Platforms
Physical Intelligence Opens Pi API Beta, Moving Robot Policy Development from Months to Days
Physical Intelligence has made its robotic foundation models available to outside teams through a hosted fine-tuning service, compressing policy development cycles from months to days for early adopters in manufacturing, agriculture, and food service.
Mainstay Communications Team · · 3 min read

In brief
Physical Intelligence launched the Pi API beta on October 13, 2026, giving external teams access to fine-tuning infrastructure for its π0 and π0.5 vision-language-action models without managing their own GPU clusters. The workflow mirrors the managed-model pattern that normalized large language model deployment: upload demonstration data, fine-tune on Physical Intelligence's compute, and serve inference through a hosted gateway. Early design partners across manufacturing, agriculture, and food service have logged working deployments, with one team achieving a functional deburring policy in four days of iteration.
The Announcement
Physical Intelligence opened the Pi API beta on October 13, 2026, making its robotic foundation models available for hosted fine-tuning and inference by outside teams. Access is gated by a waitlist; applicants describe their intended use case before being approved.
The development workflow runs four steps: record teleoperated robot demonstrations, upload the dataset, launch fine-tuning on compute Physical Intelligence manages, and connect robots to a hosted inference gateway. Teams then use observed failures to guide the next round of data collection.
The Foundation Models
The Pi API supports three models: π0, a flow-based vision-language-action model; π0-FAST, an autoregressive variant of the same architecture; and π0.5, which extends π0 with improved capability for open-world generalization. All three were pre-trained on more than 10,000 hours of robot demonstration data.
Physical Intelligence has released open-source versions under an Apache-2.0 license via its openpi repository. The Pi API handles the infrastructure overhead that most robotics teams lack: fine-tuning a vision-language-action model from scratch requires A100 or H100 GPUs with at least 70 GB of memory for a full training run, plus engineering capacity to manage pipelines most hardware teams don't staff for.
Early Partners
Design partners span manufacturing, agriculture, and food service. Almond's Axol dual-arm robot achieved a functional deburring policy for spacers used in its own assembly after four days of iterating with the Pi API. Actor, which develops autonomous heavy machinery, deployed a wheel loader model trained through the API to commercial sites where it scoops material and loads dump trucks without an operator.
Agriculture is represented by AgRobotics. In food service, CloudChef, Omakase, and Chef Robotics have applied it to ingredient handling and food preparation. Formic, Foundry, Deft, Ultra, Weave, and Dream Machines are among the manufacturing and logistics teams in the program. Active deployments span Japan, Switzerland, and the United States.
The Industry Read
Language model APIs established that most product teams shouldn't run their own model infrastructure. Physical Intelligence is applying the same premise to robot control.
The harder problem is that robotic policies handle physical state over time, with tighter tolerances than text generation. A language model that produces a wrong word can be prompted again; a robot that misjudges a part position causes downtime or damage. The iteration cycle the Pi API enables, where teams can close the loop from demonstration data to deployed policy in days, is what was previously measured in months of GPU time and engineering overhead.
Physical Intelligence was founded in 2024 by seven co-founders including Karol Hausman, formerly of Google DeepMind. The company raised $600 million at a $5.6 billion valuation in November 2025. Pricing for the Pi API beta hasn't been disclosed.
Sources
- Frontier Models for Every Robot: Announcing our API beta, Physical Intelligence
- Physical Intelligence Opens Pi API Beta for Robot Foundation Models, Unite.AI
- Physical Intelligence Opens Pi API Beta for π0, π0.5 Fine-Tuning, AI Weekly
- Physical Intelligence Opens Its Robot Brain API to Outside Teams, Alpha Signal
- Physical Intelligence Inc., Wikipedia
- Axol: A Dual-Arm Robot Built for AI, Y Combinator Launch YC
Companies mentioned
Mainstay Digital · Spatial and 3D visualization for industrial companies

