SourceRecord, train, and deploy from one place with Strands Agents, LeRobot, and Hugging Face Storage Buckets
This process reduces the time GPUs spend idle during downloads and lowers storage costs by only uploading changed bytes of large robotics datasets.
- Enables recording robot demonstrations directly to Hugging Face Storage Buckets using lerobot dataset format
- Uses Xet content-defined chunking to deduplicate data transfers during recurring syncs
- Allows training directly by streaming batches from the Hub without downloading full datasets locally
- Deploys trained policy checkpoints back to simulation or physical hardware with a single mode parameter change