Physical Intelligence opens its robot foundation models to outside developers in API beta

Physical Intelligence is opening its general-purpose robot foundation models to outside developers. In an announcement dated October 13, 2026, which surfaced on the front page of Hacker News, the company introduced a beta of the Pi API: third parties upload their own demonstration data, fine-tune Physical Intelligence's models on compute the company manages, and run the resulting policies on real robots through a hosted inference gateway.

What the API does

Physical Intelligence argues that deploying robots is hard not only because of hardware and use-case selection, but because developers need models capable enough for the task at hand. Its foundation models, π0 and π0.5, are meant to supply a base of physical knowledge that cuts down the data and engineering effort needed to adapt to new tasks, environments, and robot platforms. Even so, adapting them still required machine learning infrastructure, compute budgets, and data pipelines that many teams cannot easily build themselves.

The API is designed to remove that burden. According to the company, a developer uploads data, starts a fine-tuning job on managed infrastructure, and reaches the finished model through a hosted gateway for inference. Physical Intelligence says the newest model exposed via the API needs less data to fine-tune and delivers better capability than what partners previously had, and because the company can swap in updated models quickly, it intends to keep the API tied to its strongest releases over time.

Early deployments

To test how broadly the models transfer, Physical Intelligence worked closely with companies building autonomy systems for their own applications. One partner, Actor, operates autonomous heavy machinery: its driverless wheel loader lifts material into a dump truck at a live commercial site, with no operator in the cab. Another, Almond, builds Axol, a two-armed robot aimed at developers of physical AI; Axol was shown deburring spacers used in its own assembly using a model its team reached after four days of iterating with the API.

Other cited problems were smaller in scale but varied — a garment turned inside out, an ingredient that kept slipping off a scoop, a part that ended up misaligned. The improvement process, though, looked the same everywhere: gather demonstrations, train a policy, test on hardware, and let the failures point toward the next data to collect. The API is built so partners can run that loop themselves rather than depending on Physical Intelligence at each step.

Deployments spanned Japan, Switzerland, and the United States, and most relied on cloud-hosted inference through the API. That arrangement also exposed weaknesses, the company acknowledged, including network connectivity problems in some settings.

Why it matters

Robotics has traditionally required each application — warehouse automation, construction, surgical assistance — to solve perception, manipulation, and control largely from scratch. Physical Intelligence describes that as effectively needing a dedicated company per problem, and argues the challenge becomes unmanageable if every one of them must solve physical intelligence independently.

A shared foundation model, adaptable through an API, attempts to change those economics the way base models reshaped language and vision software: specialize a general model rather than build intelligence from zero. There is also a feedback loop for Physical Intelligence itself — wider use reveals where the current models break, informing the next generation and, in the company's framing, progress toward genuinely general physical intelligence.

The caveats are real. This is a beta with a limited set of early partners, several deployments are still early-stage, and cloud-hosted inference introduces a dependency that at least some partners found fragile. But making frontier robot models reachable without owned ML infrastructure is a notable opening of the field beyond large labs, and interested developers can apply to the beta by telling the company what they want to build.