LeLab Is Hugging Face’s New Browser-Based GUI for the LeRobot Ecosystem

The open source SO-101 robot arm created by Hugging Face is one of the least expensive ways you can get into serious robotics. For around $200 you can get all the parts needed to build one, and for a little bit more, a fully assembled SO-101 can be delivered to your door. Despite the price, which is quite low in the world of robotics, the arm is very capable. This means that useful robotic arms are now accessible to more people than ever before.

However, cost isn’t the only barrier to entry in this area. There is also a lot of complexity to deal with when it comes time to fire the machine up. Installing and configuring all of the required software packages, and learning how to use them, is more than many people care to deal with. Fortunately, Hugging Face is now making this process easier with a new graphical user interface called LeLab. With LeLab, users can easily teleoperate their SO-101, collect datasets, train policies, and much more.

LeLab is a browser-based front end for the broader LeRobot ecosystem, consolidating what would normally require numerous command-line tools into a single graphical environment. Rather than jumping between terminal windows and documentation pages, users can now perform calibration, teleoperation, dataset recording, model training, inference, replay, and dataset uploads directly from a web interface. The goal is to help newcomers go from unboxing a robot arm to training their first AI policy in a matter of minutes.

One of the first tasks most users will encounter is calibration, which can often be one of the more intimidating parts of setting up a robotic arm. LeLab addresses this with guided setup flows that walk users through calibrating both leader and follower arms without requiring any keyboard commands. Cameras can also be attached and configured through the interface to keep things simple.

Once the hardware is configured, LeLab provides tools for teleoperation. A user can manually move a leader arm while a follower arm mirrors those motions in real time. The software also includes live joint streaming and a 3D visualization that displays the state of the robot as it moves.

Through the interface, users can define task descriptions, specify the number of episodes to record, and configure timing parameters. These recordings are automatically organized into LeRobot datasets that can later be used to train machine learning models. Training can be performed locally on the user’s own computer or remotely through Hugging Face Jobs, which provides access to cloud-based GPUs. During training, progress logs and checkpoints are displayed directly within the application. Once a model has been trained, users can select it from within LeLab and deploy it back onto their robot for inference with just a few clicks.

The software also supports replaying previously recorded episodes, uploading datasets directly to the Hugging Face Hub, and training on datasets created by other members of the community. By consolidating all of these capabilities into a single interface, LeLab removes much of the complexity normally associated with robotics and machine learning, making platforms like the SO-101 more accessible than ever.