Physical Intelligence Launches Managed Robot API Beta

Physical Intelligence, also known as Pi, has opened the Pi API beta to give external developers direct access to its robotic foundation models. Rather than requiring teams to operate their own training clusters and deployment infrastructure for models like the open-source π0 and π0.5, the managed service allows users to upload teleoperation demonstrations, fine-tune base models, and run policies through a hosted cloud inference gateway.

Early design partners span construction, food service, agriculture, and manufacturing, including Actor, Almond, Chef Robotics, AgRobotics, Formic, Foundry, Deft, Ultra, Weave, and Dream Machines across Japan, Switzerland, and the United States. In one test case, Almond developed a functional policy for its dual-arm Axol robot to deburr assembly spacers after four days of iterating with the API. Other partner deployments range from Chef Robotics handling food plating to Actor operating autonomous wheel loaders without human intervention.

For robotics engineers, this platform shifts the focus from building bespoke perception systems and training pipelines toward gathering target environment data. The development loop relies on collecting task demonstrations, evaluating real-world failures such as dropped ingredients or misaligned parts, and retraining the model. However, relying on cloud-hosted inference introduces network latency and connectivity risks into the control path, requiring developers to keep critical safety controls on local hardware.

Access to the beta is currently managed through a waitlist at pi.website/waitlist, and Pi has not yet revealed its pricing structure. Developers adopting the API will still need to manage hardware safety, dataset validation, and local systems, as technical details regarding control frequencies, regional latency limits, and dataset retention policies remain undisclosed.