Nicolas Rabault's LeLab gives LeRobot users a GUI for robot training

Why it matters

Robot learning is moving from research labs toward broader developer use, and tools like LeLab try to make that shift by replacing command-line workflows with productized interfaces.

LeRobot (@LeRobotHF) launched LeLab, a graphical interface built on top of the LeRobot library, in a three-post thread on X on Wednesday.

https://x.com/LeRobotHF/status/2062192165216735385

The project credits Nicolas Rabault (@rabault_nicolas) as the builder of LeLab. The tool is not a new robot arm; it is a front end for the LeRobot workflow, meant to take users from hardware setup to teleoperation, dataset collection, policy training and deployment without requiring them to memorize command-line steps.

Source code and issues are hosted in the Hugging Face leLab GitHub repository. Installation still starts like a developer tool: LeLab requires uv, is installed from the GitHub package, and then launches from the terminal with lelab.

LeLab's app flow covers the core robotics-learning loop. Users can add robots by selecting the arm type, including leader and follower roles, calibrating each joint from the middle position, and attaching cameras. They can teleoperate the follower arm with the leader arm while watching a live 3D visualization.

For data collection, LeLab lets users define a task description, number of episodes, and episode and reset durations, then advance between episodes with the spacebar. The guide recommends more than 30 episodes. Users can also provide a Hugging Face dataset ID to train on community datasets they did not record themselves.

Training can run locally, with users choosing a dataset, policy type, batch size and step count, or in the cloud through HF Jobs after running hf auth login. LeLab says HF Jobs support comes with transparent pricing and points users to the Compute HW Guide for hardware and batch-size guidance. Training progress is visible inside the app, and checkpoints are saved automatically.

Deployment is also part of the interface: users can choose a trained model from their jobs list and run inference on the robot with one click.

The important limitation is hardware support. For now, LeLab is compatible only with SO-ARM101. That makes the first release more of a focused onboarding layer for one LeRobot-compatible setup than a general-purpose robotics control plane.

For Rabault and the LeRobot project, the bet is clear: lowering setup friction could expand the pool of people who can collect robot data, run training workflows and test policies. Whether LeLab becomes a serious entry point depends on how well the GUI handles the messy parts of calibration, data quality, training failures and deployment once users leave the happy path.