Record, train, and deploy from one place with Strands Agents, LeRobot, and Hugging Face Storage Buckets

The post is part of a series on Strands Robots, an open-source SDK from AWS (Apache 2.0) that exposes robot abstractions, simulation, and the LeRobot stack as AgentTools. It addresses the challenge of running a continuous recording-training-deployment loop: daily runs incur repeated byte transfers as the dataset grows and new checkpoints ship out, wasting bandwidth and compute.

The proposed solution introduces Hugging Face Storage Buckets, a mutable, non-versioned, Xet-backed object-storage repository type announced in March 2026. Buckets sit alongside normal dataset repositories in the same hf:// namespace and are managed with the existing hf CLI. The workflow follows four steps: record demonstrations into a bucket, store with byte-level deduplication, train by streaming from the Hub (avoiding full dataset copies), and deploy the policy back to hardware while returning new data to the loop. LeRobot's dataset format, already used by over 90,000 datasets and models from more than 8,000 publishers, ensures Strands recordings are natively compatible without conversion.

The loop is designed to support human decisions like which episodes to keep, when scene drift requires re-recording, whether today's batch is large enough for training, and which checkpoint replaces the one on the arm. The post includes a sample application, security considerations, cleanup steps, and links to resources.