The Signal · Issue 07 · Physical Intelligence
NVIDIA says its new robot model learns a task from one demonstration, 71% of the time
At the Conference on Robot Learning, NVIDIA released GR00T 2, an open world action model, plus Isaac Lab 3.0 and new industrial benchmarks. The figures are NVIDIA's own. The more important shift is that robots now rehearse delicate work in simulation before touching anything real.
Facts are attributed and linked to their sources. The analysis and opinions are the author's own.
What it is
On 9 November, at the Conference on Robot Learning in Austin, NVIDIA introduced GR00T 2, an open world action model built on its Cosmos models and trained on robot and human data, including 100,000 hours of its open-source egocentric human video. NVIDIA says it completed 40 of 50 evaluated tasks at least once without task-specific demonstrations, reached 71% average success across 50 tasks with one demonstration per task and 95% with five, and can transfer skills between different robot bodies. AI agents can call it through an API to run skills in sequence and recover from errors.
NVIDIA also released Isaac Lab 3.0, its open-source robot-learning framework, which it says reaches up to 2.5 times the training throughput of the previous version on a vision-based dexterity benchmark, and two new open industrial benchmarks, DexBench and VABAR. Partners named include Foxconn, applying GR00T 2 to factory automation for server manufacturing, and Siemens, which is evaluating it. In one research example, a single policy learned all 20 tasks on the NIST assembly task board with more than 98% training success, against 81% for separate per-task policies, using roughly 39 times less GPU time.
Why it matters
These are vendor numbers on vendor-chosen tasks, and "completed at least once" is a low bar for a factory. Still, the direction is clear. Robot learning is moving from walking to fine manipulation, such as connecting cables and assembling small parts, and the training happens in simulation first, at a scale no physical lab can match.
Simulation is becoming the place where machines rehearse. The question for anyone financing them is how often the rehearsal fails in the real world.
For lenders and insurers this matters before it matters on the factory floor. Equipment finance, warehouse automation loans and operational-risk cover will increasingly depend on robots whose skills were learned in simulation. In my view the useful due-diligence question is not "which model does it run?" but "what is its measured success rate on this task in this site's conditions, and what happens on failure?" New benchmarks such as DexBench and VABAR, with standard variations and difficulty tiers, are a step toward answers that can be compared.
What I'd watch
- Independent results for GR00T 2 on the public benchmarks NVIDIA cites, and on DexBench and VABAR.
- Production figures, not pilots, from Foxconn or Siemens.
- Whether one-demonstration success rates hold outside lab conditions.