Reading Modes
Papers & Evaluations with links to supporting source material.
556 selected reportsLast 30 days 491 recordsAll coverage 2,632 records
This research presents MPAIL2, the first method to learn visual manipulation in the real world from scratch using observational learning via Inverse Reinforcement Learning from Observation (IRLfO). It achieves 82% performance in pick-and-place tasks within 40 minutes, surpassing reinforcement learning and behavior cloning approaches with comparable interaction and demonstration budgets.
This research introduces a complete, GPU-resident pipeline for multi-robot 3, D inspection of underwater structures. The pipeline computes 6-DOF trajectories for a fleet of robots with limited cameras, minimizing a makespan/total-cost objective. It decomposes the NP-hard problem into five core stages, including visibility computation, set cover, and routing optimization, with GPU-accelerated implementations and experimental validation on 14 meshes.
This research proposes Collective Experience Learning (CEL), a paradigm where deployed robots pool execution experience to update a shared policy. The method, implemented via Failure-Guided Experience Retrieval Fine-Tuning (FERT), uses failure contexts to retrieve successful trajectory segments for supervised updates without requiring new expert demonstrations. Evaluated across five real-world manipulation tasks on robots excluded from training, the approach achieves success rates between 93.3% and 99.2%. Specifically, on multi-stage phone packaging, using a fleet of 96 robots improved success rates from 79.2% to 98.3%, validating CEL as a scalable method for improving robot policy reliability through deployment-generated experience.
This research introduces T(R,O) Grasp++, a scalable diffusion-based framework for cross-embodiment dexter, grasping in cluttered environments. The model achieves an average success rate of 88% across eight dexterous embodiments, with 0.3s inference latency and improved sample efficiency.
This research proposes Skill-RSI, a framework for creating reusable manipulation skills for embodied agents through recursive self-improvement. Skills are compiled in simulation and run as standalone code or under a cheap caller, achieving high success rates on new layouts. The study evaluates performance across 7 RoboTwin 2.0 tasks, demonstrating the effectiveness of inheritance over from-scratch rewriting and the cost-efficiency of frozen skills.
This research presents EndoVLA-OOV, a vision-language-action policy for robotic endoscopy that translates anatomical descriptions into camera motions. The policy uses a pretrained backbone adapted to continuous distal bending commands and combines images to guide search. An open-source dataset with 1,050 episodes and 262k observation-action pairs supports policy learning. Physical evaluations show 93.3% success in cross-embodiment transfer, with success rates improving from 40.0% to 90.0% with more training data.
Isaac Lab 3.0 is now generally available, offering a unified framework for robot learning with support for multiple physics, rendering, and visualization backends. Key features include Newton physics, OV libraries, and Isaac Sim integration, along with improved simulation tools and reinforcement learning compatibility.
After a decade of development in unmanned delivery, the final hundreds of meters to the doorstep remain a significant challenge. This article explores the difficulties in private spaces like residential areas, the technical and operational hurdles, and the innovative solutions from Nuwa Robotics, including their 'world-passing model' and deployment strategies. It also discusses the broader implications for the future of robot delivery and the potential for expansion into other service scenarios.
The article explores the evolving landscape of dexterous hands in robotics, noting a trend away from high degrees of freedom (DoF) toward more practical metrics like durability, cost, and operational efficiency. It highlights how industrial standards are influencing the market, with grippers and simpler hands becoming more prevalent. The text also discusses the limitations of current VLA models in controlling complex robotic hands and the role of data collection in training these systems.
At the 2026 World Robot Exhibition, Faowei Robotics showcased its force-controlled robotic arms and smart workstations, demonstrating real-world applications in manufacturing. The company emphasized industrial deployment, with solutions for tasks like shoe gluing, welding, and assembly. Faowei also announced a $1 billion funding round, signaling strong industry confidence in its embodied intelligence technology.
This paper reviews the field of physics-informed learning in robotics, exploring how integrating physical principles into AI models can reduce reliance on massive datasets. It categorizes approaches into three main types: physics-guided, physics-encoded, and physics-informed, each with distinct implementation strategies and trade-offs. The review also discusses challenges in high-dimensional systems and contact dynamics, emphasizing the need for more robust solutions in complex robotic applications.
A research team from Shanghai Jiao Tong University identifies a critical flaw in current robot control models: they rely on visual shortcuts rather than true language understanding. To address this, they introduce RoboFollow, a benchmark with high-scene-entropy scenarios that force models to depend on language for task resolution. The study shows that even top models fail under slight variations in scene layout or instruction phrasing, indicating a deeper issue in training paradigms.
Daxiao Robotics has released the embodied base model ACE-Brain-0.6, which surpassed OpenAI's GPT-6 Astra in six out of eight visual grounding evaluations. The model enhances spatial perception through approximately one million high-quality, verifiable samples. While it performs better than general-purpose models in several international benchmarks, further validation of its stability and generalization in real-world scenarios is needed.
This article observes the technical paths and scene deployment progress of five Chinese companies in the field of world models and embodied intelligence. It evaluates their core competencies and development prospects, emphasizing the transition from parameter scale to system-level performance and the importance of real-world data feedback in model iteration.
Menlo Research has released the Python SDK for its Asimov 1 humanoid robot, enabling developers to control movement, access sensors, and integrate media. While the SDK is open-source under Apache 2.0, the control layer Asimov Edge remains unverified. This marks a shift in the industry toward software-driven development, with challenges in real-world integration and safety.
This paper introduces the self-building laboratory (SBL), a system that designs, configures, and operates its own laboratory setup given a high-level scientific goal. SBL can start from an empty room or an existing self-driving laboratory and build missing capabilities including lab layout, digital twin, robot skills, and experimental workflows. The system is demonstrated across different robots and setups, scaling from a single workstation to multi-robot laboratories. Validation is provided through a real-world quantum-dot synthesis experiment, showing that SBL shifts human researchers from manually configuring device automation to specifying scientific objectives, making scientific automation more accessible, flexible, and scalable.
This research presents LAW, a closed-loop framework that combines vision-language models and world-action models to enable agentic physical intelligence. The system uses semantic subtask progress for planning, action evaluation, and execution monitoring. LAW achieves 94.21% average success on 50 RoboTwin tasks and 65% on real-robot tasks, outperforming existing methods like $pi_{0.5}$ and Fast-WAM.
SenseSwarm is a research paper proposing a decentralized swarm exploration method that uses a lightweight learned map predictor to enhance mapping efficiency. The approach integrates neural structure estimation with shared world-anticipation, allowing a swarm of robots to rapidly map unknown environments. The method is evaluated on a benchmark of real building floors and simulations, showing improved performance in map reconstruction compared to non-predictive baselines.
The 2026 Physical AI and World Model Conference in Nanjing marks a significant step in the development of embodied intelligence, with local companies like Yuanfang Intelligent and Jiajia Vision showcasing advancements in world models and physical AI. The event underscores the region's commitment to fostering an ecosystem for real-world AI applications.
The Tencent Robot X Team, Fujita Lab, and Tencent Hunyuan Team jointly released a study proposing a new embodied cognition foundational model called Hy-Embodied-RxBrain (shortened as RxBrain). This model aims to enable robots to not only describe steps in language when planning actions but also generate mental images of the world state after each step in their mind.
Article dates follow your selected timezone. Briefing editions use Hong Kong time (UTC+8).