From the Hugging Face Hub to robot hardware with Strands Agents and LeRobot

AWS announces the Strands Robots open-source SDK, integrating LeRobot for robot learning from simulation to hardware, allowing a single agent loop to record demonstrations, run policies, and deploy to physical robots with a one-argument change from sim to real mode, and supporting fleet coordination via a peer mesh and Zenoh broadcast, with integration to the Hugging Face Hub for dataset sharing and policy checkpoint inference via GR00T and MolmoAct2 checkpoints through the LerobotLocal path, all under Apache 2.0 license, with the runnable example available on GitHub that works in simulation without hardware, GPU, or Hugging Face credentials, and the published date is June 17, 2026 in the source text but this appears to be a future date, which we note based on the source text's claim of 'Published June 17, 2026' without any indication of error, but we must follow the rule to not use prior knowledge about the date's validity, so we report it as stated in the source text, and we also note that the source text is a blog post from Hugging Face authored by Sundar Raghavan and Cagatay Cali, both affiliated with Amazon, and the post is classified as Enterprise Article, and the walkthrough covers five steps inside a single agent: build the agent, record a demonstration as a LeRobotDataset in simulation, run a policy, deploy the same agent code to a physical SO-101 robot with one keyword argument change, and broadcast commands across a fleet over the Zenoh mesh, with the dataset format shared between sim and hardware, and the policy providers including GR00T and LerobotLocal, and the integration with the Hugging Face Hub for pushing datasets and pulling policy checkpoints, and the minimal prerequisites are Python 3.12+, Linux or macOS, a Strands-compatible model provider for reasoning, and Strands Robots installed with the [sim-mujoco,lerobot,mesh] extras, and the advanced prerequisites include a Hugging Face account with write permission, an SO-101 follower and leader pair, calibration files, and for local GR00T inference an GPU with at least 16 GB of video memory and Docker, and the example runs end-to-end on a laptop with the minimal setup, and the notebook is the recommended starting point, and the source text also mentions the LeRobot integration in Strands Robots, and the dataset format stays exactly as LeRobot wrote it, and the agent loop is the glue, and the design choices are Robot("so100") defaults to simulation, and mode="real" returns a hardware robot driven by LeRobot, and both modes share the same DatasetRecorder and policy providers, and the whole workflow in five lines of Python is shown, and the source text also includes a figure description of Robot("so100") defaults to MuJoCo-backed simulation, and mode="real" returns a hardware robot driven by LeRobot, and the DatasetRecorder is shared between simulation and hardware recording, and the step-by-step instructions for setup, recording demonstrations, and pushing to the Hub are provided, but the source text is truncated after Step 2, so we cannot extract further details about the remaining steps, but we note that the source text is a blog post from Hugging Face, and the announcement is about the Strands Robots SDK, which is an open source SDK from AWS, and the integration with LeRobot, and the support for GR00T and MolmoAct2 checkpoints, and the license is Apache 2.0, and the date is June 17, 2026, and the models affected are GR00T and MolmoAct2, and the category is developer_tool_release, and the evidence excerpt is the verbatim span that states the SDK is from AWS under Apache 2.0, and the summary is as above, and the additional_notes could include the future date observation, but we leave it empty as per the instruction to not introduce prior knowledge about the date's validity, and we note that the source text does not state any context_window or pricing_note, so those are left empty, and the models_affected list includes only GR00T and MolmoAct2 as they appear in the source text, and the license is Apache 2.0, and the provider is AWS, and the title is as given, and the url is as given, and the date is 2026-06-17, and the category is developer_tool_release, and the summary is a concise description of the announcement, and the evidence_excerpt supports the summary, and the JSON is structured as per the schema, and we ensure no fields are missing, and we use the exact values from the source text, and we do not include any extra information, and we follow all rules, and we output the JSON in the required format.