NVIDIA releases GR00T 2 world action model and Isaac Lab 3.0 at CoRL
NVIDIA releases GR00T 2, an open world action model built on Cosmos, at CoRL, and launches Isaac Lab 3.0, Isaac Lab-Arena 0.4, and benchmarks such as DexBench and VABAR.
Highlights
UC Berkeley and NVIDIA propose agentic robotics, using AI agents to generate and improve robot programs in simulation
Ken Goldberg's team at UC Berkeley and NVIDIA robotics engineering lead Spencer Huang propose the "agentic robotics" method, letting AI agents generate, test, and correct robot control programs in simulation without additional data collection.
Paper analyzes agentic manipulation policy: AMP1 achieves 85.6% zero-shot success rate
The paper proposes an agentic manipulation policy (AMP), in which a general AI model directly controls the robotic arm and adjusts actions as the task progresses; the authors' reference implementation AMP1 runs without modification on single-arm, dual-arm, and mobile manipulation.
arXiv cs.ROSource ↗Learning-augmented robotic automation achieves cable insertion and welding on a motor production line
The research proposes a hybrid Learning-Augmented Robotic Automation system, integrating a learned task controller and a neural 3D safety monitor into a traditional industrial process, automatically completing flexible cable insertion and welding on a motor production line. Each task uses less than 20 minutes of real data; the system ran continuously for 5 hours and 10 minutes without physical fences, producing 108 motors, with a product-level quality inspection pass rate of 99.4%, a takt time close to manual work with smaller cycle time fluctuations, and weld quality comparable to skilled operators.
arXiv cs.ROSource ↗
Model releases/updates
NVIDIA releases GR00T 2 world action model and Isaac Lab 3.0 at CoRL
NVIDIA releases GR00T 2, an open world action model built on Cosmos, at CoRL, and launches Isaac Lab 3.0, Isaac Lab-Arena 0.4, and benchmarks such as DexBench and VABAR.
NVIDIA releases robot foundation model GR00T 2, simultaneously predicting future video and actions
NVIDIA releases the robot foundation model GR00T 2, which can simultaneously predict future video and actions. The model emphasizes learning what the world will become, not just what to do next, with the introduction given by Jim Fan at CoRL Austin.
- NVIDIA releases GR00T 2 world action model and Isaac Lab 3.0 at CoRL
Boston Dynamics puts wheels on Atlas, validating actuator limits with whole-body reinforcement learning
Boston Dynamics control engineer Andy Wu introduces that the team replaced Atlas's feet with wheels as a research testbed, driven by whole-body reinforcement learning policies, to test actuator and motor models under extreme conditions.
Paper research
UC Berkeley and NVIDIA propose agentic robotics, using AI agents to generate and improve robot programs in simulation
Ken Goldberg's team at UC Berkeley and NVIDIA robotics engineering lead Spencer Huang propose the "agentic robotics" method, letting AI agents generate, test, and correct robot control programs in simulation without additional data collection.
Paper analyzes agentic manipulation policy: AMP1 achieves 85.6% zero-shot success rate
The paper proposes an agentic manipulation policy (AMP), in which a general AI model directly controls the robotic arm and adjusts actions as the task progresses; the authors' reference implementation AMP1 runs without modification on single-arm, dual-arm, and mobile manipulation.
arXiv cs.ROSource ↗Learning-augmented robotic automation achieves cable insertion and welding on a motor production line
The research proposes a hybrid Learning-Augmented Robotic Automation system, integrating a learned task controller and a neural 3D safety monitor into a traditional industrial process, automatically completing flexible cable insertion and welding on a motor production line. Each task uses less than 20 minutes of real data; the system ran continuously for 5 hours and 10 minutes without physical fences, producing 108 motors, with a product-level quality inspection pass rate of 99.4%, a takt time close to manual work with smaller cycle time fluctuations, and weld quality comparable to skilled operators.
arXiv cs.ROSource ↗Dexmate breaks down the full process of coding agents executing robot tasks
Dexmate has Codex (GPT-6 Astra) and Claude Code (Opus 5.5) control the Vega wheeled dual-arm mobile manipulation robot through MCP tools; the fruit-into-basket and door-opening tasks all succeeded, while the higher-precision bowl-rack task had no successes under the original tool interface.
Brief updates
A2-MEAS: an embodied multi-agent framework with code as the extension interface
The authors propose the embodied multi-agent framework A2-MEAS, which uses code as the extension interface, combining corrective language, spatial evidence, execution history, and action suggestions into corrective code, with MetaBrain providing supervision and runtime guidance during training.
arXiv cs.ROSource ↗Apple publishes HALO humanoid robot research: using a Unitree G1 plus a Mac mini to achieve autonomous navigation in unfamiliar homes with pure simulation data
Apple's research team publishes the HALO humanoid robot research, using a Unitree G1 as the body, carrying a Mac mini M4 Pro running the Qwen model, completing tasks such as sitting down, crawling under tables, and pushing boxes in 6 unscanned real homes, with an average success rate of 91% across 360 tests.