The Pi API turns fine-tuning and inference for robot foundation models into a hosted service. On October 13, 2026, Physical Intelligence announced on its official blog that the Pi API has entered beta: developers upload their own robot data, launch fine-tuning on compute and infrastructure that Physical Intelligence manages, and run inference on their robots through a hosted gateway, with evaluation data guiding the next iteration. Applications are open, and teams are asked to describe what they want to build.
What it removes is exactly the part robotics teams lack most
The promise of general robot foundation models such as pi0 and pi0.5 was always to cut the data and effort needed to adapt to new tasks, environments and machines. But less is not none: putting a foundation model onto your own robot still takes ML infrastructure, compute and complex data pipelines — precisely the part most robotics companies are weakest at. The Pi API applies the same logic as hosted large-model inference: take away the heaviest, least differentiating engineering, and leave three actions — upload data, launch fine-tuning, call inference. The company says the latest model served through the API fine-tunes more data-efficiently and is more capable for its partners, and because the underlying models can be updated quickly, future models will keep arriving through the same door.
Early customers are already running real workloads
To show this is more than a lab demo, Physical Intelligence published progress from early partners. Actor uses it on autonomous heavy machinery: an unmanned wheel loader scoops material and loads a dump truck in a real commercial deployment, with no operator in the cab. Almond's dual-arm robot Axol, after four days of iterating with the API, started deburring spacers used in its own assembly — the goal is for Axol to help build itself. These deployments span Japan, Switzerland and sites across the United States, most of them running cloud-hosted inference, and the company is candid that some sites hit network connectivity problems. What the cases share is not the task but the loop: collect demonstrations, train, test on the robot, and let failures decide what data to gather next — a loop that used to require an in-house team and is now compressed into an API.
Stepping back, the significance is larger than one company adding a product surface. Robot foundation models have mostly existed as papers, demo videos and open weights, which ordinary teams could not necessarily run even when they had the weights. With hosted fine-tuning and hosted inference, the competitive bottleneck moves from having an infrastructure team to having good task data. The undisclosed parts matter just as much: the announcement gives no pricing, latency figures or general-availability date, and whether cloud inference is good enough for latency-sensitive scenes will only be settled by more teams in real conditions. Google, by contrast, embeds its models in its own ecosystem, as with Gemini Robotics ER 2's high-level planning for robots; Physical Intelligence is betting on models as a service. Which route produces paying robotics customers first should become visible within the next year.
