Physical Intelligence opens its robot foundation models to developers

Physical Intelligence, the robotics lab known for its π0 and π0.5 foundation models, has opened a beta API that lets outside developers fine-tune and run those models without building their own training stack. The idea is the same one that reshaped language models: start from a general model that already has broad physical behavior, then adapt it to your robot and task with a modest amount of your own data. Developers upload demonstrations, Physical Intelligence runs the fine-tuning on infrastructure it manages, and the resulting model is served back through a hosted gateway for inference. The company frames the workflow as a loop: collect demonstrations, train, test on the real robot, and use the failures to decide what to collect next.

The early deployments are the interesting part, because they are real jobs rather than lab demos. Actor is using the models to run autonomous heavy machinery, including a wheel loader that scoops material into a dump truck in a commercial site in Utah with no operator inside. Almond, which builds a dual-arm robot called Axol, says its team iterated with the API for four days to get a model that deburrs the spacers used in Axol's own assembly. More than a dozen other companies are listed across agriculture, cooking, and manufacturing, with most running cloud-hosted inference across sites in Japan, Switzerland, and the United States. Physical Intelligence notes that network connectivity was a real constraint for some of them.

For now this is a beta behind a waitlist, with only a small set of companies given early access and no pricing announced. The π0 and π0.5 weights remain available through the openpi repository, so teams that want to run fully local can still do that.

Why it matters

If you work in robotics, this lowers the bar from training a policy yourself to fine-tuning a capable base model on a few days of demonstrations, the same shift that made language models usable for small teams. The open question the deployments expose is practical: whether you can tolerate cloud-hosted inference and its connectivity requirements on a factory floor, or whether you need the local weights instead.