Hugging Face's logo

Datasets:
lerobot
/
droid_100

Tasks:
Robotics
Modalities:
Tabular
Time-series
Video
Formats:
parquet
Size:
10K - 100K
ArXiv:
Tags:
LeRobot
Libraries:
Datasets
pandas
License:
Dataset card Data Studio Files
xet
 Community
2

droid_100: 25 episodes missing recoverable language labels, 32 truncated to 1 of 3 available annotations, 19 unmarked failures

#2
by ankurcw - opened
Discussion
ankurcw

Summary

While working with lerobot/droid_100, I ran a complete pass over all 100 episodes and cross-referenced them against the official DROID language annotations published in KarlP/droid (droid_language_annotations.json, the 3×-annotation set released Dec 2024). I found three label-quality issues, all verified by watching the episode videos:

Issue Count What it is
Dropped labels 25 Episodes ship blank in the port, but full 3× annotations exist in KarlP/droid
Truncated labels 32 Episode carries 1 label where 3 distinct annotations are available
Unmarked failures 19 Episodes from /failure/ paths in raw DROID, shipped with no success/failure flag

Net effect on task diversity: the port exposes 47 distinct task strings; 174 distinct human annotations are available for these same 100 episodes.

How I verified

  • Matched each port episode to its raw DROID episode via episode_id_to_path.json, normalizing on the stable INSTITUTION/{success,failure}/DATE/TIMESTAMP path tail. Port episode order matches RLDS order 1:1 (confirmed by comparing existing labels).
  • For all 25 dropped-label episodes, I watched the video and confirmed a recovered annotation correctly describes the demonstration. 25/25 matched.
  • Outcome (success/failure) read directly from the raw DROID file_path.

Two things I'd flag rather than auto-fix

  • Episode 42: marked success, but the demo appears to fail (the remote is dropped from the gripper mid-motion). The success flag looks questionable.
  • Episode 64: of the 3 recovered annotations, variant 1 has the wrong movement direction; variant 2 is accurate. A minor annotation-quality issue in the source set, noted for transparency.

Scope I'd propose for a fix

Conservative, in order of confidence:

  1. Restore the 25 dropped labels (all 3 annotations each).
  2. Expand the 32 truncated episodes from 1 → 3 annotations.
  3. Add a success/failure field per episode from the raw path.
  4. Leave untouched but note: 16 success episodes that have no annotation in KarlP/droid (likely in the ~5% the annotation set doesn't cover) — flagged as "unlabeled," not fabricated.

Offer

I have the full per-episode audit (CSV) and a script that reproduces the join end-to-end. Happy to open a PR in whatever format you prefer for v3.0 task metadata — wanted to confirm the preferred approach (regenerate via LeRobotDataset tooling vs. direct metadata edit) before submitting, so the contribution lands cleanly. Let me know and I'll send it.

The 25 dropped-label episodes (blank in port → recovered + video-verified)

All 25 confirmed by watching the episode video against the recovered annotation.

Ep Outcome Restored primary label Note
7 success Pick up the orange block on the right and put it on top of the orange block
8 success Put the metal object with the orange handle in the blue cup
14 success Move the orange towel to the left all 3 variants fit
15 success Flip over the container in the cupboard
16 success Pick up the blue block and put it in the white mug on the right
17 success Put the cloth on the table all 3 variants fit
31 success Press a button on the white remote control
32 success Put the green packet inside the plastic bowl
36 success Put some of the blue bowl's contents on the paper, then fold the paper
37 success Use the paper towel to clean the surface
38 success Move the fleece to the left a bit all 3 variants fit
42 success Move the black remote control to the left ⚠ outcome questionable — remote appears dropped mid-motion; success flag may be wrong
43 success Move the brown object to the right
45 success Place the calculator on the coaster only variant 3 identifies the object as "masking tape" — illustrates why keeping 1 of 3 loses information
46 success Place the small bottle on top of the air fryer
51 success Place the lid on the silver pot
58 success Pick up the blue cup and put it in the pink cup
63 success Press down on the remote
64 success Move the blue spoon forward variant 1 has wrong direction; variant 2 preferred
72 success Lift the pink bowl and place it into the green bowl, pick up the yellow cup and place it into the stacked bowls, then place them on top of the microwave long-horizon
78 success Move the white container backwards
88 success Put the silver lid on the stove
94 success Open the door to the fridge
95 success Move some letter tiles to the left
99 success Pick up the scoop in the yellow bowl and move a scoop of its contents into the styrofoam cup in the blue bowl, then return the scoop long-horizon

Each of these has 3 distinct annotations available in KarlP/droid; only the primary is shown here for brevity. Full per-episode audit (all 100 episodes, 3× annotations, defect classification) and a script reproducing the join end-to-end are available — happy to push to a public repo.

Sign up or log in to comment