Some recent WAMs address this by predicting a single future frame without generating the full video, but this approach neglects how to progress toward the goal. We present ProWAM, a progressive world action model that jointly predicts actions and an ordered sequence of sparse visual sub-goals, providing explicit visual guidance to anchor action generation throughout task execution. This design scales naturally, as sub-goal prediction can be learned from large-scale action-free videos, allowing the video backbone to offload complex visual planning from the action policy. For efficient action generation, ProWAM executes a single video-backbone forward pass to cache sparse sub-goal features, eliminating iterative full-video generation and requiring only lightweight action denoising during replanning. Across extensive evaluations, ProWAM achieves superior out-of-distribution robustness. On simulation benchmarks, it sets new state-of-the-art results on LIBERO-Plus (85.8%) and randomized RoboTwin (75.7%), outperforming the strongest baseline with relative gains of up to +35.9%. On RoboCasa365, ProWAM achieves a 48.1% success rate and 18.2% on the challenging Composite-Unseen split, ranking 4th overall. Crucially, in zero-shot real-world experiments, ProWAM achieves 70.0% success, outperforming the strongest baseline by +15.0 (from 55.0% to 70.0%, a +27.3% relative gain) in novel scenes. These results demonstrate the value of progress-indexed visual foresight for closed-loop control. Our program is in https://sii-ferenas.github.io/ProWAM-page.

Project Page: https://sii-ferenas.github.io/ProWAM-page

A robot demonstration records two things in the same frames: what happened to the objects, and how one particular arm made it happen. We condition on the first. A demonstration is compiled into an effect program: the 3D keypoint trajectories of the objects that moved, two points marking where each was held, and the configuration the scene ends in, with the demonstrator removed. PEWAM, a 71.5M-parameter world-action model, generates effect, robot execution, action and terminal state as four streams with independent flow-matching times, so clamping a program and sampling the execution turns inference into programming, re-solved closed loop from the live scene. On held-out LIBERO-Goal tasks, one demonstration's program completes 40 of 90 episodes, where the same backbone given a goal image or language, and published demonstration-conditioned methods, complete at most 19; on three of Meta-World's held-out classes it exceeds the best published results, though not on the five-class mean. Because a program is a set of coordinates, a person can edit it: the placement follows a shifted terminal state and the grasp turns with rotated contact points. The same program runs on four robot arms without retraining, and after a push, re-solving completes 23 of 60 episodes where replaying the demonstration completes 6. On a Franka arm, fine-tuned on real demonstrations of other tasks, programs compiled from single human videos complete 36 of 40 trials, against 22 for the same backbone conditioned on the video's last frame as a goal image.

Long-horizon compositional manipulation has become increasingly important for real-world robot deployment, where a single task involves multiple coordinated subtasks. Existing world-action models (WAMs) jointly predict short-horizon visual futures and actions, but typically lack explicit subtask-level reasoning. We propose Visual Goal-conditioned Action Reasoning (ViGAR), a hierarchical framework that factorizes manipulation into a visual subgoal planner and a subgoal executor. Given the current observation and global instruction, the subgoal planner predicts a visual subgoal for the next subtask. The subgoal executor then jointly generates future visual trajectories and actions conditioned on the predicted subgoal. Both components share a pretrained world-model representation, enabling task-level planning and action generation to benefit from common physical knowledge. Moreover, our framework naturally supports in-context learning: using a global goal image as context can induce different subtask decompositions and behaviors without parameter updates. On the RoboTwin Clean2Random benchmark, ViGAR achieves 82.00% and 67.02% success rates under the Clean and Random settings, respectively, surpassing the strongest baseline by 12.86 percentage points in average success rate. Real-world robot experiments on five compositional and two in-context learning tasks further confirm the effectiveness of ViGAR.

World Action Models (WAMs) generate both future video and robot actions, offering two connected outputs for post-training supervision. How can a pretrained WAM learn from a stronger Teacher on the histories it encounters during execution? We present WAM-OPD, which collects Student rollout histories and queries a Teacher for paired video and action targets. The Student learns from both targets while retaining its one-step video and action generation at deployment. Across 12 RoboTwin 2.0 tasks, WAM-OPD improves average success from 33.8% to 65.7%; across four real-robot tasks, it improves average success from 51.4% to 64.6%. With the collected Student histories held fixed, joint video-action supervision achieves the highest observed success on all three ablation tasks, while either modality alone also improves performance. A separate comparison with Teacher-generated histories finds task-dependent differences between the two history sources. These results demonstrate the value of paired video-action supervision for improving WAM policies without increasing their deployed sampling budget.

World action models (WAMs) combine robot action generation with future state prediction. Existing WAMs typically predict videos or learned visual latents, which represent interaction geometry only implicitly and may retain appearance information unrelated to control. We introduce SkeleWAM, a compact WAM that represents a manipulation scene as a sparse 3D skeleton composed of robot joints, object centers, and interaction points. Constructed online from current RGB-D observations and robot proprioception, the skeleton provides a unified geometric state for action generation and future skeleton prediction. Future skeleton prediction provides additional geometric supervision for action learning without requiring visual reconstruction. At inference, SkeleWAM generates actions directly from the current skeleton and language instruction, while Medoid Action Consensus (MAC) serves as an auxiliary consensus strategy for stochastic action samples. On LIBERO-Plus, SkeleWAM achieves an overall success rate of 85.9% with 57.1M parameters, outperforming Cosmos-Policy by 3.7 percentage points. These results demonstrate that sparse 3D robot--object structure provides an effective state space for robust and parameter-efficient world action learning. The project is available at https://skelewam-project.github.io/.

Active vision manipulation requires a policy to control both its camera and its end-effectors, yet camera motion determines which evidence remains visible within finite observation windows. Acquiring a new view can displace task-critical cues, while retaining a view forgoes potentially useful observations. We formulate this as an evidence-aware retain--acquire problem and present ActiveWAM, a unified world--action model that learns observation and manipulation jointly. To this end, we propose training-time inversion which constrains a frozen video prior by task-bearing source evidence and visible temporal changes, eliminating the need for test-time inversion or candidate ranking. At deployment, the policy generates bimanual and pan/tilt actions-including stay and reacquisition behaviors-from view-aware history, and updates context from newly measured RGB observations. Future-video prediction serves as a co-training signal, while action generation requires neither future-video decoding nor optimal viewpoint annotations. We introduce RoboTwin-AV, a 50-task benchmark with executable pan/tilt control and automatically generated demonstrations. ActiveWAM improves TAVIS out-of-distribution success by up to 17.0 percentage points over the strongest baselines, achieves 20.0 additional points over Fast-WAM on RoboTwin-AV, and outperforms it by 26.7 points on real-world physical kitchen tasks.

World Action Models (WAMs) predict visual futures and robot actions, yet they remain susceptible to task-incomplete imagination, where plausible, action-consistent predictions omit the transition needed for task completion. In this paper, we show that this failure is not inherent to the world model backbone, but emerges when adapted for short-chunk control, which can repeatedly favor plausible local continuations over task-completing transitions. To address this, we introduce Completion Aware Guidance (CAG), a training-free sampling method that guides generation toward task completion. Across representative WAMs, CAG improves success from 64% to 70% on a RoboTwin 2.0 subset and from 69% to 75% in zero-shot simulation, while reducing task-incomplete imagination from 79% to 40%.

Generating useful robot-training data requires more than visually plausiblescenes: objects must support interaction, placements must remain physicallyvalid, and tasks must admit repeatable execution. We present\textbf{Awomo-SimDataEngine}, an agentic system that connects asset and scenegeneration to robot demonstration synthesis. Shared asset services providerigid and articulated objects, including structure-grounded part and jointgeneration with ISArt. Scene generation supports two complementary routes:Unravel reconstructs editable scenes from images, while SimForge buildssingle-room and multi-room environments from text. A graph-native harnesscoordinates construction, validation, andbounded repair, routing failures to the responsible module while retainingunaffected scene state. PolicyForge binds validated worlds to tasks and robotembodiments to produce replayable demonstrations. Evaluations cover assetgeometry, scene quality, and downstream policy learning. On MuJoCo-basedLIBERO-Plus, co-training with Isaac Sim demonstrations improves the overallsuccess rate of a World-Action Model (WAM) from $77.17%$ to $89.43%$. Goal and spatialsuccess improve by $31.66$ and $6.25$ percentage points, respectively.These results support the utility of the generated data for cross-simulatorpolicy training, with more limited gains on long-horizon tasks.

Vision Language Action (VLA) models condition actions directly on current visual and language context, without an explicit account of how the scene evolves under candidate actions. World Action Models (WAM) attempt to address this limitation by predicting future states, but existing designs keep prediction and policy learning architecturally separate, connecting them only through the predicted output, whether through pixel space video generation or a latent forecasting module trained independently of the policy. We present Devol-ONE, a Mixture of Transformers architecture that unifies vision language understanding, latent world dynamics prediction, and action generation within a single autoregressive framework. Instead of encoding vision language tokens once and feeding them to the action expert, Devol-ONE runs autoregressive prediction jointly across a vision language stream and a V-JEPA pretrained dynamics stream, attending to the vision language key-value cache at every layer to forecast future latent states under language guidance. The action expert is in turn shaped continuously by semantic reasoning and predicted physical dynamics rather than by a fixed representation computed in advance. Extensive experiments are conducted on LIBERO, LIBERO-PLUS, RoboTwin2.0 along with real-world evaluation on Flexiv single-arm and dual-arm setups. Ablation studies show the effectiveness of dynamic stream prediction and layer-wise unified attention to validate our model architectural coherency.

Humanoid whole-body manipulation requires coordinated whole-body dynamics, yet large-scale trajectories from a target robot are expensive to collect and difficult to scale. In contrast, whole-body motion from human and humanoid sources is abundantly available, although such data cannot be directly used as embodiment-specific robot actions. This work asks whether these scalable motion resources can instead provide a transferable predictive prior for humanoid world-action modeling. We introduce WholeBodyWAM, a humanoid world-action model that learns whole-body dynamics from large-scale heterogeneous motion before target-robot training. We curate UniMotion-4K, a motion corpus spanning more than 4K hours from human videos, native 3D motion datasets, and heterogeneous humanoid platforms, and canonicalize these diverse sources into a unified motion space. A language-conditioned Motion Expert is then pretrained to predict future whole-body motion without target-robot action supervision. During robot post-training, the pretrained Motion Expert is integrated with Video and Action Experts through asymmetric Mixture-of-Transformers (MoT) attention, enabling predictive scene dynamics and whole-body motion to jointly inform embodiment-specific action generation. Experiments show that WholeBodyWAM consistently benefits from increased motion-pretraining scale, improves future-motion prediction and downstream task performance, and transfers effectively to real-world humanoid manipulation. Moreover, the pretrained motion prior substantially improves data efficiency under limited target-robot demonstrations.

Project page: https://zbzyjya.github.io/WholeBodyWAM/

World Action Models (WAMs) leverage pretrained video foundation models for robot manipulation, but their large backbones and video-action co-prediction are expensive. Although existing acceleration techniques offer many ways to reduce this cost, selecting and composing them requires substantial engineering for each model and hardware platform. To tackle this bottleneck, we present WAMJET, an agentic harness that accelerates WAM inference by equipping coding agents with reusable optimization guidance and measurement and validation tools. WAMJET follows a bottleneck-driven workflow where the agent profiles inference, modifies targeted code, validates effects, and iteratively refines the acceleration stack as bottlenecks shift, while preserving action quality. Experiments span six WAMs, three coding agents, and two GPU architectures. WAMJET achieves up to 9.95x lossless speedup over upstream implementations. Approximation and hardware-aware optimization yield additional latency reductions, with comparable success rates. The results show that WAMJET can produce effective acceleration stacks for WAM deployment.

8 pages, 3 figures, project page: https://github.com/liulixinkerry/WAMJET/blob/main/assets/blog.md

World Action Models (WAMs) connect visual prediction with robot control, but supplying predictive context often requires expensive future-video generation. Direct policies avoid this cost but lack an explicit interface for accessing future-indexed predictive information. We introduce ForeWAM, a World Action Model that separates forecasting from rendering to expose and shape latent predictive context for efficient control. Its core mechanism, Future-KV, performs a single Video DiT prefill over the current visual latent and noise-initialized future slots, then reuses the resulting key-value states throughout action denoising. To make this context relevant to control, we introduce dynamics registers supervised by latent actions from a frozen teacher during training, encouraging representations of interaction-induced transitions. This reusable context supports a lightweight, single-layer action decoder. We evaluate ForeWAM on LIBERO, LIBERO-Plus, RoboCasa, and real-world manipulation tasks. Without additional policy-level embodied pretraining, ForeWAM improves RoboCasa success by 9.7 percentage points over Fast-WAM at the same budget of 50 demonstrations per task, reaching 59.2%. With a single-layer decoder, it achieves 77.6% success on LIBERO-Plus and reduces policy-query latency to 88.7 ms on an NVIDIA A800, delivering a 6.27-fold speedup over Fast-WAM. These results show that latent predictive computation provides useful foresight for robust, efficient control without explicit future-video generation.

World action models (WAMs) jointly predict actions (intent) and visual future (foresight). Standard training adds noise to recorded actions and video simultaneously, but such training paradigms introduce a mismatch: perturbed actions imply counterfactual future visual, while the noised video remains tied to the GT recording. In low-noise regime, the scene geometry and even the dynamic behavior remain clearly visible from the noisy future frames despite the added noise. We present CtrlWAM, which executes perturbed actions in a simulator and pairs them with their noised visual consequences for joint WAM learning. To accommodate the different denoising requirements of video and actions, we introduce warped video--action noise schedules that aim to keep visual layout responsive as action predictions evolve. We further extend the action interface from ego-only control to a variable number of agent streams, allowing a unified model to represent predicted or commanded futures for multiple agents. Driving experiments show more accurate action forecasts, closer agreement between generated video and actions, and better following of supplied commands; robotics experiments show stronger motion fidelity and controllability. Matched controls support the benefit of off-path renders for command following and manipulation fidelity. Together, these findings contribute to a more controllable world action model. Project page: https://ctrl-wam.github.io/

Project page: https://ctrl-wam.github.io/

World action models improve robotic manipulation by conditioning actions on predicted futures, yet existing tactile variants largely inherit video-generation pipelines that reconstruct future tactile observations through iterative denoising. Such prediction can become unreliable under deployment drift: small changes in contact position or force may substantially alter tactile pixels even when the underlying contact evolution remains predictable. We introduce TacDyn-WAM, a heterogeneous visuo-tactile world action model that predicts implicit tactile dynamics rather than reconstructing future tactile observations. It learns TacRep, a dynamics-aware tactile target space trained through masked spatio-temporal prediction on tactile clips and regularized by relational structure distillation. A visual expert and an Implicit Tactile Dynamics Expert predict future visual and tactile representations in separate target spaces while interacting through joint attention; the tactile expert predicts future representations and their changes at multiple horizons in a single forward pass, and a read-only tactile memory supplies the current tactile state. On UniVTAC, TacDyn-WAM achieves an average success rate of 81.5% using only the provided demonstrations, reaching state-of-the-art-level performance and remaining competitive with models pretrained on large-scale visuo-tactile trajectories. Ablations confirm the benefits of both tactile pathways and TacRep over pixel-reconstruction and static alternatives. On five real-robot tasks, TacDyn-WAM reaches 71.0% average success, and modest-scale pretraining raises it to 85.0%, further validating our method.

Egocentric human data provides a scalable source of experience for robot learning, but varies substantially in human-robot alignment, behavioral coverage, and available supervision. Existing work shows favorable scaling with increasing human data, but it remains unclear which data properties drive downstream robot gains and how to use such data throughout the training pipeline. We present a systematic study of egocentric human data with different alignment and supervision under a unified world-action model framework. With the model backbone fixed, we disentangle the effects of human-robot alignment, data duration and task diversity, action supervision, and data usage strategies. We find that aligned human demonstrations substantially improve out-of-distribution generalization and reduce target-task robot data requirements; data duration and task diversity affect downstream capabilities differently; and video-only supervision remains effective without action labels, providing a strong foundation for subsequent video-action training. We validate these findings through closed-loop policy evaluation on both real robots and RoboDojo. Rather than treating data duration as the sole scaling axis, Ego4WAM shows how alignment, task diversity, available supervision, and usage strategy jointly shape the value of egocentric human data for robot learning.

Vision-language-action (VLA) policies emphasize semantic understanding, whereas world-action models (WAMs) learn predictive representations of environment dynamics. Systems that expose a policy to both sources often still concentrate action computation on a single expert. We present EWAM, an action-centric unified embodied model whose asymmetric joint attention lets action tokens read semantic, current-visual, predicted-future, and action information at every layer while the perceptual experts retain their distinct roles. Without layer-wise supervision, EWAM develops an emergent depth-wise specialization: action queries attend mainly to vision-language features in shallow layers, to predicted future frames in intermediate layers, and to action tokens themselves in deep layers. This handoff replicates across tasks and is stable across denoising steps. Checkpoint tracking and causal interventions show that it is learned and that action generation depends on it. EWAM is pretrained in two separate regimes, one on cross-embodiment robot trajectories and one on human egocentric video. In simulation and real-robot experiments, it surpasses existing VLA, WAM, and hybrid baselines. Human egocentric data improve both cross-embodiment transfer and real-robot robustness, and subtask-phase supervision improves long-horizon completion. Together, these results suggest that unified embodied learning can induce an ordered internal progression from semantic understanding, through visual foresight, to action formation.

World action models (WAMs) are large embodied policies that jointly predict future video and the actions to execute, emitting a fixed-length action chunk per inference call. Such a policy allocates its computational budget uniformly in time, unable to execute for longer over free-space motion or to spend more inference on contact-rich manipulation, which limits the throughput a WAM can reach when served in the cloud. We present SplineWAM, which adaptively compresses the action trajectory into a fixed-size window of cubic B-spline parameters, fitting the knot times to the characteristics of the motion. One parameter budget then decodes into chunks of varying temporal resolution and duration, and both the executed span and the interval until the next policy call follow from the prediction itself. Aligning the video supervision to the fitted knot times of the demonstration rather than to a uniform grid concentrates the supervised frames where the action trajectory is complex. For asynchronous deployment we introduce Jacobian-Pullback Real-Time Chunking (JP-RTC), which imposes chunk continuity on the decoded raw actions the robot executes rather than on the spline parameters, and corrects the parameters through the decoder so that the executed prefix agrees with the actions already committed. On LIBERO-Plus and RoboCasa, SplineWAM improves success rate over an action chunking WAM by $8.2$ and $4.4$ points while cutting policy calls per episode by 22% and 26%. On three bimanual real-robot tasks under asynchronous execution, it leads or matches the baseline while decoding 1.2 to 1.6 times as much executed motion per call.

Project website: https://splinewam.github.io/

Pretrained video foundation models encode rich semantic and spatiotemporal priors for embodied navigation, yet converting these priors into UAV motion typically requires expensive future-video synthesis and geometric reconstruction. We investigate whether the motion implicit in future visual prediction can instead be recovered directly from the predictive representations of a frozen video model. To this end, we present DiffWAM, a geometry-conditioned navigation world-action model that directly transforms multi-level predictive features into continuous camera trajectories. Its Grid-Motion module preserves spatial-temporal motion associations, while Latent2Pose grounds them with first-frame geometry to recover metrically meaningful 3D motion. Complete video rollouts and geometric reconstruction are required only for offline supervision, eliminating future-video decoding and multi-frame reconstruction during deployment. We further introduce FastDreamer, which overlaps predictive and geometric computation with ongoing flight and performs timestamp-aware asynchronous trajectory handoff for continuous UAV execution. DiffWAM achieves a trajectory RMSE of 0.3492 m and an endpoint success rate of 74.40% on the 1,000-sample DiffWAM-1000 benchmark, while representative real-world experiments demonstrate complex behaviors including constrained traversal, orbiting, S-shaped flight, and multi-stage navigation. An onboard DiffWAM-Flash implementation further reaches 1.08 s model-pipeline latency on NVIDIA Jetson AGX Thor. These results demonstrate that predictive video representations can be efficiently grounded into continuous 3D motion, providing a direct alternative to generate-then-reconstruct navigation pipelines. Project page: https://zzmmzzm.github.io/diffwam.github.io/.

32 pages,10 figures, 8 tables

Precise grounding matters. It specifies which object is the target and where that object is, even in clutter and for tiny objects, and it has to be fast enough for closed-loop control. Yet vision-language-action (VLA) and world-action models (WAMs) take perception from general-purpose vision-language and video-generation backbones, which still fail in these settings. We introduce GroundingPI, a 4B grounding foundation model that generates points and boxes as quantized coordinates in a shared vocabulary. Training combines multimodal and spatial pretraining, supervised fine-tuning, and reinforcement learning with GRPO, using supervision from public datasets and dedicated data engines. Against 44 baselines across 34 grounding benchmarks spanning 11 perceptual capabilities, GroundingPI establishes a new state of the art, averaging 73.68%, above the larger GPT-6 Astra (71.54%). As a downstream visual backbone, GroundingPI improves performance on robotic manipulation and autonomous driving. On RoboTwin 2.0, it outperforms every mainstream backbone we evaluate in all four out-of-distribution settings, by up to 24.8% relative to the strongest backbone. On RoboCasa-GR1, GroundingPI trained with 50% of the demonstrations outperforms those baselines trained with 75%. On nuScenes, used as the visual backbone, GroundingPI attains an average open-loop L2 error of 0.296 m. We systematically analyze GroundingPI's pretraining in scale and data composition. Downstream autonomous driving and robotic manipulation improve as the pretraining is scaled. Analyzing the data recipe across these 11 perceptual capabilities shows dense grounding's substantial benefits for both, and OCR's potential as a catalyst for perceptual learning. These results support grounding as a perceptual foundation, and dedicated perceptual pretraining as a promising direction for foundation models of physical intelligence.

64 pages, including supplementary material. Project page: https://groundingpi.github.io/ Code: https://github.com/groundingpi/GroundingPI Model: https://huggingface.co/GroundingPI/GroundingPI

World-action models (WAMs) couple predictive visual modeling with action generation, typically relying on iterative denoising with a fixed denoising steps. However, manipulation tasks contain actions chunks with varying sensitivity to generation errors: critical actions require precision, while less sensitive actions allow faster generation with fewer denoising steps. Here we introduce AnyStep World Action Model, a general framework for tunable-budget prediction and scene-dependent computation allocation. Our budget-aligned teacher-trajectory distillation trains interval-conditioned flow maps using explicit frozen-teacher transitions and shared low-rank adapters, supporting action generation from one-step prediction to multi-step refinement. Building on this capability, a lightweight risk-benefit scheduler predicts teacher-curvature-based difficulty and budget-specific student-teacher fidelity from a single one-step preview, selecting the smallest budget predicted to satisfy risk-adaptive fidelity requirements. We evaluate our framework on three widely used WAMs Motus, FastWAM, and LingBotVA using RoboTwin 2.0. Our method reduces average denoising steps by 60.2%, 49.8%, and 85.28%, respectively, while maintaining baseline task success rates. In particular, our AnyStep training substantially improves model performance under a one-step denoising budget, increasing task success rates by 7.07%, 12.08%, and 8.94% on Motus, FastWAM, and LingBotVA, respectively. Experiments on six real-world manipulation tasks further validate its effectiveness.

Mobile manipulation requires precise navigation to a manipulation-ready pose followed by reliable object interaction. These two stages differ in action spaces and visual requirements, which complicates unified policy learning. In addition, collecting diverse real-world navigation data with explicit manipulation-ready pose supervision remains costly and difficult to scale. We introduce UniWAM, a unified mixed-stream world-action model with separate action encoders and output heads for navigation and manipulation, sharing a common backbone. This design supports joint representation learning on independently sampled navigation and manipulation data. UniWAM supports independent inference for either stream and batch-parallel inference for both. We further introduce Manipulation Anchor Pose (MAP) supervision for where to stop and how to orient for manipulation. An automated pipeline constructs MAP-Data from large-scale 3D scenes, yielding over 1.5 million episodes and 7,500 hours. MAP-Data provides per-frame target-object bounding boxes and image-plane MAP coordinates as auxiliary navigation supervision. Together with projected end-effector trajectories for manipulation, these prediction targets provide stream-specific image-plane supervision for action learning from egocentric observations. With large-scale MAP-Data, UniWAM outperforms the strongest external baselines on our MAP-Bench by 30.1% in position error and 44.0% in heading error. Across 24 real-robot tasks, UniWAM achieves leading results in MAP navigation and mobile manipulation, with competitive manipulation performance. We have released code, data, and benchmark.

UniWAM Technical Report

In interactive scenarios, an autonomous driving system is required to generate ego actions under the influence of other agents' behaviors. Existing World Action Models (WAMs) typically model other agents as components of the world model rather than as decision-makers that fundamentally shape the action of the ego agent, which impairs their performance in dense interaction scenarios. We introduce Reciprocal World Action Models (ReWAM), a game-theoretic world action modeling framework that captures the reciprocal influence between the ego agent and other agents by representing them as conditional responders whose actions are mutually influenced. We instantiate this framework with a Level-$k$ response hierarchy, where role-specific ego and other action DiTs exchange compact strategy tokens through cross-agent attention while remaining grounded in a shared representation of the future driving world. To learn the response policy of the ego agent from demonstrations, we formulate expert actions as samples from the best response distribution and jointly optimize the entire hierarchy using conditional flow matching. Our framework is evaluated on the NAVSIM dataset and achieves state-of-the-art performance compared to baselines. The improvement is particularly significant in interactive scenarios, validating that modeling reciprocal responses provides a more effective foundation for interaction-aware world action generation.

World-action models (WAMs) jointly generate future world states and actions through iterative denoising, using shared weights to process heterogeneous semantic streams of video, proprioceptive, and action tokens. Quantization reduces inference cost, but comparable numerical errors in different streams can have markedly different effects on final actions, making numerical accuracy alone insufficient for reliable control. We introduce SteerQuant, a 4-bit quantization framework for WAMs that steers errors toward computations with less influence on final actions. It maps how each stream's quantization errors affect final actions and uses this map to guide shared channel scaling. Activation scaling is further calibrated for each stream and denoising step to accommodate changes in activation ranges and action impact. This adapts quantization to different stream requirements without duplicating weights or increasing bit-widths for selected streams. To reduce the extra kernel launches and memory traffic introduced by scaling, we develop Rudder, a 4-bit inference engine for WAMs that fuses scaling and output compensation into low-bit kernels. Under W4A8 and W4A4, SteerQuant maintains mean LIBERO success within 0.8 percentage points of full precision, while delivering up to $2.23\times$ denoising speedup over BF16 across three WAMs with reduced peak GPU memory usage. On a real dual-arm robot, W4A8 deployment achieves a $1.35\times$ end-to-end inference speedup while maintaining average task success relative to BF16.

Yunhan Wang and Haodong Wang contributed equally to this work

World-action models (WAMs) leverage pretrained video models to improve generalization in robot control by jointly predicting future visual states and actions. This capability comes at a substantial inference cost, as dense future-frame tokens are repeatedly processed during denoising. Prior methods address this by token pruning that prioritizes visual fidelity to reduce denoising costs in video diffusion models. However, these methods do not use action relevance to determine which future-frame tokens to retain during joint denoising in WAMs. In this paper, we propose Sparse-WAM, a training-free framework for action-guided sparse imagination that selectively processes future-frame tokens to accelerate WAM inference. We observe substantial overlap in the spatial distribution of attention from action tokens to future-frame tokens (action-to-future attention) between consecutive denoising steps, despite continued updates to the future representations. Motivated by this, we develop Action-Guided Token Selection to retain frame-specific action-relevant regions together with cross-frame context. However, a naive implementation can incur attention-scoring and token-packing overhead that offsets the computational savings from pruning. We therefore introduce Pilot, an efficient engine that reduces sparse inference overhead through lightweight scoring and cross-step reuse of token selections. On LIBERO with FastWAM-Joint and RoboLab-120 with Cosmos 3 Edge, Sparse-WAM achieves inference speedups of approximately $2.0\times$ and $1.8\times$, respectively, over dense eager inference on an NVIDIA RTX 4090, while largely preserving task performance.

Xinling Xie and Haodong Wang contributed equally to this work

Vehicle-infrastructure cooperation can complement onboard sensing with broader and more informative observations of the traffic environment, providing valuable support for end-to-end autonomous driving. However, existing cooperative driving methods mainly exploit roadside information to enhance the representation of the current scene, while the future consequences of prospective driving actions are rarely modeled explicitly. This limits the ability of the planner to anticipate how its decisions may interact with the evolving traffic environment. To address this issue, we propose V2X-WAM, a cooperative world action model that tightly couples cooperative scene understanding, action generation, and future-world reasoning. V2X-WAM constructs a reliability-aware spatiotemporal representation from vehicle- and infrastructure-side observations, while compressing infrastructure information into a compact quantized message for efficient communication. Based on the resulting cooperative representation, a multimodal planner generates prospective trajectories, which explicitly condition future occupancy and dynamic-flow prediction. The predicted world consequences are then fed back to refine the planned trajectory, forming a closed interaction between action and future-world evolution. Experiments on a large-scale real-world cooperative driving dataset demonstrate that V2X-WAM consistently improves planning accuracy and safety over representative end-to-end cooperative driving methods, while achieving stronger future-world prediction and substantially lower communication overhead. Ablation studies further validate the effectiveness of the proposed design.

Long-horizon robot manipulation chains many contact-rich skills into one multi-stage task. Vision-language-action (VLA) and world-action models (WAMs) increasingly master individual skills, yet the chain still fails: errors compound beyond the policy's ability to correct, and one subtask silently constrains the next. A promising pathway freezes the VLA and puts an LLM coding agent in charge: it plans in language, moves in free space with analytic primitives, invokes the VLA only for contact-rich segments, and writes adaptation into language memory. Yet applied to long horizons, this recipe breaks twice. (1) Its competence comes from whole-task exploration at test time, whose cost is exponential in the number of stages: if one stage needs T episodes, a K-stage task needs on the order of T^K, and a failure does not reveal which stage caused it. (2) It has no representation of transitions: the VLA primitive carries an exit but no entry condition, and a subtask can succeed in a form its successor cannot use. We present BATON to address both failures. Against (1), BATON makes the subtask the unit of exploration: each subtask is explored in the cheap short-horizon regime and its solution stored in memory; a long-horizon trajectory is then composed from these solutions rather than discovered whole. Exploration cost becomes linear (KT), and each failure is attributed to one stage. Against (2), BATON equips exploration with a transition-aware memory. Within a subtask, a verifier agent governs the invocation transition: the VLA is invoked only after the wrist view confirms the scene is ready. Across subtasks, a handoff transition restores an entry state disturbed by the predecessor's residue, and a lookahead transition selects the strategy whose outcome the successor can inherit. On the RoboMemArena benchmark, BATON improves task success by 37.7% and cumulative success by 29.7% over the SoTA.

Video generative models provide rich physical priors for robot learning, yet existing world-action models (WAMs) face a fundamental trade-off: synchronous video generation at control rate is latency-prohibitive, while abandoning test-time visual imagination sacrifices task success. We show that visual imagination achieves both real-time inference and superior success rates when generated asynchronously off the critical path and consumed directly in latent space. We introduce GlanceWAM, which decouples imagination from control on a single shared video DiT backbone: an asynchronous proposer glances ahead on a slow clock to imagine a single lookahead frame seconds into the future in the background, while an action head decodes action chunks at control rate (48 ms) purely in latent space without blocking. Enabled by a non-interfering attention mask that isolates video representations and staleness-robust horizon training that accommodates asynchronous lookahead aging, GlanceWAM breaks the speed-success dilemma. Trained purely on demonstrations, it attains 72.2% on the 24-task RoboCasa kitchen benchmark (vs. 67.1% for synchronous Cosmos Policy) and 99.0% on LIBERO while cutting per-chunk control latency $24\times$ relative to synchronous world-action models (48 ms on one A100). In single-arm and bimanual real-robot manipulation, it achieves higher average success than $π_{0.5}$ without any robot-data pretraining. Code is available at https://github.com/linhanwang/GlanceWAM.

Add real-robot experiments

Human--robot co-transportation of deformable objects requires predicting object deformation during motion, since obstacle clearance depends on both the grasp points and the unactuated interior. We present a hierarchical planning framework that combines a learned cloth model with a reduced-order whole-body model of a dual-arm mobile manipulator. A physics-residual conditional recurrent variational autoencoder (p-cRVAE) predicts the full cloth configuration from grasp-point observations by learning a residual correction to a computationally efficient linearized physics model, limiting error accumulation over 40-step planning horizon. The predicted cloth dynamics are embedded in a model predictive path integral (MPPI) planner using a reduced-order representation of a dual-arm mobile manipulator that preserves the non-holonomic base constraint and arm workspace limits. An MPC layer subsequently refines the sampled motion into smooth, executable references for whole-body control. The reduced-order formulation achieves tracking performance comparable to the full 17-DoF model while reducing computation time by approximately 80%. Across four co-transportation scenarios and two carrying speeds, the proposed framework maintains cloth-obstacle clearance where a corner-following baseline results in collisions, while whole-body refinement reduces final cloth deformation from 0.93,m to 0.28,m.

Action-conditioned robot world models must respond precisely to robot trajectories while preserving realistic visual dynamics, yet learning both from heterogeneous robot videos remains challenging. Simulation offers structured motion supervision, but appearance differences hinder direct transfer, and inaccurate simulation predictions can misguide real-video generation. We present SimForcing, a simulation-guided framework that uses simulation both as a source of transferable motion knowledge and as a controllable reference for prediction. First, we transfer motion knowledge from a simulation teacher through latent-motion distillation, aligning temporal changes in latent space to internalize motion priors while mitigating the influence of appearance differences. Second, we introduce multi-block simulation conditioning with condition dropout to exploit predicted simulation trajectories without relying excessively on their accuracy. Our simulation-conditioning classifier-free guidance scheme unifies these two ideas by balancing predictions based on internalized motion knowledge with those additionally guided by simulation latents. The jointly trained student generates both simulation conditions and real-domain videos, requiring no additional world model at inference. On Bridge, SimForcing achieves the best PSNR, SSIM, LPIPS, and FVD among the compared methods without external embodied pretraining. Evaluation on InternData-A1 further supports its applicability across robot datasets. Moreover, using our trained world model to initialize a vision-language-action model improves LIBERO success, suggesting its utility for downstream policy learning. \url{https://github.com/Wang-Xiaodong1899/SimForcing}

Code: https://github.com/Wang-Xiaodong1899/SimForcing

Vision-based autonomous navigation for low-cost robots remains a fundamental challenge, primarily due to the significant gap between simulated training environments and real-world operational conditions. Direct policy transfer from simulation is often ineffective, while training exclusively on real data is impractical. We propose a hybrid transfer learning framework that effectively bridges the sim-to-real gap by combining domain randomization with feature-level domain adaptation. Our method employs a dual convolutional variational autoencoder architecture with a shared decoder, trained on an extensive set of 45225 simulated images and a minimal set of only 4556 real-world samples. This architecture learns a compact, common latent representation space that aligns the distributions of both domains. The adaptation process is further enhanced by two complementary data augmentation techniques designed to expand the limited real-world data. Experimental evaluation demonstrates that our method achieves an average success rate of almost 91% on image classification tasks for real-world indoor navigation, significantly outperforming both simulation-only and real-world-only training. We validate these findings through a direct, real-world deployment, where the proposed policy successfully guides a low-cost robot in a reactive exploration task. Furthermore, we validate the model's efficiency through a rigorous computational estimation, confirming its suitability for resource-constrained embedded platforms such as the Raspberry Pi 4 and NVIDIA Jetson Nano. This work presents a practical solution for developing effective and efficient navigation policies for low-cost robotic systems.

30 pages, 13 figures. Published in Image and Vision Computing under a CC BY 4.0 license

General purpose language models are increasingly able to control robotic hardware. Understanding the capabilities and safety of these models when embodied is therefore increasingly important for understanding their societal impact and risks. To this end, we introduce Inspect Robots, a modular, open-source framework for developing and running evaluations of embodied agents. Inspect Robots pairs customizable, reusable abstractions for specifying physical evaluations and analyzing their results with infrastructure that automates evaluation setup, execution and termination. We demonstrate Inspect Robots by using it to evaluate the capabilities and safety of six policies based on frontier language models. Inspect Robots has seen significant early uptake, receiving nearly 100,000 downloads in the three months since its release.

Submitted to The Science of Physical AI Safety (SPAIS) Workshop at CoRL 2026

Robotic guide dogs should understand their surroundings, objects, and potential risks. Prior research has focused on raw sensor data from cameras and 2D or 3D LiDAR, which precisely measure distance points rather than provide a semantic understanding of the scene. While these physical measurements are effective for robot-centric collision avoidance and robot safety, they are not suitable for human-centric guidance. The system should recognize the type and relevance of obstacles and explain them, clearly and actionably, in terms of their spatial relation to the user. We present complete on-device perception modules that fuse a 360 camera and a 2D LiDAR for reliable collision avoidance, with moving-object detection and tracking for human-centric guidance. Finally, in walking-impossible situations, a vision--language model delivers pathway explanations as a safety mechanism to reduce user anxiety. In experiments, verification of fused 360 camera--LiDAR depth shows reliable near-range perception but inherent mid-range bias, while the system as a whole sustained real-time performance under 55 W. On the real-world egocentric GuideDogQA benchmark, our system achieved 83.8% accuracy, compared with 67.1% for GPT-4o. These results demonstrate that practical human-centric guidance with real-time on-device inference is feasible even on quadrupeds.

accepted in ACCV 2026

GAMBIT is an opening chess move in which a player sacrifices a piece, typically a pawn, to gain a positional advantage later in the game. Analogously, in multi-robot coordination, individual robots may need to forgo locally reward-maximising behaviours to improve overall team performance. Such self-sacrificial behaviours are difficult to capture with manually designed heuristics, particularly in dense, interaction-rich environments. Focusing on double-integrator continuous dynamics, this work studies how to learn such coordinated heuristics over motion primitives for multi-robot trajectory execution. Our framework, GAMBIT, first learns coordinated motion-primitive selection through imitation learning and subsequently fine-tunes the policy through reinforcement learning. We further introduce a safeguarded rollout mechanism with backup trajectories that guarantees collision-free execution at all times. Experiments demonstrate that GAMBIT substantially outperforms a range of baselines, including centralised motion planners and decentralised reactive planners, while exhibiting strong scalability. In particular, it coordinates over a thousand robots with planning latency below a few hundred milliseconds in continuous domains.

Instruction following is central to language-conditioned robot policies: language should determine what to do when the same scene permits multiple valid actions. Yet successful execution alone cannot establish whether a policy follows the instruction or infers the task from the scene. We study this ambiguity through scene-preserving instruction interventions, using valid target substitutions, arbitrary nouns, and unrelated sentences while holding the scene fixed. We evaluate vision-language-action (VLA) policies and world-action models (WAMs) in simulation and in real-world experiments. Our analysis addresses three questions: (a) Does task success imply instruction following? When instructions request a different visible object, all evaluated policies predominantly approach and pick up the incorrect original target associated with the scene. (b) Is this failure caused by language insensitivity? Instruction perturbations affect task performance. A layerwise action lens shows intermediate action predictions respond to these perturbations. Linear probes accurately recover instructed targets, indicating modified instructions are encoded despite rarely determining target selection. (c) Why does encoded language fail to control action? Attention analysis indicates weak instruction-token contributions to action generation. Target-token attention can remain focused on the original object, revealing a mismatch between target encoding and visual grounding. UMAP and shared non-negative matrix factorization show target information remains accessible within representations increasingly organized by scene identity. Our findings expose a grounding gap concealed by nominal success and provide a diagnostic framework. They further establish a concrete criterion for progress: policies should reliably follow valid changes in user intent, even when they conflict with scene-favored behavior.

Expressive humanoid interaction requires speech, facial animation, and body gestures to form a coherent response. However, many full-body humanoid robots produce speech and gestures without a visually expressive face, while talking-face animation and robot gesture generation are typically developed separately. We present Talk, Render, Act (TRABot), an agent-based framework comprising specialized agents for motion-atom construction, dialogue generation, motion planning, and facial animation. First, to produce natural and semantically meaningful gestures, we construct Robot-Ready Semantic Motion Atoms by segmenting long-form, G1-retargeted BEAT2 motion into units with natural gesture boundaries, human-verified communicative functions, and feasible trajectories. Second, to preserve semantic order and coordinate body motion with the spoken response, we introduce a Semantic-Conditioned Compositional Planner. Given an ordered semantic function sequence and an estimated response duration, the planner selects approved atoms to realize the longest feasible action sequence while accounting for transitions and neutral recovery. Finally, we deploy a Streaming Face-Speech-Body Integration system on a physical G1 humanoid, combining streaming dialogue audio, audio-driven facial animation, and semantically planned body motion in a unified real-time interaction loop. Quantitative and qualitative experiments demonstrate that TRAbot achieves the best overall performance among all compared conditions in terms of naturalness, expressiveness, and multimodal coherence.

Human videos offer scalable manipulation data, but the embodiment gap between human hands and robot manipulators limits their direct use. Existing video-editing methods replace hands with rendered robots, yet inaccurate interaction reconstruction and compositing can produce inconsistent grasps and implausible robot-object occlusions. We address these failures from two complementary physical aspects: interaction geometry and scene visibility. First, an interaction-aware contact reconstruction module combines hand-object segmentation with mesh-level contact prediction to recover dense 3D contacts, then converts them into temporally stabilized grasps for parallel-jaw grippers. Second, a depth-aware compositing module uses scene and robot depth to enforce physically consistent robot-object occlusions. The resulting videos preserve the interaction structure of human demonstrations in a robot-compatible form and are co-trained with robot demonstrations. Using identical human videos and robot data, we compare against robot-only training and the original Masquerade pipeline. Across four RoboTwin tasks and two Diffusion Policy visual encoders, our method achieves the highest average success rates, with especially strong gains under out-of-distribution scene variation. Real-world deployment further shows that the proposed co-training approach improves robustness to visual distractors when the task geometry is observable, while performance on depth-sensitive grasps remains limited by the single-camera setup.

Large Language Models (LLMs) can contribute useful engineering knowledge to robot design, but directly generated designs may rely on implicit assumptions and provide no guarantees of feasibility or optimality. These assumptions are critical because different reasonable modeling choices can materially change which designs are predicted to be feasible or optimal. We therefore present a framework that uses LLMs to construct explicit engineering models containing physical relationships, compatibility constraints, and objectives, allowing these modeling choices to be inspected and revised before formal optimization. The model can then be updated with additional engineering, manufacturer, or system-specific information before formal multi-objective optimization provides feasibility and Pareto-optimality guarantees with respect to the finalized model and specified design space. We evaluate the framework on quadcopter and line-following robot component-selection problems. Across 30 direct LLM design trials, none could be verified as feasible under the corresponding finalized model. Comparisons with an independently developed expert model and successive stages of model refinement further showed that changes in modeling assumptions substantially altered the predicted feasible and Pareto-optimal design sets. Together, these results show that using LLMs to construct explicit engineering models makes the underlying design choices available for inspection and revision before those assumptions determine the optimized designs. Explicit modeling therefore provides an interface for combining LLM-generated engineering knowledge, system-specific information, and formal design optimization.

Continuous asynchronous replanning is essential for real-time generative robot policies, but independent stochastic initialization can cause mode switching and inconsistent continuation across action chunks. We propose Execution-Aligned Progressive Noise (EAPN), which introduces structured stochasticity at both inter-chunk and intra-chunk levels. Across replanning steps, EAPN propagates a shared noise trajectory and aligns it with the actual execution displacement, establishing execution-aligned inter-chunk correlation. Within each action chunk, it models temporal correlation along action time. The aligned stochastic history is further combined with committed action context to condition subsequent generation, allowing new chunks to continue from execution-consistent generative states rather than restart from independent noise. We evaluate EAPN on D3IL, Kinetix, LIBERO, and real-world manipulation tasks. EAPN improves multimodal behavior consistency on D3IL and achieves an average success rate of 88.59% on Kinetix. On LIBERO, it remains robust and maintains strong task performance even under long inference delays. Real-robot experiments further achieve 90.0% success on Object Storage and 96.7% on bimanual Cloth Folding, demonstrating reliable continuous execution under asynchronous replanning.

8 pages, 7 figures, 5 tables. Project page: https://embodied.magiclab.top/works/eapn/index.html

Diffusion- and flow-based robot policies have recently become widespread in robotic Imitation Learning (IL) due to their high performance and ability to model continuous and multimodal distributions. However, the iterative denoising procedure used by these models introduces significant prediction latency, hindering high-frequency closed-loop robot control and leading to jittery, unstable motion when frequent updates to the robot's action predictions are used. Therefore, it is common practice to train models to predict chunks of actions that can be executed sequentially without feedback, even when this reduces responsiveness and may mean the most recent state information is not used. In this article, we present Adaptive Mean Flow (AMF), a flow-based IL method that enables smooth and responsive, fully closed-loop robot control. AMF uses Mean Flow, which is an accelerated form of Flow Matching (FM), to minimize prediction latency. To ensure smoothness and consistency across predictions, AMF uses a corrupted version of the trajectory from the previous step when predicting new robot actions, with the signal-to-noise ratio increasing over the time parameter of the trajectory. This discourages large changes in the prediction from one step to the next, while allowing freedom to adapt the predictions for future steps. We evaluate AMF across a wide range of simulated and real robot tasks and demonstrate significantly improved performance compared with baselines. Code: https://github.com/akselva/Adaptive-mean-flow-RoboticIL.

Ultrasound visual servoing is essential for autonomous robotic ultrasound, yet 6-DoF probe control from 2D B-mode images remains challenging due to limited and ambiguous out-of-plane motion cues. Existing methods typically rely on anatomical priors or handcrafted visual features, limiting their generalizability across imaging targets. Inspired by trackerless 3D ultrasound reconstruction, we propose Recon2Servo, a visual servoing framework that learns image-to-motion inference directly from B-mode images for 6-DoF probe control. A DINOv3 encoder with low-rank adaptation and a bidirectional relation module estimate the relative probe pose between current and target images to guide iterative closed-loop target-view alignment. The framework combines supervised relative-pose learning, reconstruction-guided closed-loop adaptation, and bounded residual pose correction to improve motion inference during servoing. Evaluations on a public dataset and an in-house dataset collected from 12 healthy volunteers using different ultrasound systems demonstrate its effectiveness in reconstructed-volume servoing. Additional real-robot demonstrations of target-view alignment and dynamic tracking on a human forearm are provided in the supplementary video: https://youtu.be/qwsOdI-GMYk.

Embodied co-painting requires a robot to repeatedly update a shared physical canvas while human intent evolves over interaction. Existing reference-driven painters or reactive assistants are typically optimized for single-shot rendering or sketch completion, limiting their ability to sustain coherent multi-round collaboration or to construct complex, content-rich scenes over time. We present CoBrush, a hierarchical framework that formulates multi-round co-painting as a coordinated semantic, spatial, and execution process. By separating high-level intent inference from spatial grounding and stroke-level control, the system supports progressive scene development on real acrylic canvases. We evaluate the framework through real human-robot painting sessions, stress tests, and user studies. Compared to single-turn baselines, our approach achieves stronger semantic alignment, more stable spatial progression, and higher perceived plausibility of robot actions. These results demonstrate that structured multi-stage reasoning improves the coherence and robustness of interactive painting and supports the progressive development of content-rich physical artworks.

Accepted to the 2026 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2026)

Robot delivery studies can overstate transport capacity when travel to pickups and human support fall outside the modeled schedule. We formulate a location aware dispatch model that couples robot admission to transporter support and shared elevators, charging, and cleaning. The model replays 33,079 observed hospital requests; missing contents, deadlines, staffing, and completion times remain explicit scenario assumptions. Two full grids compare human dispatch, a resource aware controller, and a proximity and workload benchmark in 30 paired replications per scenario. An exploratory extension tests a simpler deadline admission rule under the same operating model. Pickup travel increases demand on a pooled elevator bank. In one high staffing case, raising assumed elevator capacity from two to four reduces human only lateness from 55.11% to 3.18%, exceeding the dispatch differences. Direct policy comparisons and robot coverage distinguish admission selectivity from system service. The analysis explains why a robot can pass a deadline admission test yet delay service through staff handoffs and shared facilities. Its contribution is a reproducible evaluation of coupled dispatch workflows, with a clear separation between admission, completion, and return to availability. The results are conditional comparisons, not estimates of hospital benefit or released clinical capacity.

Natural face-to-face human--robot interaction requires a robot to understand an evolving social situation, decide when to engage, and express its intent through coordinated physical behavior. Yet existing approaches rarely close this loop: foundation-model agents provide increasingly capable multimodal reasoning and memory but remain largely disembodied, while expressive virtual agents do not face the physical constraints of real robots, and physical social robots typically address social reasoning and embodied expression only partially. We present ARISE, a unified framework that bridges Agentic Reasoning and Interactive Social Embodiment on the Sophia humanoid robot. ARISE integrates multimodal context understanding, long-term memory, and reactive and proactive interaction to determine when and what to communicate, and translates social intent into robot-native gestures coordinated with speech and mechanical facial expressions through streaming execution. Extensive evaluations on Sophia demonstrate strong perceived interaction quality, expressive and well-coordinated embodied behavior, and substantial latency reductions through streaming execution. These results highlight the importance of jointly reasoning about what to communicate, when to engage, and how to physically express social intent for natural interaction with humanoid robots. Project Page: https://robosocial.github.io/

Learning from human demonstrations is a reliable way to teach robots new tasks, but the gains from each additional demonstration shrink as the policy improves. Continued improvement can instead come from supervised deployment, where an operator places the objects and intervenes when the policy fails. We ask how to maximize improvement from a fixed budget of supervised episodes on high-precision manipulation tasks with wide ranges of object placements. We observe that failures can concentrate in a small subset of initial states, so uniform collection spends much of the operator's time on states the policy already handles. Mulligan makes the initial-state distribution a decision, starting each round's episodes at observed failures and untried states. To further improve data efficiency, we augment interactive imitation learning with a value function trained on all data, including failures that imitation discards. Across three real-world tasks evaluated on 2,550 held-out, blinded episodes and two simulated tasks, Mulligan outperforms uniform initial-state sampling at matched collection budgets, and combined with value-based action selection, HiL-IDQL+Mulligan, improves final real-task success by 10-34 percentage points. With operator interventions, the human-robot team completes 98% of collection episodes, remaining productive while the policy learns. Videos, code, and data are available at https://mulligan.page/.

Project page: https://mulligan.page/

Robots operating in human-centered environments must safely navigate among multiple dynamic obstacles to avoid collisions with people and surrounding infrastructure. Control barrier functions (CBFs) provide an effective mechanism for safety filtering, and recent CBF-based reinforcement learning (RL) methods embed such safety information into learned policies. However, executing only safety-filtered actions during training can restrict policy exploration, a limitation that becomes particularly consequential in dynamic scenes where safety depends on relative robot-obstacle motion. We propose DODGER, a safety-guided RL framework that directly executes policy-generated actions to drive training rollouts while using CBF-filtered references and constraint violations to shape the policy toward collision-avoidance behavior. We evaluate DODGER through a Dubins-car safety analysis and demonstrate goal-directed navigation among multiple dynamic obstacles in full-order humanoid simulation and real-world humanoid experiments using LiDAR-based perception, without a runtime safety filter.

The first three authors contributed equally to this work. Project Page: https://psh0823.github.io/dodger-homepage

Robotic ultrasound couples a high-rate contact-force loop with slower, delayed image feedback, so image-guided ultrasound probe rotation can perturb contact force before the resulting image response is observed. We derive a closed-form orientation-rate limit that bounds the modeled rotation-induced estimated-force excursion over a finite horizon while accounting for disturbance rejection by the fast force loop. The limit depends on local contact stiffness, force-loop gains, a conservative rotation-to-force gain bound, the excursion budget, and the prediction horizon. We implement this model in a dual-rate controller with timestamp-based delay reconstruction and joint-torque-based force estimation, and evaluate it on a curved gelatin phantom using paired controller comparisons and component ablations. Relative to unconstrained image guidance, the proposed rate-limited controller reduced first-second root-mean-square (RMS) estimated-force error by 0.40 N while increasing cue-convergence time by 0.94 s. A fixed rate cap near the analytically predicted ceiling produced no resolvable difference in force error and converged 0.32 s faster, indicating that the principal practical value of the model is the rate-design rule rather than online prediction. Delay reconstruction had no resolvable effect at the tested latency. A single-subject popliteal scan demonstrated feasibility, although the image cue was noise-limited on heterogeneous tissue.

8 pages, 4 gigures, conference

Transporting unsecured payloads with legged robots over uneven terrain requires balancing locomotion performance and payload stability, since aggressive motion can destabilize the payload even when the robot remains stable. We study quadrupedal transportation of unsecured stacked boxes on an edgeless torso-mounted board without dedicated payload sensors or active carrier mechanisms. To address this trade-off, we propose Payload-Adaptive Multi-Objective Reinforcement learning for Transportation (PAMORT). PAMORT trains a multi-objective base policy conditioned on a preference vector that weights locomotion and payload-stability reward groups, then trains a weight adjuster on the frozen policy to adapt this preference online from proprioception. In simulation, PAMORT achieves comparable or better overall transportation success than a corresponding single-objective baseline across different payload configurations, including an unseen three-box stack, despite training only with two boxes. Real-world experiments on a Unitree Go2 demonstrate zero-shot transfer to slopes and steps at or beyond the training difficulty, with mean success rates of 0.850 for PAMORT and 0.675 for the baseline across eight tasks. These results demonstrate robust unsecured-payload transportation with online adaptation of the locomotion--payload trade-off from proprioceptive information.

Vision-language-action models (VLAs) act on instructions without being able to refuse, so screening harmful requests falls to monitors. We test whether these monitors catch ordinary robot tasks requested for harmful reasons, holding the task fixed while varying only how explicitly the intent is stated. $π_{0.5}$ completes the task at every level of explicitness, as often as for harmless controls. Text guards flag nearly every blunt request but few implied ones: up to 95% of implied-harm runs end with the task done and no flag raised, and up to 90% even after recalibrating on robot instructions. Monitoring the model's activations does not close this gap. Linear probes separate harmful from harmless instructions almost perfectly in the base language and vision-language models, but this separation weakens after robot training in two model families, most for implied harm.

Under review at the SPAIS 2026 workshop (CoRL 2026)

Robotic minimally invasive surgery offers well-documented clinical benefits, but the cost and infrastructure requirements of purpose-built platforms limit access in rural and lower-resourced facilities. Recent work has teleoperated general-purpose robots for laparoscopic tasks and in vivo procedures, but relied on handheld instruments coupled through passive linkages rather than native robotic actuation. Instead, we adapt da Vinci Classic and Xi instruments onto general-purpose robots that can integrate in clinical workflows. We present DASH, a da Vinci Adapter for Serial-link and Humanoid platforms, consisting of two types of adapters that require no modification to the instruments themselves. Each adapter is compatible with many robotic platform load capacities, engages the instrument's native latch, and wirelessly identifies inserted tools to load instrument-specific kinematics and coupling matrices. Teleoperated experiments demonstrate the efficacy of DASH and its benefits for enabling research access to surgical robotic platforms.

Online control of execution speed is essential for deploying robot policies in real-world scenarios, as robots may need to speed up under time constraints or slow down to facilitate human interaction and improve safety. However, imitation-learned policies inherit the execution speed of their demonstrations, and test-time speed modification can introduce unrecoverable out-of-distribution observations, reducing task success. We observe that the directional inconsistency of action chunks reflects task-phase criticality, indicating how aggressively action step lengths can be modified while preserving task success. Based on this observation, we introduce FreeSpeed, a training-free module that post-processes action chunks from pretrained policies. FreeSpeed resamples each predicted chunk at the requested rate, then uses directional inconsistency between adjacent actions as the primary signal for rescaling. This signal adaptively determines how closely the execution speed can approach the requested speed, allowing flexible speed adjustment within the evaluated limits without compromising task success. Across three policy families and 50 simulated tasks, FreeSpeed supports online speed changes, with realized execution rates spanning 0.22x to 2.53x among settings that preserve per-task success. Across four real-world manipulation tasks, FreeSpeed achieves an average success rate of 94.0%, matching the frozen policy's 93.8%, while realizing execution rates from 0.38x to 1.97x.

37 pages, 21 figures, 14 tables. Project page: https://yuxuanhu9.github.io/FreeSpeed/

Emergency obstacle avoidance requires a mobile robot to brake or change its heading within the available distance. This paper investigates the dependence of maneuver performance and odometric error on approach speed for a differential-drive robot. A footprint-clearance analysis and a normalized wheel-motion index provide a kinematic description of five braking and turning maneuvers. The principal experiment comprises 540 block-randomized trials on tile, of which 539 are retained, at commanded approach speeds of 45, 55, and 65 cm/s. An overhead camera provides an independent pose reference. When encoder, gyro, and camera heading changes are evaluated over complete motion records, the mean encoder discrepancy increases by 4.7-5.1 degrees for the two reverse-spin maneuvers between the lowest and highest speeds; the arc and brake-assisted pivot change by less than 0.4 degrees. Larger encoder discrepancy is associated with lower avoidance success after adjustment for distance, speed, maneuver, day, and turn direction. Gyro discrepancies remain approximately 3 degrees for the reverse spins. Estimated reaction distances for 90% success and their uncertainty quantify the maneuver trade-offs. The results support speed-specific empirical characterization of emergency maneuvers and distinguish encoder error from recovery motion and measurement-window mismatch.

Autonomous robots deployed in mass casualty incidents (MCI) face the challenge of making critical decisions based on incomplete and noisy perceptual data. We present an autonomous robotic system for casualty assessment that fuses outputs from multiple vision-based algorithms, estimating signs of severe hemorrhage, visible trauma, or physical alertness, into a coherent triage assessment. At the core of our system is a Bayesian network, constructed from expert-defined rules, which enables probabilistic reasoning about a casualty's condition even with missing or conflicting sensory inputs. The system, evaluated during the DARPA Triage Challenge (DTC) in realistic MCI scenarios involving 11 and 9 casualties, demonstrated a nearly three-fold improvement in physiological assessment accuracy (from 15% to 42% and 19% to 46%) compared to a vision-only baseline. More importantly, overall triage accuracy increased from 14% to 53%, while the diagnostic coverage of the system expanded from 31% to 95% of cases. These results demonstrate that integrating expert-guided probabilistic reasoning with advanced vision-based sensing can significantly enhance the reliability and decision-making capabilities of autonomous systems in critical real-world applications.

Reinforcement learning has shown strong performance in robotic manipulation, but learned policies often degrade in performance when test conditions differ from the training distribution. This limitation is especially important for reliable robot planning and control in contact-rich tasks such as pushing and pick-and-place, where changes in goals, contact conditions, or robot dynamics can drive the system out-of-distribution at inference time. In this paper, we investigate a hybrid controller that combines reinforcement learning with bounded extremum seeking (ES) to improve robustness under such conditions. In the proposed approach, deep deterministic policy gradient (DDPG) policies are trained under standard conditions on the robotic pushing and pick-and-place tasks, and are then combined with bounded ES during deployment. The RL policy provides fast manipulation behavior, while bounded ES ensures robustness of the overall controller to time variations when operating conditions depart from those seen during training. The resulting controller is evaluated under several out-of-distribution settings, including time-varying goals and spatially varying friction patches. Crucially, under reasonable assumptions, we prove finite time RL-based convergence of a robotic arm to an object after which bounded ES achieves finite-time convergence to a bounded velocity moving target. We provide analytic formulas for both convergence times, and an ability to tune the convergence times by our choice of gains.

Efficient and robust tracking of surgical robotic instruments is important for robot-assisted minimally invasive surgery, yet remains challenging due to the complexity of surgical scenes and the unconventional geometry of surgical instruments. Keypoint-based approaches are efficient, but their performance depends on reliable feature detection. Improving these detectors with real-world supervision is difficult because accurate real-world annotations are costly to obtain at scale. To address this limitation, we introduce a tracker-guided self-training framework that adapts a model pretrained on synthetic images to unlabeled real-world videos. Given measured robot joint states, an uncertainty-aware EKF recursively corrects the instrument pose and the observable joint angles by comparing projected model features with detected keypoints, shaft boundaries, and mask-derived cues. An RTS smoother subsequently refines the resulting trajectory, which is projected into pseudo-labels for fine-tuning the feature detector without laborious pose annotations. Experiments on real-world videos demonstrate consistent improvements from self-training across all evaluated keypoint metrics, and the resulting model outperforms prior approaches in both accuracy and runtime. The code and data will be released upon publication.

We propose a force-based model for social navigation of a tour-guide robot. Social forces due to various factors like obstacles, user position and heading, have been accounted for in the model. We claim that each one of these forces makes the robot more sociable to the user and we design an experimental setup for evaluation. In the experiment, the user follows an autonomous robot to a destination in a known map, while undertaking a few simple sub-tasks in the middle, which serve as distractions. For each participant, we run several rounds of the experiment, each with different forces and a shortest path, A*-search baseline model. Using per-round subjective indicators, we propose to study the effect of our force model on constructs such as: follow-ability, perceived safety, and perceived intelligence.

5 pages, 8 figures. Accepted at the 5th Workshop on Social Robot Navigation, IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) 2026

We use fused deposition modeling (FDM) 3D printing as a case study of how manufacturing robots can use imperfect AI. In FDM, print configuration strongly affects output quality. Yet, novice users typically rely on default configurations, trial-and-error, or direct recommendations from generic AI models (e.g., ChatGPT). These strategies can produce complete prints, but they do not reliably meet specific objectives. We present a modular approach that treats an LLM as a source of tuning expertise. We embed this source of expertise within a Bayesian optimization loop. An approximate evaluator scores each candidate print configuration and returns structured diagnostics, which the LLM uses to propose natural-language adjustments that are compiled into machine-actionable guidance for optimization. On 100 Thingi10k parts, our LLM-guided loop achieves the best configuration on 78% objects with 0% likely-to-fail cases, while direct AI model recommendations are rarely best and exhibit 15% likely-to-fail cases. These results suggest that LLMs provide more value as constrained decision modules in optimization loops than as end-to-end oracles for print configuration selection. We expect this result to extend to broader LLM-based robot programming.

Accepted at IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) 2026

Most vision-language-action (VLA) models rely on current observations and lose task-relevant evidence once it leaves view, limiting performance on long-horizon, memory-dependent tasks. Existing efforts incorporate compressed historical features or sparse visual keyframes. However, isolated snapshots can leave the policy uncertain about what changed during past interactions and which action should follow. To overcome this limitation, we propose EvoMem-VLA, which constructs state-evolution memory by explicitly encoding and retaining observed changes between historical states. These change representations preserve evidence of interaction outcomes, allowing the policy to track task progress beyond isolated snapshots. Specifically, we introduce conditional delta tokenization to encode ordered frame pairs into directional, source-conditioned delta tokens, each associated with its corresponding state evidence. A shared VLM backbone supports task-adaptive routing: normal long-horizon tasks follow a direct action route, whereas multi-stage tasks use a subtask route that generates an executable subtask as an additional input for action generation. With a single jointly trained policy for each simulation benchmark, EvoMem-VLA achieves success rates of 80.7% on RMBench, 82.0% on RoboMME and 83.8% across four real-world tasks spanning two robot embodiments. These results represent substantial improvements over the previous state of the art in all three evaluation settings.

Although the advancement of vision-language models (VLMs) has endowed robots with enhanced environmental understanding and task reasoning, a comprehensive evaluation methodology is important to advance the integration of VLMs in robotic navigation and manipulation. However, current benchmarks lack a comprehensive method to evaluate diverse robotic tasks, and evaluation metrics remain relatively constrained, making it difficult to assess the embodied capabilities of VLMs in a thorough and fine-grained manner. To address this issue, we propose RMMBench, an evaluation benchmark that requires robots to understand language instructions and perform long-horizon tasks in continuous spaces. RMMBench seamlessly integrates high- and low-level embodied tasks into a unified framework, constructing a "navigation-manipulation" task suite comprising 70 canonical task scenarios that range from localized manipulation to long-horizon composite navigation. The results reveal that leading VLMs still face major challenges in spatial localization when performing mobile manipulation tasks, and also highlight the necessity of enhancing the spatial perception capability of robots during long-horizon interactions. RMMBench can be accessed at https://mxxq-stack.github.io/rmmbench-project/

Accepted to the Conference on Robot Learning (CoRL 2026), November 9-12, 2026, Austin, TX, USA

Simulation demonstrations can supplement scarce real-world data for robot policy co-training. However, the value of using data curation to actively select these demonstrations for sim-to-real co-training remains underexplored. Existing curation methods also lack a unified criterion for measuring trajectory-level utility and set-level coverage from closed-loop target behavior. To address these gaps, we present the first systematic study of data curation for sim-to-real robot policy co-training and propose Trajectory-level Utility and set-level Coverage Optimization (TUCO). TUCO uses influence functions to trace how each source demonstration affects target-domain scoring rollouts. Our key insight is that these effects can be decomposed into an overall contribution to target return and variation across rollouts, providing a common closed-loop basis for measuring trajectory utility and set coverage. We further propose a performance-aligned subset optimizer that combines these measures in a unified curation objective to reduce redundancy and select complementary demonstrations. Extensive experiments on RoboMimic and OmniReset establish the value of active simulation data curation for sim-to-real policy co-training and show that TUCO achieves state-of-the-art performance across single-simulator, sim-to-sim, and sim-to-real settings.

This study proposes a reconfigurable rocker-bogie mechanism that achieves efficient turning motion with a small number of actuators while maintaining high step-climbing capability. By installing motors at the bogie joints and actively swinging up and down bogies, the system enables switching between four-wheel and six-wheel configurations. Omnidirectional wheels are mounted on the rear ends of the rockers, allowing smooth turning in the four-wheel configuration based on a differential-drive model. Experimental evaluation using a prototype robot demonstrated that the proposed mechanism achieves zero-radius turning at a speed more than five times that of a conventional rocker-bogie mechanism equipped with six non-steerable grip wheels, while requiring only approximately 17% of the total average wheel torque. In addition, the robot successfully climbed a 40 cm step with an average climbing time of 6.4 s, confirming its high turning and step-climbing performance.

Accepted for publication in the Proceedings of the IEEE/ASME International Conference on Advanced Intelligent Mechatronics (AIM 2026)

Ergodic control drives robots to spend time in each region in proportion to a spatial distribution of interest, making it well suited for dense spatiotemporal environmental monitoring. Existing safe ergodic controllers rely on offline trajectory optimization or hierarchical architectures, which limit real-time applicability and decouple the ergodicity objective from the safety constraint. This paper presents a quadratic programming (QP)-based controller that treats ergodicity and safety jointly. We first introduce a Gaussian-kernel ergodic metric that, unlike the classical indicator-based metric, is time differentiable. This allows the exponential decay of the metric to be imposed as a time-varying control barrier function (CBF) constraint, relaxed by a slack variable, alongside hard CBF constraints for region containment and inter-robot collision avoidance. We establish that the resulting QP remains feasible from any safe initial configuration and that, up to the slack term, the ergodic metric decays exponentially. Simulations show improved performance over baseline methods, and experiments with three aerial vehicles validate the work on real hardware.

We propose a hierarchical control framework to address severe dynamic instabilities and navigational drift that arise when a two-wheeled inverted pendulum (TWIP) robot attempts asymmetric object manipulation. While two-wheeled platforms are highly manoeuvrable, their constant balancing adjustments make onboard odometry highly unreliable for precise navigation. Furthermore, the addition of a side-mounted robotic arm introduces unactuated lateral roll moments when a payload is lifted, a challenge heavily compounded on uneven terrain. To solve these coupled problems, our architecture divides the workload. An offboard vision system tracks overhead ArUco markers to provide high-latency global waypoint navigation, bypassing odometry drift. Simultaneously, a low-latency onboard control loop rejects active physical disturbances using inertial and encoder data. In our physical experiments, this dual-loop approach enabled the custom-built robot to navigate accurately, reject transient impacts from speed bumps, adapt to a dynamic seesaw ramp, and carry a payload securely without falling over its narrow wheelbase.

6 pages, 9 figures, 2 tables. Accepted for presentation at the 2026 IEEE International Conference on Intelligent Navigation and Systems (CINS), Dubai, UAE

Equipping humanoid robots with coherent and adaptable personas is crucial for fostering natural, engaging, and trustworthy human-robot interaction (HRI). However, existing approaches often rely on static, hard-coded identities that lack the flexibility to adapt to individual user contexts. In this paper, we present PACE (Persona Adaptation through Conversational Elicitation), a novel framework for the interactive generation and deployment of structured personas on the Ameca humanoid robot. Our system introduces an Interactive Persona Elicitation Pipeline, enabling the robot to dynamically synthesize a tailored, psychologically grounded identity through user Q&A. This elicitation process feeds into a persona prompt compilation phase, generating a structured persona prompt built upon multi-perspective dimensions. We detail the Embodied System Integration required to translate this structured specification into expressive, multimodal humanoid behaviors. Through a comprehensive empirical HRI evaluation, we assess the impact of dynamically generated personas on user trust, perceived anthropomorphism, persona consistency, personal relevance, and interaction quality compared to a generic baseline. These contributions establish a scalable pathway for deploying personalized, interactive, and reliable identities in embodied humanoid assistants. Video demo is available at: https://lipzh5.github.io/PACE/

Given a set of point robots $\mathcal{R}$ in the Euclidean plane and a target circle $\mathbf C$ enclosing all robot positions, the \textsc{Min-Max Uniform Circle Formation (MMUCF)} problem requires the robots to move to distinct positions on $\mathbf C$ such that the final configuration forms a regular $n$-gon while minimizing the maximum distance traveled by any robot. Uniform circle formation is a fundamental coordination task in swarm robotics with applications in perimeter monitoring, surveillance, boundary coverage, and pattern formation. The literature does not address the optimization of the maximum individual displacement during the formation process. In this work, we study the min--max versions of the circle formation and uniform circle formation problems, where the goal is to minimize the maximum distance traveled by any robot. We consider these problems under the $\mathcal{ASYNC}$ model, where robots are autonomous, anonymous, identical, homogeneous, oblivious, and silent, and operate under the \textit{Look--Compute--Move} model with non-rigid motion. We first give necessary conditions for a deterministic solution and then present deterministic, distributed, and collision-free algorithms that form a circle and a uniform circle in finite time while minimizing the maximum movement. The algorithms ensure that robots reach distinct positions on the circle and, in the uniform case, equally spaced positions on $\mathbf C$ under the considered model.

Generative robot policies trained on demonstrations using behavior cloning often learn actions that are sub-optimal or misaligned with respect to the downstream task. Policy improvement approaches aim to bridge this gap and improve the cumulative reward with minimal interventions on the pre-trained policy. However, an impediment to their practical deployment is the requirement of cumbersome hyperparameter tuning specific to model or task families. In this work, we present an episodic, derivative-free latent policy improvement approach that works with little to no changes across model scales from MLPs to VLAs. We demonstrate that the performance of a pretrained, frozen diffusion or flow matching policy can be improved with respect to a downstream reward by swapping the sampling of initial noise from the prior distribution (typically isotropic Gaussian) with a well-chosen, constant initial noise input---a golden ticket. We show the prevalence of golden tickets by improving the policy performance of $46$ out of $51$ tasks across manipulation benchmarks, with absolute improvements in success rate by up to $79%$ for simulated tasks, and $28%$ within $60$ search episodes for real-world tasks. Further, we find that the versatility of our approach opens up new avenues such as simultaneous policy improvement for multiple downstream rewards. Project webpage: https://lottery-tickets.rai-inst.com/.

34 pages, 14 figures, 12 tables

Learning-based models (e.g., visuomotor and Vision-Language-Action (VLA)) are increasingly explored for industrial robotic manipulation, where model predictions are directly translated into physical actions. This tight coupling between model behavior and physical execution makes hidden security vulnerabilities particularly consequential. While backdoor attacks have been widely studied in conventional AI models, their effects on deployed learning-based robotic arm manipulation systems remain less understood: a backdoored robot can behave normally during benign operation while inducing attacker-specified behaviors only when specific triggers are present, posing potentially serious risks in physical environments. In this work, we present a preliminary empirical security study of backdoor attacks and defenses in learning-based robotic manipulation on two real commercial industrial robotic arms (FANUC and xArm). We investigate whether a backdoor can reliably induce semantically incorrect manipulation behaviors while remaining stealthy under nominal task execution. We then develop an online defense pipeline that detects and neutralizes triggers at runtime, and compare its effectiveness against an offline fine-tuning defense. Beyond defense effectiveness, we further evaluate the computational latency and execution overhead introduced by the defense pipeline to assess its suitability for high-throughput industrial operation.

Accepted at the IROS 2026 Workshop on Industrial Applications of Robot Learning. Added the required Distribution Statement A for public release and distribution

In this paper, we present a distributed cascade force control system (DCFC) for multiple robots with the aim of pushing a rigid object towards a desired moving target without their inter-robot communication. These mobile robots are equipped with 360-degree vision-based soft tactile sensors utilized to determine contact location and resultant impact force. By investigating the dynamics of moving rigid objects on the flat, we proposed a distributed cascade control. The inner loop control incorporates contact force and positioning, ensuring the robots' pushing contact and application of the desired force to the object. The outer loop control coordinates the robots to push the object in a desired direction without inter-robot communication, regardless of the unknown object mass. The stability and convergence of the system are verified using the Lyapunov stability theory. We also conducted simulation and real-word experiments to validate the performance of the proposed control method, and the experimental results showcase the successful coordination of multiple robots in pushing an object towards a moving desired direction.

6 pages, 7 figures. Published in the 2024 IEEE/SICE International Symposium on System Integration (SII)

This letter proposes DNPush, a Distributed Nonparallel Pushing force controller for cooperative object transport without inter-robot communication. Each robot uses local force measurements, a preassigned force weight, and a locally expressed desired transport direction, without requiring an object model, contact locations, a CoM estimate, or moment arm magnitudes online. DNPush regulates nonparallel force magnitudes from local force direction misalignments through a reciprocal-sine law. Under the stated contact conditions, Lyapunov analysis with LaSalle's invariance principle proves convergence to a unique rotational equilibrium, where perpendicular force contributions cancel and the net torque from normal forces is zero. Object velocity converges to the desired transport direction, and suitable contact regions exist for convex polygons. Simulations isolate the reciprocal-sine scaling mechanism, five-robot coordination, and active rotational damping. Hardware experiments with soft tactile differential-drive robots achieve a velocity direction error of 0.61 deg +/- 0.79 deg over five trials with a fixed transport direction and demonstrate object transport through waypoints with three robots.

8 pages, 8 figures. Submitted to IEEE Robotics and Automation Letters

When a team of robots delegates planning and mutual trust to LLM agents, a single compromised robot can corrupt the shared outcome. We study this threat in a grounded task: a multi-robot survey in which measurements can be verified against the physical world, but every verification costs budget that would otherwise advance the mission. We treat the compromised robot as a stealthy adversary in the system-theoretic sense: it is limited not by an energy bound but by the team's own detectors. We then derive two bounds. First, the probability that the adversary's reports are verified is bounded below in terms of the degrees in the communication graph and the verification budget. Second, the map error caused by any stealthy adversary is bounded above by the value of a linear program over the adversary's bias distributions; its solution is an exchange rate between stealth budget and damage: below a critical verification level the worst stealthy attack tells rare, full-magnitude lies on the records least likely to be verified, and above it the better purchase is small biases hidden in the noise. In experiments where the honest robots are LLM agents, both bounds hold at the budget the attack actually spent. The experiments also show that which records an LLM robot re-checks is unbiased, but how much it re-checks is unpredictable.

Force-aware manipulation typically relies on specialized force or tactile sensors. We show that force-aware manipulation can instead be achieved through visual force prediction from the deformation of a compliant Fin Ray gripper. Our approach trains two models. First, we train a visual force estimator on calibration data and use it to annotate task demonstrations with force estimates. Second, we train an action--force proposal policy on these force-augmented demonstrations to jointly generate candidate robot actions and their associated forces. At test time, we sample candidate actions and the forces they are expected to produce, then execute the action whose predicted force is closest to a target from the demonstrations. We evaluate our approach on berry picking, empty-can grasping, in-hand reorientation, and plug insertion. Our results show that visual force prediction can guide inference-time action selection for contact-rich manipulation without requiring force or tactile sensors at deployment.

Memory is essential for long-horizon robotic manipulation, where successful actions may depend on past events that are no longer recoverable from the current observation. As episodes grow longer, however, retaining the full history becomes increasingly costly, creating a fundamental scalability challenge for memory-augmented policies. Existing approaches address this challenge by either storing selected past observations in a bounded memory bank or compressing interaction history into a fixed-size parametric state through Test-Time Training (TTT). Yet these formulations do not explicitly distinguish between historical information that can already be recovered from the policy's current context and information that must persist beyond it. We introduce TTT-RM, which repurposes TTT as Residual Memory, using fast weights not to generically compress history but to complement a bounded memory bank by preserving task-relevant historical information that cannot be recovered from the policy's current context. Concretely, TTT-RM learns a history decoder that reconstructs historical representations from the current observation and retrieved memory. The resulting reconstruction residual captures what this context fails to explain and serves as the learning target for TTT. The TTT slow weights are optimized against this residual target so that online fast-weight updates learn to encode complementary historical information over time. The fast-weight state is then queried to produce a residual memory representation that conditions action generation. Extensive experiments on memory-intensive simulation benchmarks and real-world tasks show that TTT-RM consistently improves across multiple memory designs, outperforms diverse baselines, and supports sustained execution on a three-minute, eight-stage stowing task.

Project page at https://hatchetproject.github.io/tttrm/

Vision-language-action (VLA) and world-action models (WAMs) often degrade under out-of-distribution task variations despite retaining partial task capability. To recover such capability, we propose RoboIRS, an inference-time internal representation steering method that uses successful and failed rollouts to train linear classifiers, select outcome-relevant intervention locations, and derive task-specific steering directions without updating policy parameters. On 15 simulation tasks with a frozen $π0.5$ policy, RoboIRS improves the average success rate from 44.4% to 66.2%, outperforming alternative inference-time intervention baselines while adding little inference time. We further validate RoboIRS on real-robot manipulation using the same $π0.5$ policy and demonstrate its applicability to a world-action model Cosmos Policy, where the average success rate improves from 35.4% to 55.4%. These results show that directly steering internal robot-policy representations can improve the performance of robot policies at inference time. Project website is available at https://rollingoat.github.io/roboirs/.

Soft robots present significant modelling challenges due to their non-linearity, complex dynamics and potentially intricate geometries. These difficulties in accurate system identification and dynamics modelling limit their applications in precise robotics tasks. Prior modelling approaches typically suffer from trade-offs in accuracy, computational efficiency, or speed. The differentiable finite element method (FEM) offers a promising balance between these desiderata for soft robot modelling, and enables the use of gradients for efficient calibration and trajectory optimisation. In this paper, we propose an end-to-end differentiable FEM-based framework designed to streamline modelling and system identification for pneumatic soft robots, enabling the generation of reliable models for downstream tasks. The framework automates the conversion of computer-aided designs into voxelised tetrahedral viscoelastic FEM meshes that accurately represent the robot's geometry and actuation mechanism. By integrating easily acquired point cloud data with differentiable FEM, the framework achieves precise identification of material and dynamic actuation parameters using a minimal experimental setup. Additionally, the differentiable nature of the model facilitates trajectory optimisation by leveraging gradients from robot dynamics and contact interactions. We validate the framework by modelling a complex bellow-shaped pneumatic soft robot and demonstrate its efficacy in real-world motion planning tasks. Experimental results indicate high modelling accuracy, with a maximum positional error of less than 3 mm, and successful application in tasks such as path following and grasping.

Published in IEEE Robotics and Automation Letters (RA-L)

Engaging K-12 students in authentic scientific research remains a significant challenge, particularly at the intersection of environmental science and robotics. We introduce the Jar Jar ROV, a low-cost, open-source Remotely Operated Vehicle (ROV) platform designed for citizen science-based water quality monitoring by middle school students. This paper presents the design of the platform and the results of a large-scale deployment with over 100 students across a US state who built, programmed, and deployed the ROVs in local lakes. The educational framework yielded high student engagement in hands-on activities, with ROV construction earning a perfect average score from mentors. Scientifically, the program established a grassroots monitoring network, generating nearly eleven thousand validated measurements of temperature, pH, dissolved oxygen, and turbidity. However, our evaluation identified a critical "engagement gap," with student interest declining sharply during more complex tasks such as electronics assembly and data uploading. This paper contributes both a validated, scalable model for integrating robotics into environmental education and a data-driven roadmap for future improvements. These enhancements focus on lowering technical barriers and creating a more intuitive link between data collection and scientific discovery, addressing a key challenge in empowering the next generation of citizen scientists.

Accepted to ICRA 2026

We introduce DreamFormer, a model-based agent that acquires language-conditioned, multi-task skills by imitating expert demonstrations within the latent imagination of a learned world model. DreamFormer first learns a task-agnostic Transformer world model from unstructured play data, then acquires task-specific behaviors by optimizing an intrinsic reward that aligns agent-generated rollouts with expert demonstrations in latent space. Since the policy is trained on-policy inside imagination, it is exposed to its own errors during training, mitigating the covariate shift inherent to offline behavioral cloning. To make long-horizon imagination affordable, DreamFormer encodes a high-resolution multi-view robotic observation into a single input token, avoiding both spatial downsampling and the multi-token representations used by prior Transformer world models. On the long-horizon CALVIN benchmark, DreamFormer outperforms LUMOS, the comparable model-based agent, on single-environment evaluation (2.52 vs 2.34 average tasks completed per chain of five) while keeping imagination rollouts tractable. Against HULC, the behavior cloning baseline, it nearly doubles performance on zero-shot transfer to an unseen environment (1.30 vs 0.67), indicating that dynamics learned by the world model transfer more readily than a directly cloned policy. This is consistent with the role attributed to internal models in biological agents, where a model of environment dynamics supports behavior in situations not previously encountered.

submitted to IEEE Transactions on Cognitive and Developmental Systems (TCDS)

Robotic long-horizon manipulation requires robots to compose perception, reasoning, and action over extended task sequences, yet existing language-conditioned frameworks often rely on dense step-wise feedback, learned action policies, or unstructured prompting, which limits robustness and generalization. We propose DAHLIA, a data-agnostic code-generation framework that treats long-horizon manipulation as episodic task planning and evaluation. DAHLIA uses progressively organized in-context examples with chain-of-thought reasoning to guide an LLM in synthesizing executable code plans from language instructions. To improve robustness under partial observability, a VLM-based reporter evaluates task outcomes only after each execution episode and provides structured feedback for replanning, avoiding the overhead of per-step inference. Experiments on LoHoRavens, CALVIN, Franka Kitchen, and real-world manipulation tasks show that DAHLIA achieves strong performance on over 30 long-horizon tasks and improves generalization to unseen scenarios, including cluttered scenes and occluded objects.

Accepted by IROS 2026 (8 pages)

Long-horizon manipulation under sparse rewards remains challenging for reinforcement learning due to delayed feedback and inefficient exploration. Existing skill-based approaches often assume a fixed parametric prior (e.g., a single Gaussian), limiting their ability to capture diverse and multi-modal skill structures required for complex tasks. We propose a Bayesian non-parametric skill prior that models temporally extended skills in a structured latent space using a Dirichlet Process Mixture, enabling adaptive skill discovery without predefining the number of components. Integrated into a hierarchical RL framework, the learned prior guides high-level skill selection while a pretrained decoder generates temporally abstracted actions, improving exploration efficiency in sparse-reward settings. Experiments on Franka Kitchen, LIBERO-Long, Meta-World, and a real robot demonstrate consistent gains in long-horizon manipulation, achieving over 0.8 success rate within 1.5M steps, whereas SAC fails to converge even after 5M steps ($<$0.1). Compared to a single-Gaussian prior baseline, our model yields an average improvement of 21.8%.

Accepted by IROS 2026 (8 pages)