Physical Intelligence Releases π0.7 Robot Brain Model with Generalization Capabilities

Apr 17, 2026

San Francisco robotics startup Physical Intelligence unveiled its π0.7 robot control model on Thursday, demonstrating compositional generalization capabilities that allow robots to perform tasks they were never explicitly trained on. The model successfully attempted cooking a sweet potato in an air fryer despite having only two air fryer episodes in its training data, achieving a 95% success rate after prompt refinement from an initial 5%. The generalist model matched previous specialist models on tasks like making coffee and folding laundry, though it cannot yet execute complex multi-step tasks autonomously without step-by-step verbal guidance. Physical Intelligence, valued at $5.6 billion with over $1 billion raised and discussions for funding at $11 billion, has declined to provide investors with a commercialization timeline, while the absence of standardized robotics benchmarks makes external validation of the company's claims difficult.

π0.7's compositional generalization and unexpected capabilities

▪Ashwin Balakrishna randomly bought a gear set and asked a robot controlled by Physical Intelligence's π0.7 model to rotate a gear, and the robot successfully performed the task.
▪Physical Intelligence's π0.7 model made a passable attempt at using an air fryer to cook a sweet potato with zero coaching.
▪Physical Intelligence's π0.7 model performed successfully at using an air fryer to cook a sweet potato when given step-by-step verbal instructions.
▪Physical Intelligence's π0.7 model demonstrates compositional generalization, which is the ability to combine skills learned in different contexts to solve problems the model has never encountered.
▪Physical Intelligence published research on Thursday showing that its π0.7 model can direct robots to perform tasks the robots were never explicitly trained on.
▪Sergey Levine stated that once Physical Intelligence's π0.7 model crosses the threshold of remixing skills in new ways, capabilities go up more than linearly with the amount of data.
▪Physical Intelligence's π0.7 generalist model matched the performance of Physical Intelligence's previous specialist models across tasks including making coffee, folding laundry, and assembling boxes.

Limitations and challenges in current robot AI development

▪Lucy Shi stated that after spending about half an hour refining how the air fryer task was explained to Physical Intelligence's π0.7 model, the success rate jumped from 5% to 95%.
▪Physical Intelligence's π0.7 research paper uses hedging language describing the model as showing early signs of generalization and initial demonstrations of new capabilities.
▪Physical Intelligence's π0.7 model is not yet capable of executing complex multi-step tasks autonomously from a single high-level command.
▪Sergey Levine stated that Physical Intelligence's π0.7 model cannot be told to make toast autonomously, but works well when walked through step-by-step instructions.
▪Sergey Levine declined to speculate on when a system based on Physical Intelligence's π0.7 findings might be ready for real-world deployment.
▪Lucy Shi stated that sometimes the failure mode of Physical Intelligence's π0.7 model is not on the robot or model, but on the team not being good at prompt engineering.
▪Lucy Shi stated that an early air fryer experiment with Physical Intelligence's π0.7 model produced a 5% success rate before prompt refinement.

Physical Intelligence's funding and commercial trajectory

▪Physical Intelligence was most recently valued at $5.6 billion.
▪Physical Intelligence is in discussions for a new funding round that would nearly double the company's valuation to $11 billion.
▪Physical Intelligence has raised over $1 billion to date.

Perspective of Physical Intelligence investors

▪Physical Intelligence has attracted over $1 billion in funding without providing investors a commercialization timeline.
▪Physical Intelligence's valuation has grown from $5.6 billion to potential $11 billion despite the π0.7 model not being ready for real-world deployment.

Perspective of Robotics research community

▪The robotics field lacks standardized benchmarks that would enable independent verification of claims like those made about Physical Intelligence's π0.7 model.