Training General Robots for Any Task: Physical Intelligence’s Karol Hausman and Tobi Springenberg

Physical Intelligence is building robotic foundation models that can enable any robot to perform any task. Karol and Tobi explain that robotics has been bottlenecked by intelligence, not hardware, and that the classical approach of breaking robotics down into perception, planning, and control was fundamentally flawed. Their newest model, PI-STAR 0.6, uses reinforcement learning to learn from experience, achieving robust real-world performance, such as robots making coffee for 13 hours straight and generalizing across tasks from surgical robots to drone flying. The model architecture is analogous to vision language models, pre-trained on robotics data and internet data, with an added action model to drive the robot.