View a dataset

A camera shows what a policy did. An action trace shows what it commanded. Put them on the same clock to inspect the moment the behavior changed.

Open the dataset playground and paste this config into its editor. It displays three episodes from the public PushT dataset:

version: 1
storage: { driver: hf, repo: lerobot/pusht }
dataset:
  format: lerobot
  episodes: { limit: 3 }
views:
  - view: videoStack
    cameras: ['observation.image']
  - view: lineChart
    series: [action]
    title: Target position
    unit: px
  - view: timeline

Hover the chart to see the corresponding camera frame. Click the timeline to seek, or play the episode. PushT's action contains target x/y coordinates in image pixels; those become the chart's two traces.

Inspect another field

Replace the views list with views: [{ view: fields }]. The inventory lists the names and shapes available in each episode. Use a listed field in series to plot it; a vector field expands into one trace per dimension.

For example, change series: [action] to series: [observation.state] to inspect PushT's observed position. Keep the camera beside it to compare the numeric trace with the image.

Connect your data

Change storage.repo to a compatible LeRobot repository. For other layouts, choose a reader, then inspect its fields before writing bindings. A camera name or tensor shape can differ between datasets.

When you save a .dreamrc at the dataset root, you can omit storage: the host supplies the root in which it found the file. Keep storage in standalone configs, such as the playground example above.

View examples show annotations, trajectories, depth, and 3D data. To embed the dataset browser in React, follow the TypeScript example.