Physical Intelligence has opened its latest robotic foundation models for training and inference through the Pi API beta. Deploying robots remains difficult, since developers must identify the right use case, build suitable hardware, and supply models intelligent enough to perform the task. General-purpose models such as π0 and π0.5 provide a strong foundation of physical knowledge, substantially reducing the data and effort needed to adapt to new tasks, environments, or robots. Adapting them, however, still requires ML infrastructure, compute, and complex data pipelines.
The Pi API removes much of that burden. Users upload data, launch fine-tuning on infrastructure Physical Intelligence manages, and access fine-tuned models through a hosted gateway for inference. The latest model available through the API offers more data-efficient fine-tuning and improved capability for partners, and because Pi can update the underlying models rapidly, the company plans to keep providing access to its newest and most capable versions.
Early collaborations tested the models across industries. Actor deploys autonomous heavy machinery, and its autonomous wheel loader scoops material and loads it into a dump truck during a real commercial deployment without an operator inside the machine. Across these partnerships, the process for improving performance was consistent: collect demonstrations, train a model, test it on the robot, and use failures to guide the next round of data collection. Almond, which builds the dual-arm Axol robot, used the API to produce a model in four days that deburrs spacers used in Axol's own assembly.
Most early deployments, from Japan to Switzerland and across the United States, ran on cloud-hosted inference. Network connectivity challenges surfaced for some users, highlighting areas for improvement. Physical Intelligence says that feedback from real-world use will guide future model development.
Source: Hacker News · Summarized by HeadlinesBriefing