Physical Intelligence, the San Francisco-based robotics startup backed by major AI investors, has unveiled π0.7, a new robotic foundation model that it says can perform tasks it was never explicitly trained to do. The release marks a notable step toward general-purpose robotics, where one model can adapt across tasks, environments, and even robot hardware with far less retraining than traditional systems require.
TL;DR
- Physical Intelligence introduced π0.7 on April 16, 2026.
- The company says the model shows compositional generalization in robotics.
- It can follow new language instructions and tackle unfamiliar tasks such as using an air fryer.
- π0.7 also transfers capabilities across different robot setups, including laundry folding on new hardware.
- The release strengthens the push toward general-purpose robot brains.
In its research release, Physical Intelligence said π0.7 delivers a step change in generalization. According to the company, the model can match fine-tuned specialist systems on dexterous manipulation tasks while also following fresh language commands and handling tasks absent from its training data.
The standout example involved an air fryer. The company said it did not collect demonstrations for that exact appliance task, yet the model was still able to make progress with a zero-shot prompt and perform the task more effectively when given step-by-step language coaching. Physical Intelligence said that after repeated coaching, the same instructions could be used to fine-tune a higher-level policy that enabled more autonomous execution.
The startup’s researchers themselves were surprised by some of the results. Co-founder Sergey Levine said the system appears to be moving past the stage of only doing exactly what the data directly taught it. Researcher Lucy Shi also noted that prompting quality had a major effect on outcomes, with one early air fryer experiment reportedly jumping from a 5% success rate to 95% after the team refined how the task was explained.
Physical Intelligence also highlighted cross-embodiment transfer as a key breakthrough. The company said π0.7 could control a UR5e bimanual robot to fold laundry despite having no laundry-folding data for that exact hardware. That matters because one of robotics’ biggest bottlenecks is the cost of collecting fresh data and retraining models for each robot form factor, gripper, and environment.
Under the hood, the company attributes the improvement to broader and more diverse data, combined with multimodal prompting. In practice, that means π0.7 can use not only language instructions, but also metadata, control-modality labels, and visual subgoals to understand both what to do and how to do it. The approach is designed to help the model recombine previously learned skills into new behaviors.
Topics for more insights:
The bigger takeaway is that robotics firms are increasingly chasing the same kind of scaling curve that transformed language models. If Physical Intelligence’s claims continue to hold up, π0.7 could strengthen the case that robot foundation models are finally starting to generalize in the real world, not just memorize narrow workflows. That would be a meaningful shift for warehouses, factories, and homes, where adaptable robots remain one of the industry’s biggest promises.
Join The Discussion