Hugging Face and AWS introduce streaming data loop for robotics training and deployment

Hugging Face and AWS announced a workflow that ties together Strands Agents, LeRobot, and Hugging Face Storage Buckets to create a continuous loop for recording robot demonstrations, training policies, and deploying them back to hardware. The system is designed to run the loop daily: collect episodes, train on the growing dataset, deploy the updated policy, and pull new recordings back to improve the model iteratively.

The workflow leverages Hugging Face Storage Buckets, a mutable, non-versioned object storage type introduced in March 2026, to store recordings with byte-level deduplication. This avoids re-uploading unchanged bytes during each sync, reducing redundant data transfers and associated costs. The Storage Bucket sits alongside dataset repositories in the same hf:// namespace and is accessible via the hf CLI, making it a seamless working layer between recording and training.

Training occurs by streaming datasets directly from the Hugging Face Hub without downloading the full dataset locally. The agent reads the LeRobot-formatted data frame by frame, decoding camera video on the fly, and processes it in the same environment where it was recorded. This streaming approach eliminates the need to copy large datasets to GPUs before training begins, cutting storage and compute overhead.

The loop is orchestrated by a single Strands agent that can both record demonstrations and stream them back for training. The agent uses the same Robot() factory to interact with supported embodiments, such as the SO-101 robot, and toggles between simulation and real hardware via a mode parameter. A companion notebook at examples/notebooks/05_streaming_data_loop.ipynb demonstrates the full workflow, from recording a demonstration to deploying a trained policy.

To implement the loop, users need Python 3.12+ on Linux or macOS, a Strands-compatible model provider (e.g., Amazon Bedrock, Anthropic API, OpenAI, or Ollama), and the Strands Robots package with LeRobot extras. For advanced use cases involving Storage Buckets or hardware deployment, a Hugging Face account with write permissions, the hf CLI, and optional GPU resources for training are required. The workflow supports both simulation and physical robotics, with calibration files needed for hardware deployment.