Physical Intelligence, the robot foundation model company, said on 13 October that it is opening a beta of its Pi API. Outside developers can upload demonstration data, run fine-tuning on the company’s own compute and then serve the resulting model to their robots through a hosted inference gateway. Access is by waitlist.
Until now the practical route into the company’s work was its open releases, π0 and π0.5, which you had to train and serve yourself. The API moves the expensive part, fine-tuning on a managed cluster, onto Physical Intelligence, and packages the loop that actually produces a working robot policy: collect demonstrations, train, test on the hardware, then let the failures decide what to record next.
The post is thin on what a buyer would want to know. It does not name the model sitting behind the API, saying only that it fine-tunes more efficiently than the published ones and that the underlying models may change quickly. There is no pricing, no general availability date and no benchmark figures. The examples are qualitative: an autonomous wheel loader filling a dump truck, and a dual-arm robot deburring parts, with early users described in the United States, Japan and Switzerland.
The announcement is on the Physical Intelligence blog.