Physical Intelligence is opening a beta service for teams that want to adapt its robot models to their own machines and tasks. The Pi API announcement, published October 13, says a team can upload robot data, fine-tune a model on infrastructure Pi manages, and call the resulting model through hosted inference. Access begins with an application, and Pi says it is onboarding in batches.

That changes where some model engineering happens. A robot developer still needs a suitable machine, a task worth automating, demonstrations of that task, and a way to test the model on the machine. Pi proposes to run the training and model-serving layer that connects those pieces.

The big change

  • What changed: Pi is offering hosted adaptation of its robot models to selected teams. Its public openpi repository already supplies older model checkpoints and examples for teams prepared to run their own training stack; the beta adds a managed route for fine-tuning and inference.
  • Why it matters: Moving training infrastructure to Pi could let a robot team spend more effort collecting useful demonstrations and evaluating behavior on its own equipment. The service does not supply the robot, define a successful task, or establish safe operation for the customer.
  • What to watch: The beta’s value will depend on repeatable task results and workable integration terms. Pi has described partner examples, while the public launch material leaves the model version, supported hardware, pricing, API contract and safety terms unspecified.

The work that moves to Pi

Pi describes a sequence: upload data, launch fine-tuning on its managed compute, then use a hosted gateway for inference. In its workflow description, evaluation data guides the next iteration. A developer collects demonstrations, trains, tests on a robot and studies failures to decide what data to collect next. The API is intended to let the developer run that loop independently while Pi operates the model infrastructure.

The company's openpi codebase shows the earlier do-it-yourself path. It includes π₀, π₀-FAST and π₀.₅ checkpoints, plus examples for converting a dataset, configuring fine-tuning and serving a policy. Its stated GPU requirements and data-preparation steps make clear why managed training might appeal to a team with robotics expertise but limited machine-learning infrastructure. Pi has not identified the beta model by version in the announcement, so the open-source models should not be mistaken for the specific model served by the API.

Pi says its latest API model has improved capability and more data-efficient fine-tuning for partners. The launch post does not publish a comparative dataset, benchmark protocol or task success rates to quantify that claim. It does say that most featured applications used cloud-hosted inference and that some encountered network connectivity challenges. That detail matters for a moving machine whose control loop depends on a remote service, though the post does not specify latency or reliability requirements.

What a team must bring

The beta application asks who on the team would train the policy, what they have done with robot data and policy evaluation, what precise task they want to accomplish, and how much teleoperation or simulation data they have. It also asks for the number and type of robots, a preferred inference location, and task success criteria. These are application questions, not published eligibility thresholds or guarantees that local inference is available. The form offers cloud, local, both and uncertain as choices; the launch describes hosted inference.

Those questions point to the part the service cannot automate for a customer: deciding what correct behavior means on a particular robot and measuring it on that robot. Pi describes early partners cycling through demonstrations, tests and failures. Its examples include Almond's Axol, a dual-arm robot shown deburring spacers used in its assembly after four days of API iteration. Pi also describes an Actor wheel loader scooping and loading material during a commercial deployment. These are Pi-selected accounts and clips, not independent measures of general reliability, safety or performance across hardware.

The current public route is to apply for beta access. Pi has not published an open API contract, request and response schema, named beta model, hardware compatibility matrix, pricing, data-handling terms or safety agreement in the materials reviewed here. A team evaluating the service would need those terms and its own repeated robot tests before it could judge whether hosted inference fits a real operation.

Related coverage

BIG CHANGE previously examined Runway's video-trained robot policy and Japan's humanoid data factory. They address related model and data work. Pi's October 13 announcement is a separate service-access event.

Sources & further reading