Physical Intelligence Launches API Beta for Fine-Tuning Robotic Foundation Models

Physical Intelligence has announced the beta launch of the Pi API, a service that provides access to its latest robotic foundation models, π0 and π0.5, for training and inference. The API aims to lower the barrier to deploying autonomous robots by handling the complex ML infrastructure, compute, and data pipelines required to adapt general-purpose models to specific tasks and hardware.

How the API Works

The workflow is designed to simplify the iteration loop for developers seeking practical autonomy:

• Upload Data: Developers submit demonstration data relevant to their specific use case.
• Fine-Tune: The API launches fine-tuning on Physical Intelligence's managed compute infrastructure.
• Hosted Inference: Developers access the fine-tuned models via a hosted gateway to run inference on their robots.

This approach allows partners to rapidly iterate on policies using the latest models without building their own training infrastructure. The company states that the latest model available through the API offers more data-efficient fine-tuning and improved capabilities compared to previous versions.

Early Deployments and Partners

Physical Intelligence collaborated with a diverse set of companies to validate the API's versatility across industries and hardware types. Early partners include:

• Actor: Deployed autonomous heavy machinery, including a wheel loader operating without an operator in Utah.
• Almond: Used the API to enable their dual-arm robot, Axol, to deburr spacers for its own assembly after four days of iteration.
• Other Partners: AgRobotics, CloudChef, Deft, Dream Machines, Formic, General Instinct, Intuition.dev, Omakase, robot.com, Ultra, and Yondu.

These deployments spanned environments from Japan to Switzerland and across the United States, demonstrating the models' ability to handle tasks ranging from agriculture and manufacturing to food preparation. Most applications utilized cloud-hosted inference, though some faced network connectivity challenges.

Strategic Goal

The launch is part of Physical Intelligence's strategy to accelerate the robotics ecosystem by providing a shared foundation of physical knowledge. By making models accessible via API, the company hopes to gather feedback from real-world applications to improve future model generations and advance toward general physical intelligence.