AWS releases Strands Robots, an open-source SDK that integrates robot control, simulation, and the LeRobot AI stack into a single agent workflow — letting researchers move from demonstration data to physical robots with one keyword argument change.

What happened: AWS released Strands Robots, an Apache 2.0 open-source SDK that bundles robot abstractions, MuJoCo simulation, and LeRobot policy tools into composable AgentTools. A single agent can now record demonstrations in simulation, push datasets to the Hugging Face Hub, run policies against the same data format, and deploy to physical hardware (like the SO-101 robot) or coordinate multiple robots via a peer mesh — all without switching tools.

Why it matters: Previously, the same task required five separate tools that didn't communicate with each other. By unifying the workflow around a shared LeRobotDataset format and a single agent loop, Strands Robots reduces friction for roboticists who want to move from training in simulation to real hardware. The example runs end-to-end on a laptop in simulation mode with no GPU, no hardware, and no Hugging Face credentials needed for the default path — lowering the barrier to experimentation.

What to watch: The sample application (hub_to_hardware.py and hub_to_hardware.ipynb) is available in the strands-labs/robots GitHub repository and runs in JupyterLab. Support for GR00T and MolmoAct2 checkpoints is included; LeRobot's own hardware recording and calibration CLIs (lerobot-record, lerobot-calibrate) handle bring-up, and the agent picks up from there.