NVIDIA just introduced GR00T 2, a new world-action AI model that helps robots learn physical skills from human videos and adapt to unfamiliar tasks with just five demonstrations. In NVIDIA's experiments, GR00T 2 achieved 95% average success across 50 evaluated robot tasks after five demonstrations and only 30 minutes of fine-tuning on a single RTX 5090. This is a major step toward robots that can learn from watching humans instead of requiring extensive task-specific programming. What makes GR00T 2 different? • World-action modeling: predicts future video and robot actions together, learning how the physical world changes as a robot moves. • Massive training dataset: 402,730 hours of human first-person video, 10,184 hours of handheld-gripper demonstrations across 2,897 tasks, and 23,500 hours of robot teleoperation. • Zero-shot capabilities: 40% average success on 50 evaluated tasks without demonstrations on the target robot. • Rapid adaptation: 95% average success after five demonstrations and a 30-minute LoRA fine-tune on one NVIDIA RTX 5090. • Dexterous manipulation: 85% success inserting NVIDIA GB300 memory modules and 65% assembling a toy truck using robotic hands. • Real-world reinforcement learning: success rates reaching 88–96% across three contact-rich manipulation tasks. • Open robotics data: NVIDIA has released 100,000 hours of human video from its Human-1M dataset on Hugging Face. The bigger breakthrough is the ability to transfer skills from human videos and other robots into entirely different robotic bodies. GR00T 2 doesn't just predict what action to take next. It also predicts what the world should look like after that action. These are NVIDIA's reported research results on selected evaluation tasks, not proof that the system can master any physical task with five demonstrations. Model code is scheduled for release in December 2026. Explore the research: GR00T 2: https://research.nvidia.com/labs/gear/gr00t-2/ ASPIRE agentic robotics: https://research.nvidia.com/labs/gear/aspire/ SimFoundry: https://research.nvidia.com/labs/gear/simfoundry/ FLORA: https://research.nvidia.com/labs/gear/flora/ Hardware prototypes: https://research.nvidia.com/labs/gear/hardware-prototypes/ Human-1M Dataset Explorer: https://huggingface.co/spaces/nvidia/PhysicalAI-Human-1M-Explorer Are we getting closer to robots that can learn almost any physical skill by watching humans? #NVIDIAGR00T #PhysicalAI #Robotics