CarolinePascal 0d6e4f3bad feat(features): route 2D camera shapes to observation.depth.<key> ...

hw_to_dataset_features now treats a camera entry whose shape has
length 2 as a single-channel depth feature: it emits the feature as
"{prefix}.depth.<bare>" with names=["height", "width"] and an
info={"video.is_depth_map": True} marker so the depth-encoder branch
in LeRobotDataset is engaged. The "_depth" hardware-side suffix (if
present) is stripped so a paired RGB + depth camera ends up as
"observation.images.<cam>" + "observation.depth.<cam>".

build_dataset_frame mirrors the routing: depth feature keys read
their value from "<bare>_depth" in the raw observation dict, with
fallback to the bare name for producers that already emit
dataset-style keys.

Tests: add tests/utils/test_feature_utils.py covering the routing
of 2D vs 3D camera shapes, the paired RGB+depth case, and the
build_dataset_frame value routing.

Made-with: Cursor

2026-05-01 00:49:09 +02:00

3.0 KiB

Raw Blame History

View Raw