Reading Modes
Open Artifacts with links to supporting source material.
782 selected reportsLast 30 days 706 recordsAll coverage 2,428 records
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.
This research presents RAX, an open-source framework for robot agents that enables continuous self-improvement through online exploration. The framework allows agents to acquire experience, synthesize executable programs, and perform complex tasks without human demonstrations. RAX demonstrates significant performance improvements in both simulation and real-world settings, with results showing a +37.4% average improvement over top baseline models.
China has established over 70 training fields for embodied AI, with additional projects in planning, driven by government policies and industry investment. These centers aim to generate high-quality data for robot training, but face challenges such as data quality, sustainability, and market dynamics. The article explores both successful initiatives and failed projects, highlighting the evolving landscape of data collection and its role in advancing robotics.
NVIDIA announced at GTC Berlin the expansion of its Isaac open robot development platform, including a humanoid robot reference platform, new tools, and datasets. The Agile ONE SN, set for early 2027, integrates simulation, training, and deployment tools, with hardware and software designed for research and development. The platform includes a Jetson Thor compute platform, WUJI Hand 2 hands, and integration with NVIDIA's Halos safety system.
RoboPrompt is a robot control framework proposed by a team from Shanghai Jiao Tong University, which allows users to intuitively guide robot actions through sparse inputs such as points, trajectories, or directional commands. The framework employs a two-stage design, translating human intent into action drafts and then refining their execution through a lower-level policy. Experiments show that RoboPrompt significantly improves the success rate of robot tasks and reduces the need for manual intervention.
Researchers from Zhejiang University propose EmbodiedMemory-Bench (EMem-Bench), a benchmark to evaluate long-term memory in embodied agents, and introduce a memory system (EMem) and model (EMem-8B) to enhance memory management. The work identifies key challenges in retaining and using memory across dynamic environments and demonstrates performance improvements across multiple models.
A research team from Shanghai Chuangzhi College, Yuance Future, Hong Kong University, and NVIDIA has developed PingpongDuo, a system that enables two Unitree G1 humanoid robots to perform table tennis doubles. The system combines perception, team coordination, and whole-body motion control, using reinforcement learning for planning and risk assessment. Real-world testing showed improved performance over traditional rule-based strategies, with the robots achieving 18 consecutive alternating returns in a human-robot doubles match.
UBTECH Thinker-WM achieved first place in three major benchmark evaluations: WorldOlympiad, EWMBench, and WorldArena1.0 Track-1, demonstrating its advantage in cross-robotic form and perspective generalization capabilities.
This research paper evaluates GPT-6-Astra's embodied intelligence capabilities across 12 simulated benchmarks and 8 real-world robot platforms. While it achieves high scores in structured tasks like LIBERO-Long++ and SIMPLE, real-world execution reveals significant limitations, including safety issues, mechanical load constraints, and failure to handle complex, unstructured tasks. The study highlights the gap between simulation and physical deployment, and proposes improvements through better execution systems and task-specific adaptations.
Jim Fan@DrJimFanSignalEditorial score6868
ETH Zurich researchers have developed a system where a WUJI robotic hand can crawl, recover from falls, and perform manipulation tasks without a separate mobility system. Using reinforcement learning, the hand learns to support its own weight and move using its fingers. The work was presented in a paper titled 'Fingers as Legs: Learning Self-Supported Locomotion and Manipulation with an Anthropomorphic Hand.'
Skild AI has developed a self-play training method that allows robots to acquire physical skills without human demonstrations. The approach, which transfers successfully to a physical humanoid robot, represents a step toward autonomous physical AI and could reduce the reliance on human-generated data for complex tasks.
Leju Robotics and Zhao Feng Machinery have deployed 15 humanoid robots in a complex industrial environment, achieving a 98% success rate in part handling and completing over 60,000 components in two months. The solution includes open-sourced real-world data and tools for skill development, aiming to lower the barriers for embodied AI research and deployment.
At the 2026 Sustainable Global Leaders Conference, Leju Robotics and Zhao Feng Machinery jointly launched an embodied intelligence solution for automotive parts loading. The 15 humanoid robots completed a full loading cycle in under 40 seconds with a 98% success rate, operating for over 1000 hours. The project also released a real-world data set for the OpenLET community to support further research and development in embodied intelligence.
At the 2026 Sustainable Global Leaders Conference, the collaboration between Luju Robotics and Zhao Feng Machinery revealed a transition from pilot testing to regular operation of humanoid robots in automotive manufacturing. The robots, operating in a dynamic production environment, completed a full workflow of preparation, material retrieval, transportation, and loading with high reliability and success rates. The project also open-sourced real production line data, addressing the scarcity of real-world industrial data and enabling continuous model improvement through real-world feedback.
RoboAtlas is a contextual Active SLAM framework that adaptively balances geometric exploration and semantic reasoning using a scalable 3D semantic mapping system. Evaluated in simulation and on a Unitree Go2 robot in large-scale real-world environments, it achieves a 100% task success rate and state-of-the-art performance on the GOAT-Bench 'Val Unseen' benchmark, with success rates of 90.6% using GPT-4o and 88.8% using Qwen2.5-VL-7B.
Spark-0 is a vision-tactile-language-action model that leverages tactile data to enhance robot dexterous manipulation. The model uses a 1,000-hour tactile dataset, HumanTouch, and combines it with visual data and simulation for pretraining. It achieves the highest mean success rate on SparkArena and improves success rates on both simulated and real-world tasks when tactile data is included.
This research presents X-Fly, a staged training framework for full-chunk aerial vision-language navigation. By executing complete predicted action chunks before replanning, the method reduces model inference time from 195 ms to 42 ms per waypoint and improves recovery from non-expert states. The approach combines Recovery DAgger labeling with Flow-GRPO refinement to jointly optimize a chunked flow controller and a Bernoulli STOP policy. On the OpenFly test-seen dataset, X-Fly achieves a 74.18% success rate compared to 16.45% for full-chunk behavior cloning and 32.04% for DreamFly.
This release introduces the lerobot/so101-tabletop-mujoco environment within the LeRobot EnvHub framework. It simulates an SO-101 arm interacting with four colored cubes and a green box on a table using MuJoCo physics. The setup includes front and wrist cameras, 6 joint positions for observation, and absolute joint targets for control at 30 fps. A corresponding Isaac Sim version is also available, allowing the same policies to run across both simulators. The environment supports seeded resets, noise injection, and headless execution via EGL, all under an Apache-2.0 license.
The LeRobot team released lerobot/so101-tabletop-isaac, an environment built natively in NVIDIA Isaac Sim 6.2 featuring an SO-101 arm, four wooden cubes and a green box on a table. The layout, cameras, observations and actions match the MuJoCo version, allowing the same policies to run against both simulators. The environment is Apache-2.0 licensed, includes bundled assets requiring no network access at runtime, and runs on an NVIDIA RTX GPU with PhysX physics and RTX rendering.
Article dates follow your selected timezone. Briefing editions use Hong Kong time (UTC+8).