2skill3d4u/droid-qc-eval

DROID QC eval set

200 teleoperation episodes from DROID 1.0.1 (100 labelled successful, 100 failed by the people who collected them), the frame grid a vision-language judge sees for each, and the verdicts of claude-opus-5-5. Built with teleop-qc, which measures how far a VLM judge of robot task success can be trusted before it replaces human review in a data pipeline.

What is in a row

column	meaning
episode_index	episode in lerobot/droid_1.0.1
grid	the image the judge sees (968x1100 JPEG): 3 cameras in columns, 6 moments in rows, first at the top
human_success	DROID's success label, set by the operator who collected the episode
instruction	DROID's language instruction, "" when missing. Never shown to the judge (see below)
frames, seconds	episode length at 15 fps
videos	where each camera stream lives in the DROID video files (JSON)
judge_*	the verdict: success, confidence (0.5 to 1), p_success, failure mode, evidence, attempts, tokens

Columns of the grid, left to right: exterior camera 1, exterior camera 2, wrist camera. judge_p_success is the confidence for a success verdict and 1 - confidence for a failure. The judge abstains instead of guessing when its answer fails a schema and consistency check twice: abstentions have null judge_success and judge_p_success.

How the episodes were drawn

100 successful and 100 failed episodes, drawn with seed 0 from the first metadata file of DROID 1.0.1 (14,153 episodes). The set is frozen as evalset/droid_evalset_v1.jsonl; a change makes a new version.

Why the judge never sees the instruction. In that metadata file, 94.6% of the episodes labelled successful have a language instruction, against 0.3% of the failed ones: the instructions were crowdsourced after collection, mostly for successful episodes. A judge shown the instruction could read the label off its presence.

Results of claude-opus-5-5

Command: teleop-qc judge --evalset evals/droid_evalset_v1.jsonl --out runs/claude-opus-5-5.jsonl (effort medium, structured outputs), then teleop-qc report.

metric	value
episodes judged	200 of 200 (0 abstentions)
balanced accuracy at p = 0.5	0.755
human successes judged successful	0.830
human failures judged failed	0.680
Cohen's kappa	0.510
AUROC (95% bootstrap interval)	0.811 (0.747 to 0.868)
Brier score	0.179
ECE (10 uniform bins)	0.064
ECE after cross-fitted isotonic recalibration	0.051
share accepted without review at 95% precision	0.000
share rejected without review at 90% precision	0.130
cost of the run	$2.92
cost per episode	$0.0146
cost per hour of episode	$2.89

Files

• data/test-*.parquet: the table above, grids included.
• evalset/droid_evalset_v1.jsonl: the frozen eval set, as read by teleop-qc judge.
• verdicts/claude-opus-5-5.jsonl: the judge's raw output, one line per episode.
• judge/system_prompt.txt, judge/answer_schema.json: what the judge is told and the JSON schema its answer must follow.

License and citation

Apache-2.0, like DROID. The frames come from DROID:

@article{khazatsky2024droid,
  title   = {DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset},
  author  = {Khazatsky, Alexander and Pertsch, Karl and others},
  journal = {arXiv preprint arXiv:2403.12945},
  year    = {2024}
}