RoboSignal Weekly · 2026-W43
ROBOSIGNAL
Science Robotics Latest Cover: 100 km/h. AI-Controlled Drone Beats Human World Champions
本周 20 developments，最受关注的是：Science Robotics Latest Cover: 100 km/h. AI-Controlled Drone Beats Human World Champions；Tencent Robot X Team Creates 'Thinking and Imagining' Robot Brain - RxBrain；One Small Leap for a Muscle, One Giant Step for Robotics。
Robot Models
NVIDIA Releases Cosmos3-H-Embodied-Nano for Surgical Robotics
NVIDIA has released Cosmos3-H-Embodied-Nano, a multimodal model trained on surgical robotics data for forward dynamics, inverse dynamics, and policy prediction. It is based on the Cosmos3-Nano model and is intended for use in surgical world modeling and imitation learning.
Research
Section overviewNew frameworks for real-time adaptation and autonomous exploration in robotics research.
Science Robotics Latest Cover: 100 km/h. AI-Controlled Drone Beats Human World Champions
The MonoRace system, developed by the MAVLab team from Delft University of Technology in the Netherlands, defeated three human FPV racing world champions in the Abu Dhabi Autonomous Drone Racing Championship, achieving the fastest single-lap time of 16.56 seconds. The system, which uses only a monocular camera and IMU, achieved state estimation and control during high-speed flight through reinforcement learning and simulation training, demonstrating robustness under extreme conditions.
Tencent Robot X Team Creates 'Thinking and Imagining' Robot Brain - RxBrain
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.
ETH Zurich — Robotics research newsFirst-party10-22One Small Leap for a Muscle, One Giant Step for Robotics
ETH Zurich researchers have created biohybrid robots with living muscle tissue and soft artificial scaffolds, enabling visible and powerful movements such as jumping, swimming, walking, and grasping. The integration of muscle and scaffold design allows for a range of motion up to eleven times greater than earlier systems, though long-term deployment remains a challenge.
Gaode Releases First Full-Body Base Model λ0 Based on Open-World Behavior
Gaode, under the Alibaba Group, officially released its first full-body base model λ0 based on open-world behavior. This model enables robots to learn full-body coordination, including movement, posture, and hand operations, solely through human action data, thus overcoming the limitations of 'local action learning.' In the SIMPLE simulation benchmark and multiple real-world evaluations, λ0 achieved state-of-the-art performance, with data collection efficiency approximately 3.4 times that of mainstream methods and a 22.5% increase in real-world task success rate compared to the strongest baseline.
Zero-Shot Real-World Deployment. RAI Institute Creates 5-Motor Bicycle Robot: 8 m/s, Half-Second Aerial Flips, 1-Meter Jump Platform
The RAI Institute's UMV robot, a bicycle-like machine with 5 motors, achieves 8 m/s speed, 1-meter jumps, and half-second aerial flips without specialized balancing hardware. All movement strategies are trained in simulation and directly deployed to the physical robot, demonstrating zero-shot transfer. The robot uses reinforcement learning to manage balance and dynamic maneuvers, with no springs, control moment gyroscopes, or dedicated stability mechanisms.
Surgical robotics — News discovery10-22Surgical AI Learns to Adapt When Smoke and Darkness Blur the Robot’s View
A research team at Chengdu University of Information Technology has developed LiteST-TTA, a framework for real-time adaptation of surgical segmentation models during minimally invasive procedures. The model uses uncertainty estimation and temporal consistency to adjust predictions in response to visual degradation like smoke and low illumination. Evaluated on the EndoVis 2017 dataset, LiteST-TTA achieved a Dice similarity coefficient of 0.8879, outperforming the fixed-source baseline in severe low-light conditions. However, the study highlights that adaptation is not universally beneficial and depends on the type and severity of visual degradation.
Isaac Lab — ReleasesFirst-party10-20Isaac Lab 3.0 Released: Multi-Backend Architecture for Robot Learning
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.
RAX: Open-Source Framework for Autonomous Robot Exploration and Skill Acquisition
Researchers from Harvard, NVIDIA, UC Berkeley, MIT, and Stanford introduce RAX, an open-source framework that allows robot agents to autonomously explore and acquire skills through online experience. RAX enables robots to write and execute code in real or simulated environments, improving generalization and task performance across various robotic platforms. The framework shows significant improvements in benchmark tests, but real-world deployment and skill reuse still require further validation.
Universal Robots — Official announcementsFirst-party10-21Teradyne Robotics and Sereact Partner to Scale Physical AI Solutions for Retail and Logistics
Teradyne Robotics and Sereact have announced a partnership to deploy AI-powered physical solutions in retail and logistics, using Universal Robots' arms with Sereact's Cortex platform. The collaboration includes live deployments at Zalando's fulfillment centers, demonstrating the ability to handle unstructured tasks like returns processing. The solution runs on NVIDIA technologies, including Isaac Sim and Jetson Thor, and is being scaled globally.
arXiv Robotics — research abstracts10-20EndoVLA-OOV: Language-Conditioned Out-of-View Target Acquisition in Robotic Endoscopy
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.
Tech Xplore — Robotics10-17Team develops embodied robotic arm 'EmArm' with human-like touch sensing
A research team at HKUST has developed EmArm, an embodied robotic arm capable of whole-arm tactile sensing and adaptive manipulation. By integrating a soft optical tactile skin with a rigid skeleton and real-time decision-making algorithms, EmArm achieves high tactile fidelity and can distinguish between different touch interactions, such as tickling and patting. The system also enables emotional feedback expressions and safe autonomous operation in complex environments.
arXiv Robotics — research abstracts10-20Refuse the Harm, Complete the Task via Graph-of-Rules Planning for LLM-Enabled Robots
This research presents RuleGraph, a multi-agent Graph-of-Rules planning framework that allows LLM-enabled robots to reject harmful parts of instructions while completing legitimate tasks. The method distributes safety enforcement across multiple reasoning stages, achieving 0% attack success rate and high task success rates on benchmark environments.
ViGAR: First World-Action Model for Full-Scale Humanoid Robots
ViGAR, developed by Zhiyuan and Peking University, is the first complete world-action model (WAM) for full-scale humanoid bipedal robots. It has been validated on the real-world AgiBot A2 platform and achieved strong performance on the RoboTwin 2.0 benchmark. The model introduces a hierarchical architecture with subgoal planning and execution, enabling long-horizon task completion through explicit reasoning and context learning.
NVIDIA Expands Isaac Tools for Humanoid Robot Development
NVIDIA has expanded its Isaac open robotics development platform with a humanoid reference platform, new datasets, and runtime frameworks to support robot learning workflows from simulation through deployment. The Agile ONE SN platform integrates simulation, training, and deployment tools, offering a full-stack solution for humanoid robotics research and development.
LoopWAM: Closed-Loop Generalist Policy Improvement from Deployment Rollouts
LoopWAM, a world action model developed by Tsinghua University, Yuanli Lingji, and Jihuan Shidai, achieves top performance on RoboChallenge Table30 V2 and other benchmarks. It leverages real-world deployment data to iteratively improve robot policies through quality-guided learning, incorporating success, suboptimal, and failure trajectories to enhance world modeling and action generation. The model uses a video-action dual-branch architecture and quality steering to guide policy improvements, achieving significant performance gains across multiple tasks.
NeuroVLA: The Global First Class-Brain Architecture VLA Embodied Large Model
Zhi Fang Ping has developed NeuroVLA, the first class-brain architecture VLA embodied large model, featuring a 'cerebral cortex-cerebellum-spinal cord' three-layer system. It demonstrates safety reflexes in 20 milliseconds, 80.2% reduced jerk, and 0.4W power consumption. The company has achieved commercial deployment across 7+ industries, with 2000 annual production capacity and a 1000-unit order from Huiké. The model's progress is supported by data and industry recognition, though real-world deployment and long-term viability remain to be validated.
arXiv Robotics — research abstracts10-21From Instrument-Mounted Demonstrations to In-Vivo Execution: Learning Bimanual Laparoscopic Appendectomy Without Robot-Collected Demonstrations
This research presents an end-to-end pipeline for training a surgical robot policy using motion data captured from standard laparoscopic instruments. The system uses inertial, time-of-flight, and Hall sensors to track instrument state without external cameras. A diffusion policy is trained on both ex-vivo and in-vivo demonstrations, achieving appendectomy in three of four live rabbits. The robot serves as a timing reference and executor, with no demonstration collection.
AGIBOT Open-Sources AGIB T WORLD Release 4: 'Physical Reasoning' Dataset Enabling Robots to Reason Before Acting
AGIBOT has released the fourth version of its AGIBOT WORLD dataset, focusing on 'Physical Reasoning' to enable robots to make decisions before acting. The dataset covers over 10 categories of logical reasoning scenarios, supporting both high-level decision-making models and low-level action policies. It includes annotations linking task goals to execution processes, aiding in the training of embodied agents for real-world tasks.
Industry
Bear Robotics Launches KINO-2 with NVIDIA Collaboration for Robot Learning
Bear Robotics has launched KINO-2, a robot-learning platform that transforms task demonstrations into deployable manipulation skills for its KR1 humanoid robot. The platform integrates with NVIDIA Isaac OS and leverages foundation models to improve on-device inference speed, supporting both in-house and external models. KINO-2 is being introduced at NVIDIA GTC Berlin, demonstrating its capabilities in warehouse management tasks.
- 2026-W42Ultra-Light, High Torque, Super Smooth Backdrive: The Most Suitable AKE Collinear Drive Motor for Lightweight Exoskeletons is Now Available Today
- 2026-W41Could Sports Be the Cheapest Scenario for Robotics Industry to Implement?
- 2026-W40RoboSignal Weekly briefing · 2026-W40
- 2026-W39RoboSignal Weekly briefing · 2026-W39