Physical Intelligence puts its robot models behind an API beta

Physical Intelligence opened its Pi API beta, which lets outside developers train on and run the company's general-purpose robotic foundation models, from the π0 and π0.5 lineage, without owning the ML infrastructure. It is the first major step in making these models available beyond the company's own deployments, and the pitch is direct: upload data, launch fine-tuning on compute Pi manages, and access the fine-tuned model through a hosted gateway.

How the API works

The workflow has three parts. You upload your demonstration data. Pi runs the fine-tuning on its own compute and infrastructure. Inference then happens through a hosted gateway, so the model lives in Pi's cloud rather than on your hardware.

The company says the latest model available through the API is more data-efficient to fine-tune and improves on what its partners were running, and that it plans to keep swapping in newer models as it ships them. Those are vendor claims, and there are no published numbers yet, but the structural point stands: the alternative to this API is building the data pipeline, the training compute, and the serving stack yourself.

Who is already using it

Pi gave a small group of companies early access to test whether they could improve autonomy on their own, and the blog post shows the results. Actor runs an autonomous wheel loader in a commercial deployment in Utah, with no operator inside the machine. Almond's dual-arm robot Axol deburrs parts used in its own assembly, using a model the team reached after four days of iteration with the API. Other names in the list include CloudChef, AgRobotics, Chef Robotics, Formic Robotics, Foundry, Frontier Machines, and Ultra.

Most of the deployments shown, from Japan to Switzerland to sites across the United States, ran on cloud-hosted inference through the API. The company also flags where it fell short: network connectivity was a real problem for some of them.

The iteration loop the API is built around

The post argues that the failure modes differ per task, an inside-out garment, an ingredient that won't stay in a scoop, a part that lands in the wrong position, but the improvement process is the same everywhere: collect demonstrations, train, test on the robot, and let the failures decide what data to collect next. The API exists so developers can run that loop themselves on Pi's latest models instead of rebuilding the stack every time a model improves.

What it means for you

If you run robots, the practical question is data gravity. Cloud-hosted inference means your demonstration data and model weights live in Pi's infrastructure, and some of the early partners already ran into connectivity limits. The beta is not open sign-up: you tell the company what you want to build and it decides. If your use case needs models to stay on site, or on the machine, this beta is not for you yet. If it doesn't, it removes the largest chunk of the self-hosting stack and is worth a look.