Overview
The Pi API is the robotics foundation model API Physical Intelligence opened on October 13, 2026, currently in beta. It addresses a specific gap: general-purpose robotic foundation models such as pi0 and pi0.5 already cut the data needed to adapt to a new task, but using them still means building your own machine learning infrastructure, compute and data pipelines. The Pi API takes that layer over.
The flow is three steps: upload your data, launch fine-tuning on compute and infrastructure the company manages, then call the fine-tuned model through a hosted gateway to run inference on your robot. Physical Intelligence says the models behind the API update quickly, so teams on the API keep receiving the latest version instead of redoing adaptation themselves.
The design follows a loop the company saw repeatedly across its collaborations: collect demonstrations, train a model, test it on the robot, then use the failures to decide what data to collect next. The Pi API hands that loop to developers. Among the early examples, Almond dual-arm robot Axol deburrs spacers used in its own assembly after four days of iteration with the API, and Actor autonomous wheel loader loads material during a commercial deployment in Utah with no operator inside the machine.
Key Features
- Upload data and fine-tune: No training pipeline to build yourself, fine-tuning runs on compute and infrastructure the company manages
- Hosted inference gateway: Fine-tuned models are called through a cloud gateway, and the robot side consumes the result
- Models keep updating underneath: The company says the models behind the API update quickly, so users keep receiving the newest version
- Broad hardware and industry coverage: Early partners include Actor, Almond, Chef Robotics, Formic Robotics and Frontier Machines, spanning heavy machinery, dual-arm assembly, food service and manufacturing
- A clear iteration loop: Collect demonstrations, train, test, and use failures to decide the next batch of data, repeatable on your own
- Cloud-hosted deployment footprint: The company says early deployments ran in Japan, Switzerland and across the United States
- Application-based beta: Participation requires telling the company what you want to build
Use Cases
- Warehouse and logistics automation where the robots exist but the model capability does not
- Manufacturing and assembly teams that need a model adapted to specific parts
- Food handling and other work that needs compliant, gentle manipulation
- Unmanned operation of construction and heavy machinery
- Research groups running embodied AI experiments who want to skip infrastructure work
Pros
- No need to build machine learning infrastructure or data pipelines, so the team stays on the task itself
- The underlying models keep updating, and users on the API receive the new versions
- Hosted cloud inference lets early teams start before solving deployment
- Early collaborations span several hardware types and industries, with real deployments behind the claims
- The iteration path is explicit, with failures directly guiding the next data collection
Pricing
The Pi API is in beta. Participation requires an application describing what you want to build, pricing has not been published, and commercial terms follow the company official announcements.
Summary
The Pi API is the robotics foundation model API Physical Intelligence opened on October 13, 2026, still in beta. What it solves is not the model but the layer below it: general robotic foundation models already need far less data, yet adapting them still requires infrastructure, compute and data pipelines, and the Pi API takes that over. The flow is upload data, fine-tune on compute the company manages, and run inference on your robot through a hosted gateway, with the underlying models updated continuously.
It fits teams that already have robots and a setting but lack model capability, especially in warehousing, assembly, food service and heavy machinery. Among the early examples, Almond dual-arm robot Axol deburrs parts used in its own assembly after four days of iteration, and Actor wheel loader handles loading in a commercial deployment in Utah.
Two things to keep in mind: it is still a beta that requires an application describing your project, and pricing has not been published, so budgets have to wait for the company official word.
Version History
- Pi API opens in beta: upload data, fine-tune a pi model, run inference on your own robots through a hosted gateway (2026-10-13): On October 13, 2026, Physical Intelligence announced the Pi API beta. General-purpose robotic foundation models such as pi0 and pi0.5 substantially reduce the data and effort needed to adapt to new tasks, environments or robots, but adapting them still requires machine learning infrastructure, compute and complex data pipelines. With the Pi API, developers upload data, launch fine-tuning on compute and infrastructure the company manages, and access the fine-tuned model through a hosted gateway to run inference. The company says the models behind the API update quickly and it plans to keep enabling access to its latest and most capable models. Early examples include the dual-arm robot Axol deburring spacers used in its own assembly after four days of iteration, and an autonomous wheel loader loading material during a commercial deployment in Utah with no operator inside the machine. Most demonstrations ran with cloud-hosted inference across Japan, Switzerland and the United States. Joining the beta requires telling the company what you want to build.