Physical Intelligence Pi0.7

Physical Intelligence's pi0.7 model represents a further step in what general-purpose robotic AI can do in practice. The capability the company highlights is compositional generalization: a robot combining skills learned in different training contexts to perform tasks it was never explicitly shown — precisely the capability gap that has separated narrow task-specific robots from the general-purpose robotic assistants researchers have pursued for decades. The company has described demonstrations where a robot combined cooking-skill training with separate step-by-step coaching to complete a multi-step cooking task it hadn't been directly shown end-to-end, calling it an early but meaningful step toward general-purpose robotic AI — framing that acknowledges both the significance and the distance remaining. Physical Intelligence is behind the pi0 family of vision-language-action models, which it introduced in late 2024, part of a wave of foundation-model approaches to robot control that also includes Google DeepMind's RT-2 and UC Berkeley's Octo. The company has attracted substantial venture backing on the bet that its compositional-generalization approach will prove more scalable than narrower, task-specific deployment strategies. The research trajectory is consistent with the broader pattern in AI: earlier pi0 versions established core dexterous manipulation capabilities, and pi0.7 adds compositional generalization, with the roadmap implying continued expansion of the skill vocabulary. The critical question is whether compositional generalization transfers from demonstration tasks to the chaotic variability of real-world environments.