Nvidia’s GR00T 2 and Isaac Lab 3.0 push robot learning from walking to precise hand work

At the Conference on Robot Learning in Austin, Nvidia announced GR00T 2, an open world action model built on Cosmos (Cosmos tools), plus Isaac Lab 3.0, an updated evaluation framework, and new open benchmarks with partners. The headline is precision: robots practicing delicate tasks like connecting cables and assembling small components in simulation, then carrying those skills to real hardware.

Nvidia’s numbers for GR00T 2 include completing 40 of 50 evaluated tasks at least once with no task-specific demonstrations, and 71% average success across 50 tasks from a single demonstration each, rising to 95% with five. Foxconn, Siemens and Nexuni are testing or evaluating it. If that holds outside a vendor blog, teaching a robot a new industrial job gets much cheaper. Two Nvidia papers, GRAIL and RoboTTT, are among six CoRL Best Paper finalists, per the company.

This is a vendor announcement, so benchmarks are claims, not verdicts. The direction is right, though. Locomotion demos impress, but factories and warehouses pay for hands that do small, exact things all day, the same bet behind Google DeepMind’s factory kitting pilot (Gemini Robotics 2.5).

Two pieces stand out. AlphaNIST is a single Isaac Lab policy handling all 20 tasks on the NIST Assembly Task Board using only a pass-or-fail signal, reported at over 98% training success versus 81% for separate per-task policies, with about 39 times less GPU time. The University of Washington and Nvidia method behind it focuses practice on tasks the policy succeeds at about half the time, an idea likely to travel. And Isaac Lab-Arena, DexBench and VABAR measure failure across thousands of variations, which must come before trusting a robot near people or expensive equipment. Isaac Lab 3.0 also integrates the open Newton physics engine for deformable objects like cables and cloth, reaching up to 2.5x the training throughput of Isaac Lab 2.3 with PhysX on one vision-based dexterity benchmark.

Berkeley’s Graph-as-Policy has agents stitch skills into robot programs, test them in simulation and revise based on failures, keeping output readable for engineers. Interpretable code is a real safety benefit. But agents writing programs for physical machines means a bad plan has physical consequences, and simulation never captures everything a shop floor throws at a robot, as Nvidia’s own lab found when classical engineering beat end-to-end learning (machines that make machines).

Workers in precise assembly and handling are the obvious exposure if this matures. And the stack, from model to physics engine to benchmarks, increasingly runs through one company’s tools, even when open source. The key test is whether independent labs reproduce these gains on hardware Nvidia doesn’t showcase. More in our Nvidia coverage and robotics coverage.