Physical Intelligence Opens Its Robot Brain API to Outside Teams
Physical Intelligence opens its robotic foundation models to outside teams, letting companies upload data, fine-tune, and run cloud-hosted inference on their own hardware.
- Physical Intelligence launched the Pi API beta for fine-tuning and hosting robotic foundation models
- Workflow: upload demonstrations, fine-tune a π model, run cloud-hosted inference on your robot
- Pi manages compute and infrastructure; latest base model offers more data-efficient fine-tuning
- Design partners include Actor (wheel loaders), Almond (dual-arm assembly), Chef Robotics, AgRobotics, Formic
- Almond trained a working deburring policy after four days of iterating with the API
- Beta is waitlisted at pi.website/waitlist; pricing not yet disclosed
Physical Intelligence puts robot policies behind an API
Physical Intelligence, or Pi, has opened the Pi API beta, giving outside teams access to the robot foundation models used in its partnerships. Teams can upload robot data, fine-tune a model on Pi's managed infrastructure, and serve the resulting policy through a hosted inference endpoint. The service applies the managed-model workflow common in language AI to robot control.
Teams adapting the open-source π0 and π0.5 models have typically managed their own training compute, data pipelines, deployment infrastructure, and model serving. A robot policy maps sensor observations and instructions to actions, such as moving an arm or controlling a gripper. Pi's foundation models supply reusable physical knowledge that can reduce the data and training needed for a specific robot, task, or environment.
From demonstrations to actions
Teams begin by teleoperating a robot and recording successful task demonstrations. Pi's infrastructure then fine-tunes a base model on that dataset, while the hosted gateway provides action predictions during deployment. The development loop has four steps:
- Collect teleoperation demonstrations on the target robot.
- Upload the dataset and fine-tune a Pi model.
- Connect the robot to the hosted inference gateway.
- Record failures, collect targeted demonstrations, and retrain.
Observed failures determine which data to collect next. Pi cites garments turned inside out, ingredients falling from a scoop, and manufacturing parts landing in the wrong position. These physical edge cases expose gaps that aggregate benchmark scores may conceal.
Pi says its latest API model improves fine-tuning efficiency and partner performance. The company also plans to release newer base models through the same endpoint, making model versioning and compatibility important considerations for teams validating policies on physical hardware.
One model, many machines
The design-partner roster spans construction, manufacturing, agriculture, food service, and logistics, providing an early test of whether one base model can adapt to different robots and control problems.
| Partner | Field | Deployment |
|---|---|---|
| Actor | Material handling | An autonomous wheel loader scoops material and loads a dump truck during a commercial deployment without an operator inside the machine. |
| Almond | Manufacturing | Its dual-arm Axol robot deburrs spacers used in assembly. Almond produced the model after four days of iteration with the API. |
| Chef Robotics, CloudChef, and Omakase | Food service | Robots perform food preparation and plating tasks. |
| AgRobotics | Agriculture | Robots harvest strawberries. |
| Formic, Foundry, Deft, Ultra, Weave, and Dream Machines | Manufacturing and logistics | Pi names these companies as design partners but has not disclosed comparable deployment details for each one. |
Most partner applications used cloud-hosted inference, according to Pi, across projects in Japan, Switzerland, and the United States. Those deployments also exposed connectivity problems, which directly affect a service that sends robot observations to a remote model and receives action predictions in return.
Cloud serving introduces round-trip latency, network dropouts, and recovery requirements into the control path. Teams evaluating the API will need to measure latency under production conditions, keep safety controls local, and define how each robot behaves when the endpoint becomes slow or unavailable.
Pi takes over the model layer
Robotics products often combine bespoke perception systems, policy training pipelines, deployment tools, and inference servers. Rebuilding that stack for each application consumes engineering time before a robot performs useful work. Pi centralizes fine-tuning and model serving around a shared base model, allowing task-specific training to focus on demonstrations from the target environment.
A small team that can teleoperate its robot and record compatible episodes may be able to test a vision-language-action model without operating its own training cluster. The team remains responsible for robot integration, dataset quality, evaluation, local safety systems, failure recovery, and the mechanical limits of its hardware.
Production questions remain open
Access currently requires joining the beta waitlist, and Pi has not published pricing. The beta details released so far leave several implementation questions for developers and researchers:
| Area | Details teams need |
|---|---|
| Data contract | Accepted sensor modalities, episode schemas, annotation requirements, dataset limits, and validation tools. |
| Robot compatibility | Supported action spaces, control frequencies, client runtimes, and hardware requirements. |
| Inference | Expected latency, regional availability, rate limits, uptime commitments, and options for degraded or offline operation. |
| Model lifecycle | Version pinning, rollback support, compatibility of existing fine-tunes, and migration behavior when Pi releases a new base model. |
| Data governance | Retention policies, customer data isolation, ownership of uploaded demonstrations, and use of customer data for future models. |
| Pricing | Charges for fine-tuning, storage, hosted inference, and network usage. |
Pi says feedback from the beta will shape future base models. The current deployments cover a broad range of machines and tasks, while production adoption will depend on predictable latency, transparent costs, stable model versions, and clear data policies. If the service meets those requirements, robotics teams can direct more engineering effort toward demonstrations, evaluation, and edge-case recovery while Pi operates the training and serving infrastructure.
Comments
