DreamLake
Browse docs

Data and views

An action vector can be plotted against time or drawn as a path. Both views read the same numbers, but answer different questions: when did the command change, and where did it move?

yaml
views:
  - view: lineChart
    series: [action]
  - view: trajectory2d
    series: [action]

The chart draws a trace for each column. The trajectory uses columns named x and y, or the first two columns when those names are absent.

Fields, payloads, views

The reader lists a dataset's fields: addresses such as action or a camera stream. Reading a field produces a payload that a view can consume. For a chart, that payload is timestamps and named numeric columns.

Stored field → format reader → timestamps + columns → lineChart

LeRobot, MCAP, and folder readers can supply that same chart input. Their file layouts differ, so support depends on the selected reader, not just on the file containing numbers.

Say what a tensor represents

A [21,3] tensor might be image keypoints with confidence scores or points in 3D. Bind it with the interpretation you intend:

yaml
- view: videoStack
  cameras: [cam_front]
  overlays:
    - { field: hand_pixels, as: keypoints, on: cam_front }

Here as requests keypoints, and on selects the image whose pixel coordinates they use. A shape check can catch an invalid array; it cannot tell whether the coordinates refer to the correct camera. Check the overlay against the image.

Share an episode clock

Within an episode, time-based views share a cursor. Hover previews a moment; seeking changes the playhead. A 3D view follows that time while its pointer controls the camera, so pair it with a chart or timeline.

Use episode-relative seconds for annotations and numeric traces. A correct shape with an offset clock will still show the wrong moment.