Physical Intelligence opens a beta API to fine-tune and host its π robotics models

Physical Intelligence (π) announced on October 13, 2026, the beta of its API, offering its latest models for both training and inference. The aim is to spare robot developers from having to build on their own the machine learning infrastructure, compute and complex data pipelines needed to adapt a robotic foundation model to their application.

According to the company, deploying robots is hard: beyond identifying the right use case and building the right hardware, you need models smart enough for the task. General-purpose foundation models such as π0 and π0.5 provide a solid base of physical knowledge and capabilities, which substantially reduces the data and effort needed to adapt them to new tasks, environments or robots. But that adaptation still requires infrastructure and compute that are difficult in their own right.

The API workflow is as follows: the developer uploads their data, launches fine-tuning on compute and infrastructure managed by Pi, and accesses the fine-tuned models through a hosted gateway to run inference on the robot. Evaluation data is used to guide the next iteration. Pi says the latest model available through the API already offers more data-efficient fine-tuning and improved capability for its partners, and that, because it can quickly update the underlying models, it plans to keep providing access to its latest and most capable models.

To test the reach of its models, the company has worked with companies that deploy robots. Examples cited include Actor, which deploys autonomous heavy machinery: its autonomous wheel loader scoops material and loads it into a dump truck during a real commercial deployment, with no operator inside the machine (footage recorded in Utah). AgRobotics, CloudChef, a dairy-sector partner and Ultra Weave also appear, among others. Almond, which builds Axol, a dual-arm robot for physical AI builders, shows Axol deburring spacers used in its own assembly, with a model obtained after four days of iteration with Pi's API; Almond's goal is for Axol to eventually build itself.

Pi points to a common pattern across these collaborations: the problems vary from task to task (a garment turned inside out, an ingredient that wouldn't stay on the spoon, a manufacturing part that ended up in the wrong position), but the improvement process was similar: collect demonstrations, train a model, test it on the robot and use the failures to decide what data to collect next. The API is designed to let developers run that loop independently with Pi's latest models. A small number of companies got early access to see whether they could improve autonomy on their own, and the company says they got started quickly and made progress.

The list of early-access participants includes, in addition to those mentioned, Argon Robotics, Chef Robotics, Deft, Dream Machines, Formic Robotics, Foundry, Frontier Machines, General Instinct, Hackathon, Intuition.dev, Omakase, robot.com and Yondu. Most of the applications shown, from Japan to Switzerland and various deployments in the United States, ran on cloud-hosted inference through the API. Pi acknowledges that these early deployments revealed areas for improvement, including network connectivity issues in some cases.

Pi also presents the API as a way to learn: the more real-world uses its models have, the better it will understand their limits and be able to improve future generations. It argues that applications such as robotic construction, robotic surgery or warehouse automation each require an entire company, and that if each had to solve the problem of physical intelligence from scratch, it would be unmanageable. By offering its models via API, it says it is accelerating the robotics ecosystem and, along the way, advancing its own understanding of physical intelligence.

The company admits it is still early days and that developer feedback will help shape future versions. Anyone wishing to join the beta must describe what they want to build, and Pi says it is hiring in research, infrastructure, hardware, people and operations. It is worth noting that the claims about efficiency and capability improvements come from the company itself, and the text provides no comparative figures, pricing or access terms.