Physical Intelligence opens robot-model training and hosted inference through an API beta

Robot developers can now apply to adapt Physical Intelligence’s models without managing the training infrastructure themselves. On October 13, 2026, the company announced its Pi API beta: developers upload robot data, train a customized model on company-managed computing, and access it through a hosted service to run on their robots.

Managed training, developer-led iteration

Physical Intelligence frames the service around a gap between having a capable robot model and making it useful in a particular setting. General-purpose models such as π0 and π0.5 provide a starting point, the company says, but adapting them still requires machine-learning infrastructure, computing resources, and complex systems for preparing data.

The API—a software interface developers use to access the service—handles the computing side of that adaptation. Fine-tuning means further training an existing model on the developer’s data. After training, a hosted gateway provides inference: running the tuned model to produce outputs used by the robot. Physical Intelligence manages the underlying training infrastructure.

Developers still lead the cycle of teaching and testing. Early-access companies collected demonstrations, trained a model, tried it on their robots, and used failures to choose what data to gather next. The service was built to let those teams repeat that process independently, rather than relying on Physical Intelligence to conduct every round of improvement.

A robot working on its own assembly parts

Almond offers one concrete example of that loop. The company builds Axol, a dual-arm robot for physical AI developers, and wants it eventually to build itself. Physical Intelligence says Almond obtained a model after four days of iteration with the API that let Axol deburr spacers used in its own assembly.

The launch also describes work with Actor, which deploys autonomous heavy machinery. In a commercial deployment, its wheel loader scooped material and loaded a dump truck with no operator inside the machine, according to Physical Intelligence. That example sits within the company’s broader collaborations to assess how its models adapt across hardware, tasks, and operating environments.

The practical failures varied across collaborations: an inside-out garment, an ingredient that would not stay in a scoop, or a manufacturing part in the wrong position. Physical Intelligence says the improvement process was nevertheless similar. Those examples put the emphasis on collecting useful demonstrations for a specific problem, not simply connecting a robot to a general model.

Cloud access has a field constraint

Most applications shown in the announcement used cloud-hosted inference, spanning Japan, Switzerland, and deployments across the United States. Physical Intelligence acknowledged network-connectivity challenges in some deployments. The examples therefore show both the reach of the hosted approach and an operational problem encountered when using it outside a controlled training environment.

The company says the latest model available through the API offers more data-efficient fine-tuning and improved capabilities for partners. It plans to keep providing access to newer models as it updates the service. Those are company claims and plans, rather than a quantified comparison of training requirements or a guaranteed schedule for model releases.

Physical Intelligence also wants those deployments to inform its own model development. It says observing real applications helps reveal current models’ strengths and limits and guide the next generation. The beta follows access for a small group of companies; developers interested in participating are invited through the announcement to describe what they want to build.