source: hugging face blog: from the hugging face hub to robot hardware with strands agents and lerobot
level: technical
strands robots is an open source sdk from aws that wraps the lerobot stack as agenttools. you compose these tools into one agent that handles recording demonstrations, running policies in simulation, and deploying to physical robots. the integration is thin: lerobot's own scripts manage hardware recording and calibration, while strands agenttools orchestrate the rest. simulation records datasets in the same format as hardware, and policies swap with a string. a peer mesh lets the agent coordinate multiple robots.
the workflow starts by building an agent with a robot tool. in simulation mode, you record a demonstration as a lerobotdataset using a mock policy for end-to-end testing. the dataset uses the same parquet and mp4 layout as hardware recordings, so lerobot's loader reads it without changes. you can push the dataset to the hugging face hub. for real demonstrations, you use lerobot's record cli directly, and the agent picks up the resulting dataset for policy training or inference.
to run a policy, you attach a policy provider like gr00t inference or lerobotlocal to the agent. the agent manages the inference container or loads a checkpoint from the hub. the same agent code deploys to a physical so-101 robot by changing one keyword argument to mode="real". for multiple robots, a built-in zenoh mesh handles discovery and coordination, letting the agent broadcast commands or stop all robots at once.
why it matters: it simplifies robot learning workflows by unifying simulation, training, and deployment under one agent, reducing tool fragmentation for ai and data science practitioners.
source: hugging face blog: from the hugging face hub to robot hardware with strands agents and lerobot