AI news · October 15, 2026
Physical Intelligence opens a beta API for robot models
Robot learning moves closer to an ordinary model workflow: upload demonstrations, fine-tune in managed infrastructure and evaluate the policy through an API.
- Axol robot example
- 4 days
- availability
- Beta
Physical Intelligence announced its Pi API beta on October 13. The service is designed to let developers upload robot data, launch fine-tuning on managed compute and access the fine-tuned model through hosted inference. The company describes the API as a way to work across tasks and hardware, rather than building a separate deployment path for each robot. Its example describes an Axol robot learning a deburring task after several days of iteration.
The shift is operational rather than cosmetic. A team can treat demonstrations, fine-tuning and inference as one experiment loop, then compare a policy on a real task. The beta is early, so hardware compatibility, data transfer, latency and safety remain open questions. Use a small repeatable task, not a broad robot programme, for the first evaluation.
What you can do with it
Try it: request access to the Pi API beta with one robot task and a small demonstration set. Define success before uploading data: completion rate, collision-free runs, operator interventions, inference latency and cost per successful episode.
Our take
Hosted robot learning could reduce setup cost for teams that lack a full training stack. It will not remove the hard part: reliable data and safe physical evaluation. The best early test is a narrow task with clear pass and fail conditions.
Links Try it
Source: Physical Intelligence — Made With Models writes the brief; the reporting is theirs.