Physical Intelligence opens its π0 robot models to everyone via an API

In short: On Oct 13 Physical Intelligence (Pi) launched the Pi API beta. Developers upload robot data, launch fine-tuning of Pi's foundation models (such as π0 and π0.5) on compute Pi manages, and run the fine-tuned model through a hosted inference gateway.

What Pi showed

• Partner deployments span heavy machinery (Actor's autonomous wheel loader in Utah), agriculture, kitchens, dairy and garment handling.
• Early-access users iterated on their own: Almond's dual-arm Axol robot deburrs spacers used in its own assembly, using a model obtained after four days of iterating with the API.
• Most showcased applications, from Japan to Switzerland to the U.S., ran with cloud-hosted inference; Pi notes network connectivity was a challenge for some.
• Access is by application; pricing has not been published.

Why it matters

Robot foundation models like VLA models have been hard to adapt without ML infrastructure. An API turns the loop — collect demonstrations (often via teleoperation), fine-tune, test, repeat — into a service, the way LLM APIs did for software. It also gives Pi data on which real tasks its models fail at.