Physical AI refers to AI systems that perceive, reason about, and act in the physical world — spanning robots, autonomous vehicles, smart factories, and energy grids. Unlike traditional rule-based robots, modern physical AI agents combine vision-language-action (VLA) models with reinforcement learning to handle novel real-world situations. Three key breakthroughs are driving recent progress: open robotics foundation models trained on massive datasets, physics-aware world models that close the sim-to-real gap, and dramatically faster GPU compute. Training involves building simulated environments with domain randomization, applying reinforcement learning through trial and error, then iteratively feeding real-world failure data back into simulation until the model generalizes reliably to messy reality.
Sort: