Physical Intelligence announced the Pi API beta on October 13, 2026, for managed robot-model fine-tuning and hosted inference. The company says developers upload robot data, fine-tune models on its infrastructure and access adapted models through a hosted inference gateway. It describes a development process of collecting demonstrations, training, testing on robots and collecting more data informed by failures. According to the company, most showcased applications used cloud inference across Japan, Switzerland and the United States, and some encountered connectivity problems. Access is through a waitlist.
NewsEditorialSource: Physical Intelligence
Physical Intelligence announces Pi API beta for robot fine-tuning and inference
Why it matters
The service could reduce infrastructure work for robotics developers, while leaving hardware integration and evaluation with each team. The reviewed material does not establish pricing, supported robot interfaces, data rights or inference latency; independent reporting was available only through search results. Interested teams should apply through the official homepage, request those details and test behavior during network interruptions before planning deployments.
Attribution
Sources
- 1Physical IntelligencePhysical Intelligence (π)
- 2Physical IntelligenceFrontier Models for Every Robot: Announcing our API betaSource published (UTC)
- 3Black ScarabPhysical Intelligence opens Pi API beta for robot training and inferenceSource published (UTC)
Read the original source for full details and context.