π0.7

PI's 2026 generalist robot model that can be steered with rich prompts and shows early signs of compositional generalization.

π0.7 is a robot foundation model Physical Intelligence released in April 2026. Simply mixing data from different robots, human videos, and autonomous runs (including failures) together in training doesn't work well. π0.7's approach is to attach a richer prompt to every piece of data: language describing the task and its substeps, metadata like speed and quality, a label for whether joint-space or end-effector control was used, and subgoal images showing what things look like after a substep is done (which, at inference, can be generated by a lightweight world model). This lets data of different quality and different ways of doing the same task all be used, and at inference the prompt specifies “how to do it” — steerability. PI reports that without fine-tuning it matches the level of task-specific π*0.6 models at folding laundry, making coffee, and folding paper boxes, and shows the first signs of compositional generalization: it learns to use an air fryer it has never seen through step-by-step language guidance, and folds laundry on a bimanual UR5e with no laundry-folding data of its own.

ExampleAsked to put a sweet potato in an air fryer, the robot only partially completes the task on a few tries when given just one instruction; guided step by step in language by a person, it finishes the task; fine-tuning a high-level policy on that guidance data then lets it generate its own substeps and complete the task fully autonomously.

Also called

pi0.7, pi07

Related

Steerability · π*0.6 · π0.5 · Compositional Generalization · Cross-Embodiment · Hi Robot

Sources

π0.7: a Steerable Generalist Robotic Foundation Model with Emergent Capabilities (arXiv 2604.15483)
π0.7: a Steerable Model with Emergent Capabilities (Physical Intelligence blog)

As of

2026-04