Bigenlight/carrot_in_pot_lighting_raw

Carrot in Pot — Raw Collection (Additional Lighting)

This repository preserves the original, unconverted data from the 2026-09-18 carrot_in_pot collection session. It is the raw source for the corresponding LeRobot-formatted dataset.

Task scene captured on 2026-09-18. The two visible softbox lights are the added-lighting setup used during this collection.

Collection condition: additional lighting

Additional task lighting was installed and used throughout this entire collection session. This is the primary difference from the earlier carrot_in_pot dataset. Keep this domain change in mind when comparing, training, or evaluating across the two datasets.

Contents

• 65 takes, stored verbatim under their original take directory names.
• Each take contains:
  • vectors.h5 — recorded vectors / robot trajectory data.
  • cam1.mp4 — camera 1 RGB video.
  • cam2.mp4 — camera 2 RGB video.
• No depth recordings are included in this session.

The take directory names have deliberately not been renamed or normalized. The raw layout is:

take_<number>_<timestamp>[/_recover|_recover_twice]/
  vectors.h5
  cam1.mp4
  cam2.mp4

Recovery-labelled takes

_recover and _recover_twice are original collection labels, preserved exactly in directory names and in the LeRobot conversion. They identify trials where the task attempt failed and was retried once or twice. They are not arbitrary filenames; retain these labels when selecting recovery-focused evaluation subsets.

Recovery-labelled takes:

take_06_20260918_202527_recover
take_07_20260918_202637_recover
take_08_20260918_195306_recover
take_08_20260918_202709_recover
take_09_20260918_202737_recover_twice
take_41_20260918_200952_recover
take_42_20260918_201017_recover
take_44_20260918_201121_recover
take_45_20260918_201147_recover

Using the data

Use these files if you need the source recordings or want to run a different conversion pipeline. For standard LeRobot training/loading, use Bigenlight/carrot_in_pot_lighting_lerobot_v3 instead.

The raw takes were converted with the repository's convert_takes_to_lerobot.py workflow and independently validated: all 65 takes passed, with 20,472 RGB frames in the converted dataset.