AI BriefingNVIDIA Unveils GR00T 2 and Isaac Lab 3.0 at CoRL to Advance Precision Robot Manipulation
·2026.11.09 23:00
Key point
The open frontier world action model, trained on 100,000 hours of human data, enables skill transfer across robot bodies and ranks among the top three on major robotics benchmarks.
Details
- NVIDIA introduces GR00T 2, an open frontier world action model built on NVIDIA Cosmos, which achieves 95% average success across 50 tasks with just five demonstrations per task and 71% with one.
- Isaac Lab 3.0, the latest open source robot-learning framework, integrates the Newton physics engine to model deformable objects and complex articulated systems, achieving up to 2.5x the training throughput of Isaac Lab 2.3 with PhysX in vision-based dexterous manipulation benchmarks.
- Two NVIDIA papers, GRAIL and RoboTTT, are finalists for the Best Paper Awards at the Conference on Robot Learning (CoRL), focusing on digital generation pipelines for 3D assets and scaling visuomotor context three orders of magnitude beyond state-of-the-art policies.
- New open benchmarks DexBench and VABAR are released to measure robot dexterity and adaptability across industrial tasks, with DexBench featuring 15 tasks across retail, logistics, manufacturing, and hospitality scenes.
Simulation and Training Tools
- Isaac Lab 3.0 allows researchers to switch between NVIDIA PhysX and Newton physics solvers and apply a new reinforcement post-training library to foundation models like GR00T.
- The University of Washington and NVIDIA developed Success-Guided Sampling (SGS), a method that focuses practice on tasks a policy succeeds at about half the time, enabling robot training to scale beyond a million parallel environments.
- AlphaNIST, built in Isaac Lab, demonstrates a single policy learning all 20 tasks on the NIST Assembly Task Board with more than 98% training success and roughly 39x less GPU time compared to separate per-task policies.
Agentic Robotics and Partners
- UC Berkeley’s Graph-as-Policy (GaP) framework uses agents to connect individual skills into robot programs, testing them in simulation and refining them based on failures to keep programs interpretable.
- Partners including Ambi Robotics, Bosch Research, Siemens, and Skild AI are integrating these technologies, with Skild AI announcing STEEL, a new rigid-body physics solver built on Newton to accelerate robot training.
- NVIDIA EgoHand, an open model for reconstructing 3D hand motion, has received over 40,000 Hugging Face downloads in the two weeks since launch.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
NVIDIA Releases GR00T 2 Foundation World-Action Models for Cross-Embodiment RoboticsRunway Releases Praxis-1 Open-Weight Robotics Model