Physical Intelligence launches robotics model API beta

Physical Intelligence says developers can upload data and launch fine-tuning using compute and infrastructure the company manages.[1] After fine-tuning, developers can run inference through a hosted gateway.[1] The company identifies π0 and π0.5 as examples of its general-purpose robotic foundation models.[1] It says these models provide physical knowledge and capabilities intended to reduce the data and effort needed to adapt robots to new tasks and environments.[1] However, the company also says adapting foundation models still requires ML infrastructure, compute and complex data pipelines.[1]

The beta is starting with limited access: Physical Intelligence says it will onboard developers in batches and asks applicants to describe what they are building.[2] The company says it plans to continue providing access to its latest and most capable models as it updates the models behind the API.[1] Physical Intelligence also says the latest model available through the API already offers its partners more data-efficient fine-tuning and improved capability.[1] The evidence gives no price or broader availability date, so teams considering a pilot will need to confirm those details with the company.

Physical Intelligence says a small number of companies received early access to test whether they could improve autonomy independently, and that those companies got started and began making progress quickly.[1] One example is Actor, which deploys autonomous heavy machinery.[1] The company says Actor’s autonomous wheel loader scoops material and loads a dump truck in a commercial deployment without an operator inside.[1] Another example is Almond, builder of the dual-arm Axol robot, which aims to build itself.[1] Physical Intelligence says Axol deburred spacers used in its own assembly after Almond’s team iterated with the API for four days.[1]

Physical Intelligence says its partner collaborations test model performance across different tasks, hardware and operating environments.[1] It describes a workflow of collecting demonstrations, training a model, testing it on a robot, and using failures to decide what data to collect next.[1] Most applications shown ran inference in the cloud through the API, while the company highlighted network-connectivity challenges for some.[1] Physical Intelligence says developer feedback will help shape future releases.[1] It describes the beta as a first step toward a foundation the community can build on and end-to-end autonomy.[1]