Physical Intelligence opens beta of Pi robot model API

On October 14, 2026, the robotics model company Physical Intelligence (Pi) announced the launch of the Pi API beta, opening model training and inference services to outside developers. According to its technical blog, the API centers on two capabilities: fine-tuning of robot models and cloud-hosted inference. Developers can upload their own task data, train adapted models on compute and infrastructure hosted by Pi, and then call the model through the API to drive a robot through tasks. They can also add data based on real-world test results and keep improving the model, which reduces the work of building their own training and inference systems.

In a case disclosed by Pi, the Almond team used the API over a 4-day iteration cycle to get the two-armed robot Axol to deburr the spacers needed for its own assembly, taking part in machining its own parts. In the video, Axol picks up a part, places it on a work surface resembling sandpaper and rubs it, flipping the part over on its own.

The report notes that this gives robotics companies a hosted-service path alongside open models: developers can push their applications forward around their own hardware and data while handing part of the infrastructure needed to train and run models to Pi. The service is currently in beta.