AWS's Strands Robots SDK unifies LeRobot from simulation to hardware deployment

Currently, building a robot that learns from demonstrations requires five separate tools—recording, training, simulation, hardware deployment, and fleet coordination—none of which talk to each other. AWS's new Strands Robots SDK solves this by exposing the LeRobot stack as AgentTools that compose into a single Strands agent. The SDK is open source under Apache 2.0, runs on Python 3.12+, and requires minimal dependencies: just MuJoCo for simulation, a model provider (e.g., Anthropic, OpenAI, Ollama), and the 'strands-robots[sim-mujoco,lerobot,mesh]' package.

The workflow is elegantly simple. Developers create a Robot("so100") object that defaults to simulation (safe, no GPU needed) and a Strands agent with that robot as a tool. The same agent code works on a physical SO-101 by changing a single keyword argument: mode="real". Five steps happen inside that call: record a demonstration as a LeRobotDataset in simulation, push it to the Hugging Face Hub, run a policy on the same robot, deploy identical code to hardware, and broadcast commands across a fleet via a Zenoh mesh. The dataset format is identical between sim and real, and policy providers like GR00T and LerobotLocal are interchangeable. A ready-to-run example is available on GitHub (examples/lerobot/hub_to_hardware.py) and runs on a laptop with no hardware, no GPU, and no HF credentials needed for the default path.