Physical Intelligence API Beta: Frontier Models for Every Robot

Physical Intelligence (Pi) has launched a beta API that exposes its robotic foundation models (π0 and π0.5) for fine-tuning and hosted inference. Developers upload task-specific data to Pi, trigger managed training on Pi’s infrastructure, then deploy models through a hosted gateway to run on robots. The workflow - collect demonstrations, fine-tune, test on hardware, and iterate - reduces the data and engineering burden traditionally required to adapt general robotic models to new tasks, environments, or hardware. Pi emphasizes rapid model updates and improved data efficiency in fine-tuning so teams can iterate without building complex ML pipelines or compute stacks.

Early partners across industries validated that the models adapt to diverse applications, from an autonomous wheel loader that loads dump trucks to a dual-arm robot (Axol) that deburrs parts after four days of iteration. Most deployments used cloud-hosted inference; this worked well but revealed practical limits such as network connectivity issues in some settings. Pi reports that real-world usage informs model improvements and accelerates ecosystem growth, and it invites teams to join the beta and applicants to apply for roles across research, infrastructure, hardware, and operations.