π0.7 (pi-0.7)

π0.7, released by Physical Intelligence on 16 April 2026, is a roughly 5-billion-parameter vision-language-action model built from a 4B vision-language backbone, a video-history encoder and an 860M action expert. Its central idea is steerability: instead of a single terse instruction, the model is conditioned during training on rich context — detailed language, visual subgoal images produced by a lightweight world model, metadata describing how a task should be executed (speed, care, quality) and labels for the control modality (joint or end-effector). The same weights can therefore be pushed toward different strategies for the same task without fine-tuning. The payoff Physical Intelligence reports is emergent compositional generalization: π0.7 recombines skills learned on separate tasks to handle situations it was never trained on, including operating kitchen appliances it has not seen and folding laundry on a robot embodiment for which no laundry data exists. On laundry folding, espresso making and box folding it matches or beats specialist policies optimised with reinforcement learning, and on a novel embodiment it reached the level of an expert human teleoperator zero-shot. Training mixes demonstrations from static, mobile, single-arm and bimanual platforms in labs, homes and the wild, human demonstration video, autonomous episodes from earlier policies, web-scale vision-language pre-training and the open DROID dataset. Unlike π0, which Physical Intelligence released with open weights, π0.7 is a closed model — the company has published the research and demonstrations but not the checkpoints.

▸Videos

No videos yet.

▸Reviews