Skip to main content

Data Tools

Use Data Tools after recording demonstrations. Work through the tools in this order:

Data Tools Workflow

Data Tools Tabs Overview

Review Episodes​

Use Review Episodes first. Check that the recorded data is usable before moving to the next tools.

Click Select Task and choose the recorded rosbag2 task folder under /workspace/rosbag2.

Review Episodes Task Selection

Open the task and replay each episode. Check the camera views and 3D viewer. Also review the joint data, task information, and playback timeline.

Review Episodes Player

Note the episode index when an episode has failed motion or missing camera data. Do the same for wrong task instructions or any other unusable data. Remove those episodes from Delete Episodes.

Delete Episodes​

Use Delete Episodes to remove failed or unwanted episodes from a recorded dataset.

Delete Episodes

  1. Select the task folder under /workspace/rosbag2.
  2. Enter the episode indices to delete.
  3. Keep Compact indices after delete enabled when the remaining episodes should be renamed back to 0..N-1.
  4. Click Delete.

Episode indices can be written one by one. Separate multiple indices with commas:

1, 3, 6, 10

Merge Rosbag Datasets​

Use Merge Rosbag Datasets when demonstrations for the same task were recorded across multiple rosbag2 task folders.

Merge Rosbag Datasets

  1. Add two or more source datasets from /workspace/rosbag2.
  2. Enter the merged dataset name.
  3. Check the output path preview.
  4. Click Merge.

The merged dataset is saved under /workspace/rosbag2.

Convert Dataset​

Use Convert Dataset after the rosbag2 dataset has been reviewed and cleaned. If needed, merge datasets before conversion.

Convert Dataset

  1. Select the task folder to convert.
  2. Set the target FPS.
  3. Select LeRobot v2.1 or LeRobot v3.0. You can select both.
  4. Click Convert Dataset.

Converted datasets are saved under /workspace/lerobot.

During conversion, Cyclo Data parses the recorded MCAP topics. It aligns actions and observations to the requested FPS. It also selects the matching MP4 frames using the camera timestamp files. Then it writes the LeRobot dataset output.

Hugging Face Upload & Download​

Use Hugging Face Upload & Download after conversion or after model training.

Data typeUse it for
DatasetUpload converted datasets from /workspace/lerobot so they can be used on an external training server.
ModelDownload trained model checkpoints back into /workspace/model before running Model Inference.

After downloading a trained model, continue to Model Inference.