Train AI Robots Without Writing Code! (Introducing LeLab)

In this talk

1. 0:00

   Set up robots and cameras

   LeLab provides a visual workflow for using the LeRobot library, from robot control and data collection to model training and deployment. The presenter configures leader and follower robots, detects their USB ports, calibrates their joints, and adds top and wrist cameras.

2. 1:22

   Practice and record demonstrations

   The teleoperation page lets users control the follower robot with the leader arm and view a 3D representation. For a task, the presenter recommends practicing first, then recording at least 30 episodes, with 50 as the suggested target. Consistent, smooth demonstrations help make the dataset useful for training.

3. 2:28

   Review and upload the dataset

   Failed episodes can be reset and recorded again, helping keep the data consistent. The example task is placing a pen in a holder; recording 50 episodes takes roughly 15 minutes. The dataset can be uploaded to Hugging Face Hub and reviewed online episode by episode.

4. 3:10

   Train and try checkpoints on the robot

   Users can train on a local computer or use GPU hardware through Hugging Face Jobs, choosing a dataset, policy, training steps, and batch size. Training progress, losses, and logs are available to monitor, and multiple runs can happen in parallel. In the example, an ACT model struggles at 3,000 steps but completes the task at 30,000, including with small changes in object positions.