Hugging Face and AWS integrate LeRobot datasets and policies with Strands Robots SDK for robotics workflows

Hugging Face and AWS announced an integration that connects the Hugging Face Hub’s LeRobot datasets and policies with AWS’s Strands Robots SDK, enabling a single agent loop to move from dataset to physical robot. The workflow uses a shared LeRobotDataset format across simulation and hardware, so datasets recorded in MuJoCo simulation match those captured on physical robots such as the SO-101. The Strands Robots SDK exposes LeRobot’s stack as AgentTools that can be composed into a single Strands agent, orchestrating recording, training, simulation, deployment, and fleet coordination.

The integration allows users to record demonstrations in simulation, push datasets to the Hugging Face Hub, run policies in simulation, deploy the same agent code to hardware with a single argument change, and coordinate multiple robots via a built-in peer mesh. The example agent performs four tasks: record demonstrations in simulation, push results to the Hub as a LeRobotDataset, run a policy in simulation, and deploy to physical hardware with a keyword argument change. For hardware-specific steps like recording and calibration, LeRobot’s own CLIs handle bring-up, while the agent manages orchestration.

The workflow is designed to run end-to-end on a laptop in simulation without hardware, GPU, or Hugging Face credentials for the default path. The provided example application at examples/lerobot/hub_to_hardware.py and hub_to_hardware.ipynb demonstrates the workflow, with the notebook running in simulation mode by default using a Mock policy. The MuJoCo-backed simulation and hardware paths share the same DatasetRecorder and policy providers, ensuring format consistency between simulated and real-world datasets.

To use the integration, users install Strands Robots with specific extras and optionally set a Hugging Face token for dataset and policy operations. The default simulation path requires only Python 3.12+, a compatible model provider (e.g., Amazon Bedrock, Anthropic API, OpenAI, or Ollama), and the Strands Robots SDK. For hardware deployment, additional requirements include a supported robot like the SO-101, calibration files, and for local GR00T inference, an NVIDIA GPU with at least 16 GB VRAM and Docker.

The integration’s design emphasizes minimal coupling: LeRobot handles hardware-specific recording and calibration, while Strands Robots provides the orchestration layer. The agent code remains identical whether targeting simulation or hardware, with mode="sim" (default) or mode="real" determining the backend. This design choice ensures that datasets captured in MuJoCo and on physical hardware use the same format, enabling seamless transitions between development and deployment.