Physical Intelligence announces pi API beta for robotic foundation models

The API currently exposes three models: π0, a flow‑based vision‑language‑action model; π0‑FAST, an autoregressive model that uses the FAST action tokenizer; and π0.5, an upgraded version with better open‑world generalization trained with knowledge insulation. Base checkpoints are pre‑trained on more than 10,000 hours of robot data and are released under an Apache‑2.0 license. Running the open models locally requires an NVIDIA GPU with more than 8 GB of memory for inference and more than 70 GB for full fine‑tuning.

Physical Intelligence highlighted early deployments with partners such as Actor, AgRobotics, CloudChef, Ultra, and Weave. The company also gave early access to the API to a small group of firms, including Almond’s Axol robot, which deburred spacers after four days of iteration. The company said the examples show the models can be used for real applications, but the effort remains early and developers’ feedback will shape future releases.