Practice Makes Precise: At CoRL, NVIDIA and Partners Reveal How Robots Rehearse Delicate Work in Simulation
NVIDIA unveiled GR00T 2, an open frontier world‑action model built on NVIDIA Cosmos that leverages 100,000 hours of open‑source human egocentric data to teach robots dexterous manipulation. The model completed 40 of 50 evaluated tasks without task‑specific demonstrations, achieved 71 % success with a single demonstration per task (rising to 95 % with five), and scales visuomotor context three orders of magnitude beyond state‑of‑the‑art policies without added inference latency. These results place GR00T 2 in the top three across vision‑language‑action and world‑action benchmarks (RoboCasa365, RoboLab, RoboDojo, MolmoSpaces), enabling practical transfer of human skills to diverse robot bodies for industrial use cases. Engineers can now integrate a ready‑to‑use API to sequence skills and recover from errors, accelerating the path from simulation to real‑world deployment.
⚡ Key Takeaways
- Completed 40 of 50 evaluated tasks at least once without task‑specific demonstrations.
- Scales visuomotor context three orders of magnitude beyond state‑of‑the‑art policies without increasing inference latency (as demonstrated in RoboTTT).
- Achieves 71 % average success across 50 tasks with one demonstration per task, increasing to 95 % with five demonstrations.
- Provides an application programming interface (API) for AI agents to run skills in sequence and recover from errors.
- Requires access to the GR00T 2 model built on NVIDIA Cosmos and the associated simulation tools.
🔧 Tools & Libraries
GR00T 2’s high‑success, low‑demo learning dramatically reduces data collection overhead for precision manipulation, letting production engineers deploy adaptable robot policies faster and with fewer custom demonstrations.
✅ Practical Steps
- Apply the concepts from this article to your own system design.
Want the full story? Read the original article.
Read on NVIDIA Blog ↗