Pi launches beta API for robots, offering training and inference through the cloud

Pi launched the beta version of its Pi API for robots on Oct. 13, 2026. Users can upload data, fine-tune models using computing resources and infrastructure managed by Pi, and run inference through a hosted gateway. The API supports π0 and π0.5, and the company said it aims to simplify adaptation for individual robots.

Pi is proposing to shift AI for robots from custom implementations at individual companies to an operational platform that can be updated and reused through an API. If the company’s description is accurate, developers can run fine-tuning and inference without building their own training infrastructure, creating potential applications for businesses with real-world robot data.

The novelty is not the robot foundation models themselves, but the emphasis on an API layer for adapting them to specific applications. By moving fine-tuning and inference with π0 and π0.5 onto Pi-managed computing infrastructure and a hosted gateway, the company has brought training, deployment and inference together in a single operational workflow. The announcement also frames the challenges of robot development as extending beyond hardware selection to ML infrastructure, computing resources and complex data pipelines. Pi says it will continually update the API with the “latest and highest-performing models,” pointing toward an approach that incorporates model updates and retraining into operations rather than offering a fixed model. A key question is whether data gathered during real-world robot use can be fed back into fine-tuning through the same API. The announcement describes a cycle of “collect demonstrations, train a model, test it on the robot, and use the failures to decide what data to collect next,” opening a loop of real-world data collection, training, testing and further collection to external users. One potential implication for the industry is that competition could shift beyond robot performance to software and cloud infrastructure for data collection and retraining. Pi said its initial deployments spanned Japan, Switzerland and the United States, making adaptation across different environments and hardware a key issue. The company also mentioned network connectivity challenges in some cases, leaving cloud inference reliability, latency and behavior when offline as unresolved questions for real-world operations. Other questions include how much performance gains can be achieved with a given volume of data, which models and applications can deliver generalizability, and how quickly API-based retraining can be repeated.

Pi said the API supports using its latest models for both training and inference. The company manages the fine-tuning compute environment, infrastructure and inference gateway. Pi also said it plans to provide continued access through the API to the “latest and highest-performing models.” The company said its initial collaborations included deployments in locations across the United States, Japan and Switzerland, and that many ran on cloud-hosted inference. It also said there were network connectivity challenges in some cases.

Pi said Japan was among the locations for its initial deployments, making how to incorporate real-world robot data into the training loop a key issue for Japanese robotics companies. Because cloud inference and network connectivity conditions could constrain operations, companies will need to assess deployment requirements in environments such as factories and warehouses where connectivity may be unstable.