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

According to Hugging Face, Strands Robots, LeRobot, and Hugging Face Storage Buckets combine to form a streaming data loop that handles robot demonstration recording, model training, and policy deployment from a single interface. Strands Robots, an open-source Apache 2.0 SDK from AWS, allows developers to compose robot abstractions, simulation features, and the LeRobot stack into a unified Strands agent. The workflow utilizes Hugging Face Storage Buckets, a mutable, non-versioned, Xet-backed object storage repository type announced in March 2026. These buckets reside in the hf namespace alongside dataset repositories and operate via the hf CLI, acting as an interim working layer between data collection and training. Because Storage Buckets are backed by Xet, they use content-defined chunking to deduplicate uploads at the byte level, avoiding the overhead of re-uploading multi-gigabyte video shards during continuous collection campaigns. Agents can record a LeRobotDataset, sync it directly into a Storage Bucket, stream the dataset back frame by frame with no full local download, and deploy trained checkpoints back to hardware by altering a single keyword argument. All stages share the same LeRobot on-disk format, ensuring compatibility across simulations and physical hardware such as the SO-101 and SO-100 arms.