NVIDIA unveils GR00T 2 world action model at CoRL: 71% success with one demo
NVIDIA Blog · rss · 2026-11-09
At CoRL, NVIDIA rolled out open robotics tech centered on dexterous manipulation and agentic robotics:
GR00T 2 world action model (built on Cosmos, trained with 100k hours of egocentric human data):
- Completed 40 of 50 evaluated tasks with zero task-specific demos
- 71% average success with one demo per task, 95% with five
- Skill transfer across robot bodies; agents can chain skills via API and recover from errors
- Top-3 across major VLA/world action benchmarks; Foxconn applies it to server manufacturing, Siemens evaluating industrial use
Two Best Paper finalists: GRAIL (humanoid loco-manipulation from world models) and RoboTTT (scales visuomotor context 3 orders of magnitude without added inference latency).
Simulation & data tools: Isaac Lab 3.0 integrates the Newton physics engine for deformables, boosting training throughput up to 2.5x; Skild AI launched the STEEL solver on Newton; a Video-to-Data workflow converts human video into simulated demos; EgoHand open model (40k+ HF downloads in two weeks) reconstructs 3D hand motion.
AlphaNIST trained a single policy to 98% success on all 20 NIST Assembly Task Board tasks (vs 81% for per-task policies) with 39x less GPU time. Partners include Bosch Research, UC Berkeley, and UW.
More from Embodied
- UT Austin to hire 50 tenure-track robotics faculty, touts top GPUs-per-student — DynamicWebPaige · 2026-11-09
- NVIDIA expands Alpamayo platform to reasoning-based whole-body humanoid autonomy — drmapavone · 2026-11-09
- DefFlow: Robots Manipulate Ropes, Cloth into Precise 3D Goal Shapes via Point Flows — drsrinathsridha · 2026-11-09
- Lamp's onboarding guide gets turned into a comic book — dee_hw · 2026-11-09
- No vision, no tactile sensors: humanoid dexterity from proprioception alone at CoRL — Jan_R_Peters · 2026-11-09
- Geometric abstraction cuts robot agent VLM queries 52.8% and speeds manipulation 33% — siddkaramcheti · 2026-11-09