Physical Intelligence opens the Pi API beta: upload robot data, fine-tune π models on managed compute, and run hosted inference

Physical Intelligence opened a beta of the Pi API on Oct 13, 2026, turning its robot foundation models into a managed service. General-purpose models like π0 and π0.5 cut the data needed for new tasks, but adapting them still required in-house training infrastructure, data pipelines and inference serving. The API collapses that into three steps: upload demonstration data, fine-tune Pi's latest π model on compute Pi manages, and run the fine-tuned policy on your robot through a hosted inference gateway. Pi frames the value as letting customers run the improvement loop themselves — collect demos, train, test on the robot, and use failures to decide what to collect next — while Pi keeps upgrading the underlying model.

There is no standardized benchmark in the announcement; evidence is partner deployments. Actor runs an operator-less wheel loader in commercial use; early-access customer Almond iterated for four days to get a policy that lets its Axol dual-arm robot deburr its own spacers. Roughly 20 partners and early users are named, with deployments in Japan, Switzerland and the US, mostly on cloud-hosted inference; Pi notes network connectivity was a problem for some.

Access is by application and pricing/API details are unpublished. Teams can prototype today with the Apache-2.0 openpi repo (π0, π0-FAST and π0.5 base checkpoints pre-trained on 10k+ hours of robot data; >8 GB VRAM for inference, >22.5 GB for LoRA fine-tuning).