AWS introduces Strands Agents to enable a continuous loop of recording, training, and deploying robot policies using LeRobot datasets and Hugging Face Storage Buckets. This integration allows agents to sync demonstrations into mutable buckets with byte-level deduplication and stream datasets directly from the Hub for training without full downloads.

  • Strands Robots exposes robot abstractions as AgentTools, allowing a single agent to record demonstrations and deploy policies across simulation and physical hardware like the SO-101.
  • Hugging Face Storage Buckets use Xet-backed content-defined chunking to reduce upload data by approximately four times by syncing only changed bytes.
  • The system supports streaming LeRobot datasets frame-by-frame for training, avoiding local copies and reducing bandwidth costs during iterative policy improvement.

This approach simplifies the robot development workflow by unifying data collection, storage, and training within a single agent-driven loop.