Browse docs
Read your files
First check whether a reader supports your recording's layout. Then check the fields you want to display. A supported container can still contain an encoding or tensor shape that a particular view cannot read.
Choose a reader
| Your data | dataset.format | Episode selection |
|---|---|---|
| LeRobot v2.0, v2.1, or v3.0 | lerobot | auto; metadata at meta/info.json |
| UMI Zarr ReplayBuffer zip or supported v3 directory store | umi | auto; use path to select a store |
| Indexed MCAP v1 recordings | mcap | auto for root files, a glob, or one path |
| Directories of videos, images, and tables | folder | A directory glob |
Zarr support covers the layouts used by the examples, not every codec or array layout. MCAP supports uncompressed, lz4, and zstd chunks with supported Foxglove, ROS 1, or ROS 2 schemas. HDF5 and RLDS/TFRecord need an export or a custom reader.
A folder of recordings
Select it with dataset: { format: folder, episodes: 'episodes/*/', fps: 30 }.
Each matched directory is an episode. Media keep their filenames; other files
use basenames as field names, so subtasks.vtt is bound as subtasks.
The reader also lists annotations/; its fields override matching names at
the episode root.
Numbered stills form a frame sequence. Set fps to the actual recording rate.
CSV and Parquet traces use a timestamp-like column when available, otherwise
row indices. For synchronization, give numeric data and annotations the same
episode clock as the media.
The shape each payload needs
| View input | Supported representation |
|---|---|
| Traces | Numeric scalars or vectors with timestamps |
| Image keypoints | [J,2] xy or [J,3] xy plus score, in image pixels |
| Depth | [H,W], [H,W,1], or a supported 16-bit PNG sequence |
| Point cloud | [N,3] xyz or [N,6] xyz plus rgb |
| Rigid pose | Translation plus quaternion, with the declared quaternion order |
| 3D points | [J,3] or flattened xyz coordinates |
| Labelled intervals | Start, end, and label; or a per-frame label column |
The reader must support how these values are stored. Use NaN for missing numeric keypoints rather than a zero coordinate. Depth uses zero for a missing reading. Check image dimensions, coordinate frames, and units in the rendered result.
Annotations: files beside the data
WebVTT/SRT supplies intervals, COCO supplies image keypoints, and glTF supplies
geometry. Declare external paths with
dataset.annotations.
COCO files need an fps value to map image indices to episode time.
Annotations may also live inside a container: for example, LeRobot keypoint features or an index column joined to a label table. The annotated dataset shows that arrangement.
Check playback
Use a browser-supported codec. Video seeking decodes from a preceding keyframe, so long gaps between keyframes can make scrubbing slow. Check seeking against your actual browser and storage service before diagnosing it as a chart delay.
HTTP storage must permit browser access and the range requests required by the
reader. Directory discovery also needs a listing service: Hugging Face provides
one; a static HTTP directory needs an
index.json manifest.