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

Physical Intelligence is building robotic foundation models capable of performing various tasks across different robotic form factors. Karol and Tobi from Physical Intelligence discuss their PI-STAR 0.6 model, which uses reinforcement learning to improve robot performance and enable real-world deployment. They highlight the importance of focusing on intelligence as the key bottleneck in robotics, rather than hardware limitations. The discussion covers the challenges of generalization, the need for diverse datasets, and the shift from demonstration data to RL-driven learning from experience. The PI-STAR 0.6 model has shown impressive results, such as robots making coffee for 13 hours straight and a 2x increase in policy throughput on various tasks.