Physical Intelligence Opens API Beta for Robot Model Training and Inference

Physical Intelligence has announced a beta API for training and running its robot models, taking another step toward supplying the intelligence behind other companies’ machines.

Key Takeaways

Show all

• Developers can fine-tune models on PI-managed infrastructure and use hosted inference.
• Access is application-based, with teams onboarded in batches.
• Robot data collection and evaluation remain part of the development process.

Announced on October 13, the service lets teams upload demonstrations, train a policy and test it on hardware, then collect further data around failures. PI says most showcased applications used cloud inference, with network connectivity proving challenging for some.

One example is Almond’s Axol dual-arm robot deburring spacers for its own assembly. PI says the team reached that demonstration after four days of API iteration. Its broader collaborations include Actor, whose wheel loader loads dump trucks in a commercial deployment.

A managed route to robot models

PI already distributes models and training tools through openpi, including π0, π0-FAST and π0.5. That repository provides checkpoints and examples for adapting models to developers’ datasets, but users must configure data processing, training and policy serving themselves.

The API’s managed approach could make experimentation more accessible to teams whose expertise lies in hardware or a particular application. That is the practical opportunity: less infrastructure to maintain while investigating whether a learned policy can reliably perform a useful job.

The launch builds on our earlier coverage of PI’s work with Weave and Ultra, which examined its models in laundry folding and industrial packaging. Those partnerships illustrated the business proposition of supplying a shared model foundation across different robot makers.

Applications are open

The beta application form says PI is starting small. It asks about applicants’ robot setups, training experience, available teleoperation data, intended simulation data and task success criteria. It also asks whether inference needs to run in the cloud, locally or both; that question alone does not establish local deployment availability.

For prospective users, the next step is applying with a concrete task and hardware setup. The rollout offers a way to test the shared-model approach against the requirements of an actual deployment.