$NVDA Debuts Open GR00T 2 Robotics Foundation Model as Physical AI Race Intensifies Across $TSLA, $GOOGL, and...

$NVDA has unveiled GR00T 2, a next-generation open foundation world-action model (WAM) designed to learn robotic manipulation skills from human demonstrations and simultaneously predict future video and physical motor actions.
Built on a diffusion transformer architecture with robot-specific output heads, GR00T 2 enables cross-embodiment deployment across parallel grippers, multi-fingered dexterous hands, and humanoid robots. According to $NVDA research benchmarks, the model achieves a 40% zero-shot success rate across 50 unseen tasks and scales to 95% accuracy with five demonstrations per task after 30 minutes of fine-tuning on a single RTX 5090 GPU. Alongside the flagship model, $NVDA introduced a 2-billion-parameter GR00T 2 Turbo variant capable of running real-time whole-body loco-manipulation inference in 124 milliseconds on Jetson Thor hardware.
The release highlights a rapidly intensifying race to build generalist foundation models for physical AI and humanoid robotics across major technology and automotive players:
- $TSLA: Scaling Optimus Gen 3 humanoid production alongside an end-to-end AI stack powered by its Full Self-Driving (FSD) neural network architecture, AI5 compute, and Grok integration.
- $GOOGL DeepMind: Advancing Gemini Robotics 1.5 and Gemini Robotics-ER 1.6 for spatial reasoning, multi-view understanding, and low-latency on-device robotic control.
- Hyundai / Boston Dynamics: Integrating Google DeepMind’s Gemini Robotics models into its commercial electric Atlas humanoid platform for enterprise and industrial deployment.
- Unitree Robotics: Pairing mass-market humanoid hardware such as the G1 with its open-source UnifoLM vision-language-action model.
- Figure AI: Running its proprietary Helix vision-language-action model across commercial Figure 03 humanoid deployments in automotive manufacturing.