Physical Intelligence’s π0.7 makes robot skills recombine

Physical Intelligence is making a robotics claim that matters because it targets the bottleneck behind many brittle robot demos: training a new specialist system for each task. In its April 16, 2026 research post, the company describes π0.7 as a general-purpose vision-language-action model that can follow new language commands and perform tasks not seen in training data.

The key phrase is compositional generalization. Physical Intelligence says π0.7 can recombine skills from different tasks to handle new problems, including new kitchen appliances and laundry folding on a robot for which no laundry-folding data was collected. The company compares this loosely to how LLMs combine known concepts in new formats, but robotics adds physical motion, robot morphology, and scene variation, so the claim is harder to validate than a text benchmark.

The UR5e transfer is the detail to watch

The most concrete example is laundry folding on a bimanual UR5e system. Physical Intelligence says the source robot and the UR5e differ substantially in size, positioning, and morphology, and that it collected no training data for this task with the UR5e setup. Even so, π0.7's success rate matched the zero-shot success rate of expert human teleoperators who had collected the original training data and then tried the UR5e for the first time. Those teleoperators averaged 375 hours of teleoperation experience.

The method is not just “more data.” π0.7 is trained with varied prompt structures: language, metadata, control modality labels, and visual subgoal images. Those prompts describe not only what the robot should do, but how it should do it. At test time, the model can receive standard language instructions, desired strategy information, and visual subgoals generated by a lightweight world model.

This should not be read as a deployed product claim. The source uses careful language: “first signs” and “initial signs,” not finished general-purpose robotics. The result still needs outside replication, standardized robotics benchmarks, and clearer failure analysis. Still, the direction is important. If one model can approach task-specific specialist performance while handling unseen combinations, the bottleneck in embodied AI shifts from building a new model for every task toward instruction design, safety boundaries, and evaluation.