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

August 13, 2026 | Source: Hugging Face Blog

Tags: AWS, Strands Robots, LeRobot, Hugging Face, robotics, open-source, AI training

AWS's open-source Strands Robots SDK (Apache 2.0) now supports a full continuous robotics training loop via Hugging Face Storage Buckets — record demonstrations, train on the Hub, and redeploy to hardware in a single agent workflow without redundant full-dataset transfers.

Details

Amazon Web Services and Hugging Face jointly published a technical guide showing how AWS's Strands Robots SDK enables a self-improving robotics pipeline. A single Strands agent records robot demonstrations, pushes episodes to Hugging Face Storage Buckets, trains a policy directly from the Hub dataset, and deploys the updated policy back to hardware — all in the LeRobot data format throughout. The key efficiency gain is around data transfer costs. Previous workflows re-uploaded entire growing datasets on each training run. Hugging Face Storage Buckets (launched March 2026) are mutable, non-versioned, Xet-backed object storage in the hf:// namespace — agents can write new episodes incrementally without transferring the full dataset. The LeRobot format is the interoperability layer: over 90,000 datasets and 8,000 publishers already use it on the Hub, so Strands Robots recordings can be consumed by any LeRobot-compatible tooling without format conversion. The robot catalog supports SO-100/SO-101 arms plus humanoids, mobile bases, and hands. Strands agents automate pipeline decisions — which episodes to retain, when to retrain, which checkpoint to deploy — replacing manual engineering judgments that previously required direct intervention. The Strands Robots SDK is Apache 2.0 licensed for full commercial use.