DreamLake
Browse docs

Dataset examples

Choose the example closest to how your data is stored. Each demo reads the .dreamrc published with its dataset; inspect the config, then change the field bindings in the playground.

1. LeRobot with annotation features

This recording includes an egocentric camera, hand keypoints, and subtask labels. Scrub to check whether the hand overlay and task interval agree with the image.

Keypoints are numeric features inside the dataset. Subtasks use an integer column with a label table. Use this arrangement when your annotation pipeline can write features alongside the original measurements.

Resolving dataset…

2. A folder of video and sidecar files

The ceramics recording keeps a video, a COCO keypoint file, and WebVTT labels in each episode directory. It produces similar views without putting the data in a LeRobot container.

The folder reader discovers annotations/. The config binds hand_keypoints as keypoints and subtasks as segments; changing a filename requires changing an explicit binding to that field.

Resolving dataset…

3. A folder of MCAP recordings

Each file is one nuScenes scene. Camera, lidar, and transform schemas provide the data for synchronized media and 3D views. Use the timeline to change time and drag within a 3D panel to inspect another angle.

This example uses nuScenes data under CC BY-NC-SA; consult the dataset README before reusing the recordings.

Resolving dataset…

Try another modality

The view examples isolate depth, point clouds, trajectories, and motion reconstruction. Start there when the question is how to display one field rather than how to arrange a complete dataset.