Skip to content

Technical Directions

VLA Latest coverage

Robotics coverage relating to vla.

286 selected reportsLast 30 days 265 recordsAll coverage 622 records

updated

Signal archive — Page 10

2026-11-02Mon181–200 records
  1. arXiv Robotics — research abstracts75

    Language-Conditioned Behavioral Cloning for Autonomous Endoscope Guidance

    This research paper proposes a language-conditioned behavioral cloning approach to automate endoscope guidance during minimally invasive surgery. The authors define six clinically relevant guidance behaviors and collect 235 subtask demonstrations. Preliminary experiments on phantoms show the policy achieves 100% success on focus, overview, and stop behaviors, 80% on search-tip and follow-tip behaviors, and 50% on horizon leveling. The results indicate feasibility for learning heterogeneous surgeon-requested camera behaviors with a single visuomotor policy, though failure modes suggest future work is needed to incorporate temporal context for more demanding tasks.

  2. arXiv Robotics — research abstracts75

    Movens-Zero: Scaling Humanoid Motion Foundation Models with Large-Scale Human Videos

    This research presents Movens-Zero, a scalable architecture for humanoid motion foundation models trained on a large-scale dataset of physically executable motions derived from internet videos. The system supports joint learning across text, speech, music, and video modalities to enable real-time closed-loop execution. Experiments demonstrate consistent gains in generation and execution quality when scaling both the volume of verified data and model capacity.

  3. arXiv Robotics — research abstracts75

    META-Robot: Self-Cognition and Capability Scaling in Agentic Robots

    This research presents META-Robot, a framework that integrates self-cognition with capability scaling for LLM-based robotic agents. By using action-level predicates to make execution outcomes explicit, the system enables agents to evaluate their own limitations, extend evaluation criteria, and guide structured skill expansion. Proprioceptive feedback grounds this process in the robot's physical state. Experiments in simulation and on real robots show improved task performance, better generalization to new tasks, continued skill expansion, and effective use of non-visual proprioceptive signals.

  4. arXiv Robotics — research abstracts75

    Visual-Inertial RGB-D SLAM with Encoders for a Differential Wheeled Robot

    This study introduces a novel tightly coupled visual-inertial encoder RGB-Depth (RGB-D) SLAM system designed for differential wheeled robots. By incorporating a generative measurement model for encoders and deriving expressions for maximum a posteriori (MAP) estimates, the approach aims to overcome limitations of purely vision-based simultaneous localization and mapping (SLAM), such as temporary feature point scarcity and fast motion. Tests on an in-house dataset demonstrate accurate and robust real-time state estimation with a root mean squared error (RMSE) of approximately 2-7 cm. The source code and dataset containing encoder information are published for verification.

  5. arXiv Robotics — research abstracts75

    $\nabla$Mimic: Ultra-Efficient Humanoid Control with Differentiable Simulation

    This paper introduces $\nabla$Mimic and NablaEngine, a differentiable GPU simulation engine for the Unitree G1. The authors address the limitations of zeroth-order methods like PPO and unstable first-order approaches by using exact simulator gradients bounded by gradient clipping and multistep regression. NablaEngine computes MuJoCo's full contact dynamics and adjoints as fused megakernels, achieving analytic gradients that match float64 finite differences and are 11.5 times faster than the available MuJoCo Warp adjoint. On whole-body tracking of all 40 LAFAN1 clips with one policy on one GPU, $\nabla$Mimic reaches a score of 0.82 in 20 minutes (matching the state-of-the-art world-model method) and 0.874 in 1.8 hours, whereas the standard first-order baseline SHAC stops due to exploding gradients within minutes. The code is released as open-source infrastructure.

  6. arXiv Robotics — research abstracts75

    RTPolicy: Real-Time (32.9 Hz) Multi-View Visuomotor Policy for Robotic Manipulation on Edge Devices

    This research introduces RTPolicy, a multi-view visuomotor policy optimized for low-latency robotic manipulation on edge devices. By combining view-aware feature fusion with visual compression and a flow-matching action generator, the system achieves real-time inference at 32.9 Hz (30.35 ms latency) on an RDK S600 device while maintaining high performance. On the RoboTwin~2.0 benchmark across 50 tasks, RTPolicy achieves a mean success rate of 52.26%, outperforming the strongest baseline ($\pi_0$ at 46.4%). Under asynchronous streaming inference on a real bimanual platform, it sustains a 72.0% mean success rate. The method reduces computational costs for visual representation and action generation while preserving necessary visual information.

  7. arXiv Robotics — research abstracts75

    VLA Policies Do Not Know When Not to Act: Unintended Task Execution and Start Suppression

    This research investigates the behavior of vision-language-action (VLA) policies when given missing, meaningless, or incorrect instructions. Using six model checkpoints in the LIBERO-Goal simulation environment, the authors found that robots consistently initiated movement even with empty inputs and frequently completed unspecified tasks. While a gating mechanism successfully suppressed motion on non-task inputs for five of the six models, fine-tuning approaches failed to reliably keep robots still. The study concludes that current VLA architectures lack the ability to distinguish between valid commands and noise, posing potential risks in real-world deployment where input reliability cannot be guaranteed.

  9. arXiv Robotics — research abstracts75

    DiffWBC: Unlocking Steerable General Humanoid Motion Tracking via One-Step Diffusion Whole-Body Control

    This paper introduces DiffWBC, a generative whole-body control framework for humanoid robots that unifies general motion tracking, test-time steering, and simulation-free failure discovery. The method utilizes a one-step diffusion model to achieve low local tracking error (20.1 mm on BONES) and zero-shot adaptation to various tasks including hurdle navigation and actuator faults without retraining or physics simulations.

  10. arXiv Robotics — research abstracts75

    HandMimicX: Learning Human-Like Dexterous Manipulation via an Object-Centric Policy from Egocentric Videos

    This paper introduces HandMimicX, a method for learning dexterous manipulation from human videos by formulating reinforcement learning objectives around object motion rather than hand trajectories. The approach uses egocentric RGB-D videos to acquire skills that generalize across spatial configurations while maintaining physically plausible, anatomically natural hand-object interactions. Experiments in simulation and on real robots demonstrate efficient learning from human demonstrations and successful transfer to new environments.

  11. arXiv Robotics — research abstracts75

    HIPC: Humanoid Iterative Planning and Control via Goal-Conditioned Motion Closure

    This research introduces HIPC, a real-time framework for humanoid robots that unifies generative motion planning with closed-loop control. It addresses issues like spatial drift from open-loop paradigms and discontinuities from direct state feedback by using goal-conditioned planning, deviation-decayed recovery, history re-anchoring, and latency compensation. Experiments demonstrate that HIPC achieves state-of-the-art goal-reaching accuracy while maintaining robustness against physical disturbances and the ability to recover from falls.

  12. arXiv Robotics — research abstracts75

    CONSORT: Uncertainty-Constrained Multi-Robot Planning for Cooperative Localization in GPS-Denied Environments

    This research introduces CONSORT, a method for centralized multi-agent path planning that achieves cooperative localization in GPS-denied environments. By optimizing mean-state trajectories and propagating covariance only for terminal constraints, the approach handles uncertainty dynamics more tractably than prior belief-space formulations. Evaluated across 50 randomized scenarios, CONSORT solves 80% of cases where a strong baseline solves 54%, while maintaining comparable path lengths. Marine simulations using independently developed factor-graph estimates confirm that executed trajectories remain collision-free and terminal uncertainty stays within 6% of the planner's prediction.

  13. arXiv Robotics — research abstracts75

    TinyAstral: Transferring Self-Evolving Harnesses to Tiny Generalist Manipulation Agents

    This paper presents TinyAstral, a two-stage approach that first improves a large teacher model's harness through recursive self-improvement (RSI) and then distills the enhanced capabilities into a compact student agent. On RoboTwin, this method raises the teacher's success rate from 33.4% to 63.6%, with the final 4B student reaching 62.8%. The harness gains transfer to other frozen planners without retraining, and the student outperforms open 4B agents and a 27B harness-RSI model while using about one-fifth of the teacher's accelerator time.

  14. arXiv Robotics — research abstracts75

    Veloxity: Towards Memory-Safe UAV Embedded Firmware using the Rust Programming Language

    This paper introduces Veloxity, an open-source embedded autopilot for unmanned aerial vehicles (UAVs) written in the memory-safe Rust programming language. The authors address the fragmented adoption of memory-safe languages in aerospace by providing a minimal, modular foundation that integrates with established autonomy ecosystems. They compare Veloxity to equivalent C++ autopilots using simulation, hardware testbenches, and outdoor waypoint-following flight tests, demonstrating functional parity and improved performance. These results suggest that modern memory-safe languages are a viable option for reliable and performant firmware in autonomous aerospace systems.

  15. Chinese robotics — Research news discovery75

    Feijie Kesi Releases Fysiverse-UniWAM Unified World Action Model with Fudan University's CITLab

    Feijie Kesi, in collaboration with Fudan University's Cognitive and Intelligent Technology Laboratory (CITLab), has released the Fysiverse-UniWAM unified world action model. Unlike traditional approaches that concatenate navigation and operation models, UniWAM learns navigation, stopping positions, and manipulation actions within a single framework. Key innovations include the Manipulation Anchor Pose (MAP) for precise docking and a Mixed-Stream architecture sharing video/action backbones while maintaining separate input/output heads. The model achieves state-of-the-art results on MAP-Bench benchmarks, outperforming models like GR00T and π0.5 in navigation accuracy and manipulation success rates. It demonstrates effective sim-to-real transfer without requiring expensive robot pretraining, with open-source code and data available.

  16. Robotics — Chinese web discovery75

    Stanford alumni found RoboScience, raising hundreds of millions in capital from industry investors

    RoboScience, founded by Stanford AI Lab alumni Tianye and Shaolin, announced the completion of its A2 round financing led by Amber Capital, Hongruida Investment, and Shanghai Semiconductor Industrial Investment. The company has developed a full-stack embodied AI technology stack including the Visics general-purpose embodied large model with a VLOA architecture, the RoboMirage high-precision simulation engine, and the REX G1 omnidirectional humanoid robot platform. Key technical milestones include the demonstration of cross-body generalization where different end-effectors could be swapped and utilized within 30 seconds without retraining, and the public release of world model reasoning processes. The company aims to reduce data acquisition costs and enable scalable deployment across various hardware platforms and scenarios.

2026-11-01Sun
  1. IT Home — Robotics75

    AI Performs Brain Surgery on Mice for Epilepsy Treatment

    Exin Therapeutics, a biotech startup, claims to have successfully used an AI agent to perform brain surgery on a mouse to treat epilepsy. The AI selected gene therapy vectors, dosages, and brain regions using an electronic spreadsheet, three research papers, and a camera-equipped robotic arm. The AI injected 250 nanoliters of medication while monitoring for bleeding. Human oversight and safety protocols, including an emergency stop button, were in place. The AI was limited to the injection phase, with craniotomy and suturing performed by human surgeons. The team initially attempted to use Anthropic's Opus 5.5 model, which refused the task due to concerns about incomplete execution.

  2. Machine Heart — Robotics on WeChat75

    Stelledge's Auriga Framework Leads RoboDojo and RoboCasa365 Benchmarks Under Zero-Shot Evaluation

    Stelledge Intelligent introduced the general-purpose embodied Harness framework Auriga, which achieved leading performance on the RoboDojo and RoboCasa365 benchmarks under a strict zero-shot evaluation protocol. Without task-specific pre-optimization or expert demonstrations, Auriga demonstrated strong cross-task migration capabilities by leveraging a unified physical execution system that supports perception, action planning, and online recovery across diverse robotic environments.

  3. Chinese robotics — Hardware and sensing75

    Humanoid Robots Enter Factories, but Success Depends on 'Brain' Integration

    The article discusses the deployment of humanoid robots like Apollo in manufacturing environments, emphasizing the integration of physical AI and multimodal reasoning for task execution. It highlights collaborations between Google DeepMind, Apptronik, and NVIDIA, as well as challenges such as energy efficiency and data collection for training physical AI systems. The article also notes the growing interest in humanoid robots across global manufacturing sectors.

  4. Chinese robotics — Research news discovery75

    Fysiverse-UniWAM: A Unified World Action Model for Navigation and Manipulation Without Large-Scale Robot Pre-Training

    A research team from Fysics AI and Fudan University's CITLab introduces Fysiverse-UniWAM, a unified world action model that integrates navigation, mobile manipulation, and fixed-position dexterous tasks. The model achieves strong performance on 24 real-world tasks, with notable improvements in positioning and orientation accuracy compared to existing baselines. The approach uses MAP (Manipulation Anchor Pose) and Mixed-Stream architecture to enable joint learning of navigation and manipulation, demonstrating the potential of alternative methods to large-scale robot pre-training.

Timezone · UTC

Article dates follow your selected timezone. Briefing editions use Hong Kong time (UTC+8).