NVIDIA's Physical AI Stack Demoed on Humanoid

- Project GR00T, which stands for Generalist Robot 00 Technology, is a foundational model intended to enable robots to learn skills by observing human actions and understanding natural language. NVIDIA's stated goal is to solve one of the "most exciting problems in AI today" and enable a leap towards artificial general robotics. - The entire system is designed to run on the NVIDIA Jetson Thor, a new computer for humanoid robots based on the Blackwell GPU architecture. This System-on-a-Chip (SoC) is capable of up to 2,070 FP4 TFLOPS of AI compute, a 7.5x performance increase over the previous Jetson AGX Orin, allowing complex models to run at the edge. - The GR00T N1 model features a dual-system architecture inspired by human cognition: a fast "System 1" for reflexive actions and a slower, more deliberate "System 2" vision-language model for reasoning and planning. - The Reachy 2 robot is an open-source platform developed by Pollen Robotics, which was acquired by the AI company Hugging Face in April 2025. Each of its two 7-degree-of-freedom arms can lift up to 3 kg, and it runs on ROS2, making it a flexible research platform. - Training these models relies heavily on synthetic data generation using NVIDIA's Omniverse platform. NVIDIA generated 780,000 synthetic trajectories—equivalent to 6,500 hours of human demonstration—in just 11 hours, and found that combining this with real-world data improved model performance by 40%. - NVIDIA's strategy is not to build its own humanoids but to provide the core compute hardware and AI models for the industry. Key players like Boston Dynamics, Agility Robotics, Apptronik, and Figure AI are leveraging the GR00T platform. - The company is releasing many of its physical