Image: website.pi-asset.com
Beta API opening access to Physical Intelligence's models so more teams can adapt them to their robots and applications.
Key specs
| Specifications | Not stated |
|---|
Technical details
The Pi API beta lets teams upload their own robot data, launch fine-tuning of Physical Intelligence's latest robotic foundation models on PI-managed infrastructure, and run the resulting policies through a hosted inference gateway. It targets the standard iteration loop of collect demonstrations, train, test on robot, and use failures to guide the next data collection. Most early-access deployments ran with cloud-hosted inference. Access is via an application/waitlist; endpoints, pricing and model versions were not published in the announcement.
| Architecture | Managed cloud service exposing Physical Intelligence's latest π-series robot foundation models for fine-tuning and hosted inference (specific model version behind the API not disclosed). |
|---|---|
| Inputs outputs | Input: customer-uploaded robot demonstration data; output: fine-tuned π models served through a hosted inference gateway that the robot queries for actions. |
| Embodiments supported | Customer robots across industries; early-access examples include an autonomous wheel loader (Actor), Almond's dual-arm Axol and other partner platforms. |
| License weights | Proprietary; weights not released (API access only). |
| Hardware requirements | Fine-tuning runs on PI-managed compute; inference is cloud-hosted via the API gateway (PI notes network connectivity was a challenge for some deployments). |
| How to try | pi.website/…/api-beta |
| Sources |
