AI BriefingNVIDIA Releases GR00T 2 Foundation World-Action Models for Cross-Embodiment Robotics
·2026.11.10 09:00
Key point
NVIDIA releases GR00T 2, a foundation world-action model that predicts future video and robot actions, enabling zero-shot transfer from human demos and real-time control of Unitree G1 humanoids via the GR00T 2 Turbo variant.
Details
- NVIDIA GEAR Labs released the GR00T 2 whitepaper, introducing a Foundation World-Action Model (WAM) that jointly denoises future video and robot actions. Code is available in December.
- Cross-embodiment adaptation allows robots to perform tasks from demos of other embodiments; human demos enabled Dexter hands to install filters with 65% success, while handheld UMI demos allowed YAM arms to assemble toy dump trucks with 60% success.
- Fast adaptation via LoRA fine-tuning on a single RTX 5090 raised success rates from 40% (zero-shot) to 95% (5-shot) on 50 unseen tasks, using less than 3 minutes of robot data per task and a 30-minute fine-tuning process.
- Agentic control combined with GR00T 2 achieved 97-98% success in 100 assembly trials for toy planes, tractors, and Hot Wheels, outperforming the post-trained policy without the agent (71-84%) and supervised post-training alone (23-69%).
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Runway Releases Praxis-1 Open-Weight Robotics Model