The Race to Create General-Purpose Robots | Karol Hausman & Lachy Groom on TBPN

Physical Intelligence aims to develop foundational AI models capable of controlling robots to perform diverse, long-horizon tasks in novel environments. Unlike previous robotic systems that required task-specific programming or localized data collection, this approach leverages end-to-end learning to achieve generalization, enabling robots to handle complex chores like cleaning kitchens or folding laundry. While simulation remains effective for locomotion, real-world data collection across varied environments is essential for mastering physical manipulation. The current milestone, PI05, demonstrates a significant leap in reliability, moving beyond static, pre-programmed behaviors toward adaptive intelligence. Although consumer-ready deployment remains a long-term goal requiring higher reliability, the integration of vision-language models allows these robots to understand intent and interact with their surroundings, marking a shift toward embodied AI that learns from experience rather than rigid, hard-coded instructions.