Nvidia debuts GR00T 2 robotics model, 71% on 50 tasks
TL;DR
- The 95% five-shot success rate covers only pre-screened tasks; the 40% zero-shot baseline is the more honest measure of general capability.
- Pretraining drew on 402,730 hours of human video, making data scale the primary competitive moat rather than novel architecture.
- Agent-monitoring overhead caused robots to idle in 38% of episodes, a real-time latency cost absent from Nvidia's official announcement.
Nvidia says its new GR00T 2 world action model completed "40 of 50 evaluated tasks at least once without task-specific demonstrations," averaged 71% success across those 50 tasks with one demonstration per task, and reached 95% with five. The model was introduced at the Conference on Robot Learning in Austin, Texas. Nvidia's own write-up describes it as an "open frontier world action model built on NVIDIA Cosmos," trained on 100,000 hours of Nvidia open-source human egocentric data.
Alongside the model, Nvidia released Isaac Lab 3.0. In vision-based dexterous manipulation benchmarks the company cites, Isaac Lab 3.0 with the Newton physics solver "achieved up to 2.5x the training throughput of Isaac Lab 2.3 with PhysX." On the new AlphaNIST benchmark, a single policy covering 20 tasks hit 98% training success versus 81% for separate per-task policies, using roughly 39x less GPU time.
Two industrial benchmarks debuted with the release: DexBench, with 15 tasks across four industrial scenes, and VABAR, which evaluates robots across six industrial tasks with standardized variations. Nvidia says GR00T 2 "ranks among top three models across vision language action and world action model categories."
"To me, agentic robotics is the most exciting technological development since the internet in 1995," said Ken Goldberg, a UC Berkeley robotics professor leading the VABAR work. Collaborators listed in the announcement include Ambi Robotics, Bosch Research, Flexion, FieldAI and Siemens.
The release lands in a thick stretch of robot-learning news: our robotics tracker has logged 191 stories in the last 90 days, including Nature Will's 50B model for robot hands the day before. The Nvidia post gives no public release date or license for the GR00T 2 weights or the 100,000-hour dataset.
What others are reporting
-
Humanoids Daily Read →
Places GR00T 2 against OpenAI's GPT-6 Astra; names Foxconn, Nexuni, and MicroAGI as early commercial partners beyond Siemens and Bosch.
Robotics is no longer an elitist sport anymore (Jim Fan, NVIDIA director of robotics)
-
Artiverse Read →
Sources UC Berkeley's Ken Goldberg on the imitation-learning plateau; flags hardware bottlenecks including 0.3mm semiconductor tolerances no current approach has solved.
The imitation learning paradigm has hit a plateau (Ken Goldberg, UC Berkeley)
-
Unite.AI Read →
Scrutinizes benchmark methodology: the 95% figure was drawn from pre-screened tasks; robot idled 38% of episodes due to agent-monitoring latency not in Nvidia's framing.
-
RuntimeWire Read →
Frames Isaac Lab 3.0 and shared benchmarks (DexBench, VABAR) as open infrastructure for smaller robot builders, positioning this as a stack release beyond a single model.
GR00T 2 tests whether human video can teach humanoids dexterous industrial work with only a handful of demonstrations.
Originally reported by blogs.nvidia.com
Read the original article →Original headline: Nvidia Unveils GR00T 2 Robotics Foundation Model at CoRL, Hits 71% One-Shot Success on 50 Tasks