Question: mapping between DreamZero-DROID-Data episodes and raw DROID 1.0.1 trajectories / frame indices

[Candy-Crusher]

Hi, thanks for releasing the preprocessed DreamZero-DROID dataset and the conversion script.

I am trying to align external per-frame / per-clip annotations derived from raw DROID 1.0.1 with the released GEAR-Dreams/DreamZero-DROID-Data LeRobot dataset.

From the docs, I understand that the released dataset is generated from DROID 1.0.1 by:

• applying droid_sample_ranges_v1_0_1.json to remove idle frames,
• filtering failed episodes,
• filtering episodes without language annotations,
• reindexing the remaining episodes into LeRobot format.

I have two questions:

1. Episode-level mapping:
   Is there a released mapping from each processed LeRobot episode, e.g.
   data/chunk-XXX/episode_YYYYYY.parquet / videos/.../episode_YYYYYY.mp4,
   back to the original DROID trajectory path, e.g.
   gs://xembodiment_data/r2d2/r2d2-data-full/.../trajectory.h5?

   If not, is the intended way to reproduce this mapping by rerunning scripts/data/convert_droid.py on raw DROID 1.0.1 and reconstructing the kept_registry?

2. Frame-level mapping:
   For a processed episode, are the video frame indices exactly the concatenation of the raw DROID frame ranges in droid_sample_ranges_v1_0_1.json?

   For example, if:
   keep_ranges = [[0, 24], [31, 171]],
   should processed video frame indices map as:

   • processed frames 0..23 -> raw frames 0..23
   • processed frames 24..163 -> raw frames 31..170

   In other words, is the released video frame index always the idle-filtered raw DROID frame index with the gaps removed?

This mapping would be very useful for aligning external annotations such as PointWorld-DROID flow clips, whose HDF5 groups are named like {start}:{end}.

Thanks!