Physical Intelligence Opens Robot Model API: Strawberry Picking, Server Assembly, New Actions Trainable in Just 4 Days

Beating AI News Flash: U.S. robotics AI company Physical Intelligence (Pi) has launched a beta version of its Pi API, beginning to open its robot foundation model to more enterprises. Developers only need to upload robot operation data to fine-tune the model on cloud computing power provided by Pi, and then control their robots through the API. There is no need to train a model from scratch or build complex training and inference infrastructure on their own.

Pi demonstrated real-world applications with multiple partners, including AgRobotics' strawberry-picking robot, an autonomous loader for automated truck loading, and robots for food processing and server assembly. Among them, startup Almond used the Pi API and, after only 4 days of repeated training and testing, enabled a dual-arm robot to learn to grind the parts needed for its own assembly.

Pi had previously open-sourced robot models such as π0 and π0.5, but developers still needed to deploy and train them on their own. This time, it has further packaged model fine-tuning, computing power, and inference services into an API, attempting to become a shared intelligent foundation for different robots. However, beta access still requires an application, and pricing has not yet been announced.