Physical Intelligence shows robot model with LLM-like generalization, flaws included

Physical Intelligence introduced π0.7, a robot foundation model built on Gemma3 with an action expert and richer task context.

The company says one general model can now match older specialists, transfer between robot arms, and learn new tasks from human step-by-step coaching.

A highlighted air-fryer demo is presented as evidence of compositional generalization.

But observers question whether the model is truly generalizing or benefiting from structurally similar training examples, renewing contamination concerns.