Robotics Model Firm Physical Intelligence Launches Pi API Beta for Cloud Fine-Tuning and Inference

Robotics model company Physical Intelligence (Pi) announced the launch of the Pi API beta on October 14, offering external developers access to model fine-tuning and cloud-hosted inference services. This move aims to lower the hardware barriers for robotics AI development, enabling teams to train adapted models without building their own high-performance computing clusters.

Previously, developers had to download open-source models and configure training pipelines locally, with full fine-tuning typically requiring GPUs with over 70GB of VRAM. The Pi API hosts this infrastructure in the cloud; developers upload task data to initiate fine-tuning on Pi-managed compute and then call the model via API to drive robot execution.

Pi disclosed several demo cases, including Almond team iterating a dual-arm robot's deburring capability in four days, and Argon Robotics completing a kitchenware organization model tuning in 35 hours. However, these results are vendor-reported demonstrations without publicly available details on independent benchmarking conditions.

Notably, cloud inference introduces new time overheads. Action updates involve image processing, uploading, model computation, and result return, meaning on-site network conditions directly impact robot response speed. Pi acknowledged that some projects encountered network connectivity issues. Access to the API currently requires application approval and is being rolled out in small batches.