NVIDIA unveils GR00T 2 robot model and Isaac Lab 3.0 simulation framework at CoRL

NVIDIA Corp. introduced GR00T 2, an updated open world action model for robots, and Isaac Lab 3.0, a new version of its open-source robot learning framework, at the Conference on Robot Learning (CoRL) in Austin, Texas.

The company said it trained GR00T 2 on a mix of robot and human data, including 100,000 hours of its own open-source egocentric data, to help robots learn dexterous manipulation. NVIDIA said the model completed 40 of 50 evaluated tasks at least once without task-specific demonstrations, and reached 71% average success across 50 tasks with one demonstration per task, rising to 95% with five.

NVIDIA also said GR00T 2 can learn from human demonstrations and transfer skills between different robot bodies without robot demonstrations. AI agents can use its programming interface and supporting tools to run skills in sequence and recover from errors, according to the company. It claimed the model ranks among the top three across vision-language-action and world action model categories on the RoboCasa365, RoboLab, RoboDojo and MolmoSpaces benchmarks.

Isaac Lab 3.0 and physics simulation

Isaac Lab 3.0 lets researchers use agents to set up training environments with simulation-ready assets, switch between the PhysX and Newton physics solvers, and pick renderers for perception-based training. It also adds a reinforcement post-training library for foundation models such as GR00T, NVIDIA said.

Closer integration with Newton, an open-source physics engine, lets Isaac Lab model deformable objects such as cables and cloth, as well as articulated systems such as multi-jointed robotic hands. On a vision-based dexterous manipulation benchmark, Isaac Lab 3.0 with Newton reportedly reached up to 2.5 times the training throughput of Isaac Lab 2.3 with PhysX.

New benchmarks and partner use

NVIDIA released a developer preview of version 0.4 of Isaac Lab-Arena, an open-source framework for evaluating policies across thousands of variations in parallel. Two new open benchmarks built on it, DexBench and VABAR, aim to measure dexterity and adaptability on industrial tasks. DexBench, developed by RLWRLD with NVIDIA, has 15 tasks across four scenes in retail, logistics, manufacturing and hospitality. VABAR, from UC Berkeley and NVIDIA, covers six industrial tasks.

Skild AI announced STEEL, a rigid-body physics solver built on Newton to speed robot training with Isaac Lab. Flexion showed how Isaac Lab and Newton help train its Reflect robot-intelligence model, and FieldAI is using Newton, Isaac Lab and Isaac Sim to train and validate policies for unstructured environments. Nexuni is testing GR00T 2 for robotic laundry, Foxconn is applying it to server-manufacturing automation, and Siemens is evaluating it for industrial uses, NVIDIA said.

NVIDIA also said it hosted a partnership forum for AI startups in South Korea, where it announced an investment in Point2 and planned further investments in Motif Technologies, RLWRLD and Upstage.