Reading Modes
Papers & Evaluations with links to supporting source material.
556 selected reportsLast 30 days 491 recordsAll coverage 2,632 records
This research explores how agents fail to maintain goal alignment during distribution shift, identifying three distinct failure modes based on where the goal is lost within the agent's internal processing. The study shows that recovery methods vary depending on the depth of goal loss and demonstrates how steering and downstream wiring can improve goal reach in different scenarios.
This paper introduces DiffGIPC, a GPU-accelerated differentiable framework for inverse problems involving contact-rich interactions between deformable and nearly rigid bodies.
Robota-Embed is the first vision-language-action (VLA) embedding model that jointly encodes task, execution stage, and future continuation for robots. It achieves 47.33% clip-level and 81.77% episode-level R@1, outperforming action-free Gemini Embedding 2. The model supports training-free and training-based policy conditioning, improving success rates in simulated and real-world tasks, and enables task-local context evolution with reduced policy token usage.
This research introduces $S^3$-Nav, a framework combining online mapping, persistent object-relation memory, and goal-conditioned navigation for robots operating under partial observability. Using RGB-D inputs and SLAM, the system maintains a structured state of object records and confirmed relations without needing a pre-existing map. A goal-conditioned evidence interface selects relevant information to guide high-level decisions on exploration, approach, and stopping. The framework validates provisional policy decisions against persistent memory and geometric evidence before execution. Evaluations across multiple environments demonstrate improved closed-loop navigation reliability and reduced language-model token consumption when using capacity-controlled evidence exposure.
The Best Paper Award at IROS 2026 was given to research on long-term memory for robots, which introduced the LT-Mem framework. This framework addresses the issue of robots being unable to track object trajectories when entering the same room multiple times. The framework updates memory using a fluctuating score and performs well on the LT-VQA dataset, outperforming several baseline models. The research was conducted by the APRL lab at DGIST in South Korea, which was established less than a year ago. The two main authors enrolled in July 2025.
Researchers at Delft University of Technology developed an AI system that allows a racing drone to navigate at extreme speeds using only a single camera and an IMU, outperforming former world champion human pilots. The system uses reinforcement learning and direct motor control to achieve high-speed navigation without GPS, with potential applications in various autonomous robotic systems.
At IROS 2026, Clearpath Robotics showcased six research papers that leverage their Husky and Jackal platforms to address real-world robotics challenges. These include heterogeneous robot collaboration using generative AI, disease detection in orchards, robust LiDAR-inertial-visual odometry in coal mines, safety-constrained reinforcement learning for navigation, 4D-aware urban navigation from internet videos, and a novel calibration framework for event cameras and odometry.
This article introduces the Recursia framework, which achieves embodied recursive self-improvement (RSI) by organizing the system as an internal and external dual-loop structure, enabling generalization in robot zero-shot long-sequence tasks. Recursia achieved success rates of 28.4% and 53.4% in LIBERO-Pro Long tasks, respectively 5.8 and 15.1 percentage points higher than NVIDIA ASPIRE. The study emphasizes the importance of integrating execution feedback into high-level decision-making and explores the challenges of maintaining task success across different execution contexts.
UNCAP is a new framework for cooperative autonomous driving that enables vehicles to communicate using natural language to share observations and reduce uncertainty. The approach uses vision-language models and uncertainty calibration to determine which messages to trust and how to integrate them into driving decisions. Evaluated in CARLA using the OPV2V benchmark, UNCAP shows significant improvements in communication efficiency, safety, and decision-making compared to traditional methods and prior language-based approaches.
CylindTrack is a depth-aware cylindrical tracking-by-detection framework for panoramic multi-object tracking in mobile robots. It addresses coordinate discontinuities and depth estimation fluctuations by jointly modeling panoramic topology and temporal depth consistency. The framework improves identity consistency and trajectory continuity, achieving HOTA scores of 33.67 and 31.12, and IDF1 scores of 40.45 and 34.33 on QuadTrack and JRDB datasets, respectively. It runs at 28.56 and 21.34 frames per second on an NVIDIA GeForce RTX 3090 GPU.
This research proposes a method to distill world-model representations into compact robot policies by aligning student policies with a frozen world model during training. The approach achieves performance improvements on tasks like LIBERO and RoboCasa-GR1, with a 0.8B student reaching 97.9% on LIBERO and improving from 48.2% to 50.5% on RoboCasa-GR1. The method is also effective on real hardware, demonstrating broad representational prior rather than fragile network alignment.
This research proposes Diffusion ReRoll, a diffusion-based framework for robotic sequential prediction that enables revisable denoising over horizons. By selectively re-noising regions that reach a predefined noise threshold, the method allows for iterative cross-horizon revision, improving performance in tasks like planning, policy learning, and navigation. The results show significant gains over existing methods in both simulated and real-world settings.
This paper presents SPARK, a framework that maintains manipulation plans by anchoring symbolic behavior trees to current 3D observations, enabling local recovery from failures when objects move. Evaluated on LIBERO-PRO without ground-truth poses or task-specific practice, the system achieves 75.1% mean success across position and task settings and 69.7% on held-out language and object settings. It also demonstrates 68% success across 11 physical task-embodiment settings on three different robots using an open-weight 27B planner with 4-bit weights.
This research presents UniT, a framework that enables human-to-humanoid policy learning and world modeling by creating a unified physical language. The method uses a tri-branch cross-reconstruction mechanism and a shared latent space to achieve state-of-the-art data efficiency and zero-shot task transfer, validated on both simulation and real-world deployments.
This paper presents Approximate Imitation Learning, a novel framework that reduces the computational cost of training event-based control policies for quadrotors in cluttered environments. By separating representation learning from policy search, the method achieves a 28x speedup in training time and enables real-world validation with high-speed flight.
LLM2R is a new method for 3D part assembly that uses a pretrained decoder-only large language model (LLM) to predict the rigid SE(3) transformations of parts. It encodes parts into geometric tokens using a Point-BERT encoder and a Q-Former, then processes the sequence with a causal LLM adapted with LoRA. A learnable summary token block aggregates the entire assembly, and a part-aware pose head predicts rotation and translation. LLM2R achieves the highest part accuracy and success rate on PartNet categories, the best rotation and translation errors on PartNeXt, and the lowest translation error and fastest inference on Breaking Bad.
Industrial Next, a San Francisco-based startup, is revolutionizing factory automation by enabling robots to switch tasks in minutes rather than weeks. Their AI-driven robotic stations use proprietary foundation models and high-fidelity simulations to reconfigure production lines without traditional programming. The company partners with BMW Group Technology Office and has already deployed its technology in smartphone manufacturing for complex, delicate tasks.
Google DeepMind introduces Gemini 4 Argon, a general-purpose physical reasoning model that significantly advances robotics through capabilities in robot control, real-to-sim-to-real workflows, physical understanding, and research self-improvement. Argon outperforms other frontier models on the Embodied Reasoning (ERQA) benchmark, achieving 87% in 3D physical reasoning tasks. It enables complex robotic tasks like mobile manipulation, zero-shot control, and video world model steering, while also supporting simulation reconstruction from real-world data and human videos. The tool emphasizes safety, iterative improvement, and cross-embodiment applications, marking a milestone in the path toward physical AGI.
On October 12, 2026, the 'Industrial Embodied · Smart Manufacturing Without Borders' Industrial Embodied Intelligence Industry Innovation Summit Forum was held at the National Exhibition and Convention Center. The forum unveiled the Physical AI Industrial Embodied Operating System and the world's first Industrial Embodied Intelligence 6S Store, aiming to address technical bottlenecks and commercialization challenges in industrial scenarios for embodied intelligence.
OSARO has launched SightWorks, an AI-powered perception software designed to enhance robotic arms' ability to perform complex pick-and-place tasks in high-variability warehouse environments. The platform uses deep learning to recognize objects, identify grasping opportunities, and adapt to different object shapes and end-effector configurations, reducing the need for extensive SKU onboarding. SightWorks also incorporates automated learning and cloud connectivity for continuous model retraining and performance optimization.
Article dates follow your selected timezone. Briefing editions use Hong Kong time (UTC+8).