AWS Strands Agents stream robot training data via Hugging Face Storage Buckets

AWS has released Strands Robots, an open-source SDK under Apache 2.0 that wraps robot abstractions, simulation, and the LeRobot stack as AgentTools for building autonomous data-collection agents. In a new walkthrough, AWS engineers demonstrate a streaming data loop that records demonstrations into a Hugging Face Storage Bucket—a mutable, non-versioned, Xet-backed object-storage repo type launched in March 2026—then trains and deploys policies on the same physical robot (SO-100) without ever downloading the full dataset.

The key advance is efficiency: instead of copying entire recordings to GPUs before each training run, the agent streams frames straight from the Hub, decoding camera video on the fly. Syncing only changed bytes cuts repeated transfer costs, and the on-disk format stays exactly as LeRobot wrote it, compatible with the 90,000+ datasets and 8,000+ publishers already using that standard. The agent handles the mundane decisions—which episodes to keep, when to re-record, whether today's batch is enough, and which checkpoint to deploy—by inspecting previous results before issuing the next command. With one keyword-argument change, the same Robot() object switches from simulation to real hardware, closing the loop between recorded demonstrations and deployed policies.