NVIDIA on Monday introduced GR00T 2, a world action model it describes as open, and Isaac Lab 3.0, an update to its open-source robot-learning framework, at the Conference on Robot Learning in Austin, Texas.
NVIDIA said the advances are meant to help researchers train robots on complex tasks and move from lab demonstrations to industrial deployment. GR00T 2 is slated to be available by the end of the year, according to an NVIDIA glossary page2; the announcement itself gave no pricing.
All performance and benchmark figures are NVIDIA's own and have not been independently tested.
Simulation has long been used for locomotion, but attention at CoRL has shifted to dexterous manipulation alongside a new paradigm the company calls agentic robotics, which integrates frontier agents to produce code for physical systems. The new tools build on NVIDIA Cosmos, which Automatica reported in November5 added four open-weight checkpoints for interactive simulation.
GR00T 2: Built on Cosmos, Trained on Human Video
GR00T 2 is built on Cosmos and trained partly on 100,000 hours of open-source human egocentric data, according to NVIDIA's CoRL announcement1. The model is designed to let robots learn dexterous manipulation and transfer skills between different robot bodies without any robot demonstrations.
NVIDIA said GR00T 2 averaged 71% success across 50 tasks with a single demonstration per task, rising to 95% with five. It completed 40 of 50 evaluated tasks at least once with no task-specific demonstrations. On benchmarks, NVIDIA reported that GR00T 2 ranks among the top three models in all vision-language-action and world-action-model categories for RoboCasa365, RoboLab, RoboDojo and MolmoSpaces, with multiple leaderboard wins. The glossary page separately states2 it ranks No. 1 on MolmoSpaces and RoboArena.
NVIDIA said Foxconn is applying GR00T 2 to server-manufacturing automation, Siemens is evaluating it for industrial use cases, and Nexuni is testing it for robotic laundry handling. MicroAGI is using a version post-trained with reinforcement learning to perform Ethernet cable swapping on a YAM robotic arm, with the workflow to be demonstrated at CoRL.
Simulation Stack Adds Newton Physics
Isaac Lab 3.0, the latest release of the open-source framework, tightens integration with the open-source Newton physics engine, which can now model deformable objects such as cables and cloth. NVIDIA said the combined stack delivered up to 2.5 times the training throughput of Isaac Lab 2.3 with its PhysX solver on a vision-based dexterous manipulation benchmark.
The update adds a reinforcement-learning post-training library for foundation models such as GR00T and allows researchers to switch between PhysX and Newton physics solvers. A new developer preview of Isaac Lab-Arena 0.4, the latest version of the evaluation framework, lets developers run policies across thousands of variations in parallel to find failure conditions.
Partners including Skild AI, Flexion, FieldAI and Dyna Robotics described how they use the tools. Skild AI announced a new rigid-body physics solver called STEEL built on Newton. Flexion showcased how Isaac Lab and Newton help train its Reflect model, and NVIDIA said Dyna Robotics will demonstrate parallel simulation training on its semi-humanoid Taku robot.
NVIDIA also released an open Video to Data workflow that turns human video into simulated robot demonstrations and the EgoHand model for 3D hand-motion reconstruction, which has seen more than 40,000 downloads on Hugging Face in two weeks.
Two Papers Are CoRL Best Paper Finalists
Two NVIDIA papers are among six finalists for the Best Paper Award at CoRL. Neither has been independently replicated.
GRAIL presents a digital generation pipeline3 for 3D assets and simulator-ready scenes to train humanoid loco-manipulation. It uses a video world model to generate interactions without rebuilding physical environments and deploys policies trained on the synthetic data onto a real Unitree G1 robot.
RoboTTT introduces a training recipe4 that scales a robot policy's visuomotor context to 8,000 timesteps—three orders of magnitude beyond current state-of-the-art policies—without increasing inference latency. Built on GR00T N1.7, not GR00T 2, it improved overall performance by 87% over a single-step baseline and completed a five-minute, ten-stage assembly task that no baseline finished, according to the project page.
NVIDIA is working with partners including Bosch Research, Siemens, Ambi Robotics, UC Berkeley and the University of Washington. Two new industrial benchmarks, DexBench and VABAR, are being released to measure dexterity and adaptability using Isaac Lab-Arena and Newton.