PHYSICAL AI
NVIDIA brings GR00T 2 and Isaac Lab 3.0 to CoRL, and robots get a rehearsal studio
November 9, 2026 · Updated November 10, 2026
#PhysicalAI #Simulation #ProductLaunch #Research
Every good performer rehearses, and NVIDIA wants robots to do the same, just with a lot more takes. On November 9, 2026, the company introduced GR00T 2, an updated open world action model, and Isaac Lab 3.0, the latest version of its open-source robot learning framework. The reveal came at the Conference on Robot Learning (CoRL) in Austin, Texas.
GR00T 2 was trained on a mix of robot and human data, including 100,000 hours of NVIDIA's open-source egocentric footage, which is basically a lot of people's-eye views of work being done. By NVIDIA's own testing, it finished 40 of 50 evaluated tasks at least once with no task-specific demonstrations. Give it one demonstration per task and it averages 71% success. Give it five and that climbs to 95%. It can also carry skills learned from humans across to different robot bodies, no robot demos required.
Isaac Lab 3.0 is the practice room. It works with the Newton physics engine so thousands of simulated robots can train at once, and NVIDIA says it reached up to 2.5 times the training throughput of Isaac Lab 2.3 with PhysX in a vision-based dexterous manipulation benchmark. Partners joined the show: Skild AI announced STEEL, a rigid-body physics solver built on Newton, while Flexion showed how it trains its Reflect model. MicroAGI is using GR00T 2 for Ethernet cable swapping on a YAM robotic arm, a task that rewards a steady hand.
The fine print
- Announced: November 9, 2026, at CoRL in Austin, Texas
- Headline numbers: 40 of 50 tasks with zero demos, 71% with one demo, 95% with five (NVIDIA's own tests)
- Training data: 100,000 hours of open-source egocentric human data, plus robot data
- Simulation speed: up to 2.5x the training throughput of Isaac Lab 2.3 with PhysX in one benchmark
- Partners named: Ambi Robotics, Bosch Research, FieldAI, Flexion, Siemens, Skild AI, RLWRLD and others
- Still unclear: NVIDIA's announcement gives no general release date or licence details for GR00T 2
What it means for automation buyers
This is a tools-and-research story rather than a product you can order for the shop floor next quarter. Still, if robots can learn fine handling tasks like cable connection and small-part assembly from a few demonstrations, programming time for cobot cells could shrink over the coming years. For now, treat the success rates as lab results from the vendor, ask integrators how they validate skills in your own conditions, and keep an eye on when the model becomes available. Meanwhile, proven cobot applications with predictable payback remain the safer purchase.