NVIDIA unveils GR00T 2 and Isaac Lab 3.0 to teach robots delicate work

Four things, all aimed at robots that need a finer touch.

• GR00T 2: an open model that helps robots learn skills like handling small objects.
• Isaac Lab 3.0: the latest version of NVIDIA's open source tool for training robots in simulation.
• Two benchmarks: DexBench and VABAR, open tests for robot skill.
• Agentic workflows: ways to let AI agents write and test robot programs.

NVIDIA says the focus is moving from locomotion, which means walking, rolling and hopping, to manipulation, which means using hands and arms. Two NVIDIA papers are also among six finalists for the Best Paper Awards at CoRL, out of thousands of submissions.

Most chatbots work with words. A robot has a harder job. It has to see its surroundings, guess how its actions will change them and then do many steps in a row without dropping anything.

A world action model is built for that. NVIDIA says it lets a robot perceive, anticipate and carry out complex tasks with a high degree of precision. GR00T 2 is NVIDIA's new one, an open model built on its Cosmos platform.

It learned from many kinds of data, including 100,000 hours of NVIDIA's open source egocentric data. Egocentric means filmed from the person's own point of view, like a camera on your head while you work.

NVIDIA shares several numbers from its own tests:

• Out of the box: it completed 40 of 50 evaluated tasks at least once without task-specific demonstrations.
• One example per task: 71% average success across 50 tasks.
• Five examples per task: 95% average success.

It can also learn from human demonstrations and move skills between different robot bodies, without needing demonstrations on the robot itself. AI agents can use its application programming interface, a doorway that lets one program control another, to run skills in order and recover from errors.

On public leaderboards, NVIDIA says GR00T 2 ranks among the top three models on the RoboCasa365, RoboLab, RoboDojo and MolmoSpaces benchmarks, with several wins. These are NVIDIA's own figures.

In real life Think of a new warehouse worker. Show them a task once and they get it right most of the time. Show them five times and they almost always do. That is the pattern NVIDIA reports for GR00T 2.

Several companies are trying the model in real settings.

• Nexuni: testing it for robotic laundry handling.
• Foxconn: applying it to factory automation for server manufacturing.
• Siemens: evaluating it across industrial uses.
• MicroAGI: using a version trained further with reinforcement learning, where a robot learns by trying and getting rewards, to swap Ethernet cables with a robotic arm.

MicroAGI builds and calibrates its training worlds automatically with Isaac Lab 3.0 and Newton, and will demonstrate the approach at CoRL.

A simulation is a virtual world that follows the rules of physics. A pretrained model gives a robot a core set of skills. The simulator lets researchers refine those skills at huge scale before touching real hardware.

The robot can break as many virtual plates as it needs, and nobody has to sweep. Failures in simulation become lessons, and researchers can run thousands of simulated robots at the same time.

Isaac Lab 3.0 lets researchers use agents to set up training worlds quickly. It lets them switch between two physics engines, NVIDIA PhysX and Newton, and pick different renderers for training that depends on vision. It adds a new library for reinforcement post-training of models such as GR00T.

Newton is an open source physics engine, the code that decides how virtual things move, bend and collide. Tighter integration with Isaac Lab now lets it model soft objects such as cables and cloth, and robot hands with many joints.

NVIDIA shows a simulated strawberry squishing under a robot's grip. That matters because real transfers from simulation fail when the physics is off.

By the numbers: in a vision-based dexterous manipulation benchmark, Isaac Lab 3.0 with Newton achieved up to 2.5 times the training throughput of Isaac Lab 2.3 with PhysX. Throughput means how much training the system gets through in the same time.

Partners are building on it. Skild AI announced STEEL, a new rigid-body physics solver built on Newton to speed up training with Isaac Lab. Flexion, FieldAI and Dyna Robotics are also using Isaac Lab or Newton to train robots for warehouses, hotels, food service and unstructured real-world spaces.

AlphaNIST, built in Isaac Lab, is the example NVIDIA points to. One policy, which is the robot's decision-making brain, learns all 20 tasks on the NIST Assembly Task Board. The parts are millimeter-sized.

The robot gets only a yes-or-no signal on whether it succeeded. Even so, it reached more than 98% training success, against 81% for separate per-task policies, using roughly 39 times less GPU time.

The method behind it, Success-Guided Sampling, comes from the University of Washington and NVIDIA. It puts practice time on tasks the robot succeeds at about half the time, which are the ones it is still learning. NVIDIA says this lets training scale beyond a million parallel environments.

Learning a skill is one thing. Doing it reliably when conditions change is another. NVIDIA's Isaac Lab-Arena is an open source framework for testing policies across thousands of variations at once. Version 0.4 arrives at CoRL as a developer preview.

Two new open benchmarks use it:

• DexBench: from RLWRLD with NVIDIA. It has 15 tasks across four scenes: retail, logistics, manufacturing and hospitality.
• VABAR: from UC Berkeley and NVIDIA. It covers six industrial tasks with standard variations, difficulty tiers and success scores.

On the programming side, UC Berkeley's Graph-as-Policy (GaP) uses agents to connect single skills into full robot programs. The agents test these in NVIDIA's simulators and fix them after failures. Engineers can read and check the programs before deployment. Ambi Robotics, Bosch Research and Siemens are exploring it.

"To me, agentic robotics is the most exciting technological development since the internet in 1995," said Ken Goldberg, a UC Berkeley robotics professor and a lead on VABAR and GaP.

• If you work in robotics: the new model, simulator and benchmarks are open, so you can test them yourself.
• If you work in a factory or warehouse: expect more pilots for tasks like cable handling and sorting, since Foxconn and Siemens are already testing.
• If you are a student: the University of Texas at Austin already uses Isaac tools to teach robot skills before hardware is available.
• If you are just curious: this will not put a robot in your home soon. It is groundwork for industry.