Strands Robots SDK Wraps LeRobot Into One Agent Loop From Hub to Hardware

Five separate tools to record demonstrations, train, simulate, deploy, and coordinate multiple robots. That's the old way. AWS's Strands Robots SDK (Apache 2.0) collapses that pipeline into one Strands agent loop, wrapping the LeRobot stack as AgentTools. The integration is deliberately thin: LeRobot's own CLIs handle hardware recording and calibration, and the agent orchestrates everything from simulation to fleet management behind the same API.

Five Tools In, One Agent Out You have a robot, a folder of demonstration data on the Hugging Face Hub, and a new task. Today that means juggling a recorder, a trainer, a simulator, a deployment script, and a coordination layer. Strands Robots exposes Robot("so100") which defaults to a MuJoCo-backed simulation. Flip mode="real" and the same agent code drives a physical SO-101. Both modes share the same DatasetRecorder, so a dataset captured in sim and one captured on hardware use the identical on-disk LeRobotDataset format. No format conversion, no custom glue.

Record, Train, Deploy in Five Lines of Python The example agent does four things: record new demonstrations in simulation, push the result to the Hub as a LeRobotDataset, run a policy in simulation against that same format, and deploy the same agent code to a physical robot with one keyword argument change. The whole workflow in five lines: ```python

from strands_robots import Robot
from strands import Agent
arm = Robot("so100") # mode="sim" - safe, no hardware
agent = Agent(tools= )
agent("Pick up the red cube")

## From Laptop Simulation to Fleet Coordination When you need more than one robot, a built-in Zenoh peer mesh fans the agent out to remote robots. The same agent code that ran on your laptop now coordinates a whole fleet. For hardware deployment, LeRobot's own CLI commands (lerobot-record, lerobot-calibrate) handle bring-up; the agent picks up from there. Policy inference uses a common interface: GR00T and LerobotLocal serve predictions, and MolmoAct2 checkpoints run through the LerobotLocal path. You swap policies with a string. The runnable companion lives in the strands-labs/robots repository at examples/lerobot/hub_to_hardware.py and hub_to_hardware.ipynb. The notebook ships with a Mock policy and zero hardware dependency. Clone it, run it on your laptop, and you've gone from the Hugging Face Hub to robot hardware without breaking a sweat.

---

Source: [From the Hugging Face Hub to robot hardware with Strands Agents and LeRobot](https://huggingface.co/blog/amazon/strands-lerobot-hub-to-hardware)
Domain: huggingface.co