Observed facts: Physical Intelligence says developers can upload data, fine-tune π models on company-managed compute, and run inference through a hosted gateway. The company describes an iteration loop of collecting demonstrations, training, testing on a robot, and using evaluation to choose the next data to collect. It also reports that Almond iterated for four days to obtain a model used by its Axol robot to deburr spacers.
Mechanism: The API bundles compute, data-driven fine-tuning, and hosted inference so a robotics team can test a task-specific policy without assembling all of that ML infrastructure itself. The loop ties failures on the physical system back to the next training dataset; the robot test is part of development, not merely the final deployment check.
Why now: Physical Intelligence says the beta is intended to let developers adapt its latest models to applications and hardware, while partner deployments provide feedback for improving the models. The announcement makes the adaptation path available to a wider developer cohort, though participation remains a beta.
Implications: Teams may be able to spend less effort building fine-tuning and serving infrastructure and more on demonstrations, robot tests, and task-specific evaluation. The central operational question is whether this loop produces repeatable gains across different robots and environments, rather than only promising early examples.
Evidence boundary: These facts and results are reported by Physical Intelligence in its launch post. The source does not provide independent benchmarks, controlled comparisons, or enough detail to establish general performance; it notes network-connectivity challenges for some cloud-hosted deployments.
Watch: Watch for beta availability and terms, reproducible results across partner tasks and hardware, and whether developers can export or otherwise control fine-tuned models and deployment behavior.
New branch: No prior New Runtime record was supplied for this specific API beta; this is a distinct record about a hosted adaptation workflow for robotic models.