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

Physical Intelligence is addressing the robotics bottleneck by developing general-purpose foundation models that move beyond traditional modular architectures—perception, planning, and control—toward end-to-end learning. By training models on diverse datasets and integrating reinforcement learning from real-world experience, these systems achieve robust performance across varied tasks like coffee preparation and box folding. The PI-STAR 0.6 model demonstrates that robots can learn from their own operational data, enabling them to recover from failures and adapt to new environments without manual rule-writing. This approach prioritizes intelligence over specific hardware configurations, aiming to create versatile agents capable of performing complex physical tasks. As these models scale, the transition from simulated training to real-world deployment creates a self-reinforcing data loop, significantly expanding the potential aperture for autonomous robotic applications in both domestic and industrial settings.