Physical Intelligence Unveils π0.7 for Adaptable Robots

San Francisco-based robotics startup Physical Intelligence announced on April 30, 2025, that its latest AI model, π0.7, enables robots to perform tasks they were never explicitly trained for, marking a potential breakthrough in creating adaptable, general-purpose robotic systems.

The model's key innovation lies in compositional generalization, where it combines previously learned skills to tackle novel problems. Unlike traditional robotic AI, which requires vast, task-specific datasets and retraining for each new function, π0.7 draws on diverse pretraining to synthesize solutions. Researchers at Physical Intelligence, a two-year-old company valued at $5.6 billion after raising over $1 billion in funding, described the results as an early step toward scalable robotic intelligence. The company is reportedly in talks for additional investment that could double its valuation.

A standout experiment involved an air fryer, an appliance the model had encountered in only two brief training instances: one where a robot pushed it closed and another where it placed a bottle inside. Despite this limited exposure, π0.7 formed a working understanding from broader data, attempting to cook a...