Datasets:
Independent 5-dim QC of LeRobot-compatible datasets (13 embodiments, 15 sets) - incl. a pre-registered utility criterion
Hi! We're an independent data-QC team (Zhiyong Flywheel / Nautilus) running a five-dimension quality-assessment pipeline over LeRobot-compatible datasets on the Hub, and your challenge_data sample was the latest set we ran. Wanted to share our readings and open the door for disputes/collaboration.
What we measure (frame-level, auditable): visual quality (freeze/blur/black-face), compliance (face detection + credential scan + redaction), hand landmarks, semantic validity, native resolution - plus a pre-registered downstream-utility criterion.
Readings on example_data/robot_data_sample (9 episodes, 3 cameras, 4535s):
- Overall very solid: zero FAIL on the four visual checks (freeze mean 2.5%, max 6.9% - healthy teleop; black frames 0). Resolution 1280x720 across all 54 files - notably better than the 4-of-14 datasets we measured earlier that ship at <=128px shortest side.
- One edge case worth flagging: our quality-dim face sampler flagged a few face frames (up to 0.9% of sampled frames on one head-camera episode) while our full-frame compliance redaction pass (YuNet @0.85) detected zero faces. A detector-disagreement edge case on household-domain data - possibly interesting for your data governance loop. Happy to share per-frame outputs.
- Decoder note: AV1 decode silently fails on some platforms (cv2 returns 0 frames without error); we ship dual-path decode (cv2 -> PyAV fallback) after being bitten once.
Broader context: we've now QC'd 15 sets across 13 embodiments (PushT, ALOHA, DROID, Franka, xArm, SO-100/101, GR-1 humanoids, this PrimeBot bimanual set, etc.). One pre-registered finding: feeding the same held-out pool to a fixed fine-tuned VLA consumer at three input scales shows -30.5pp accuracy at 96px (n=200 paired, McNemar p<1e-5) - while a zero-shot 4.2B VLM on the identical pool shows essentially none (0.825 -> 0.830). Resolution damage looks like a data x consumer property, not a property of the pixels alone.
Offer: while we scale coverage, we do independent five-dimension QC reports free of charge for community datasets. If you'd like a full report on official_data batches (or want to dispute any reading above), reply here - our criteria lock, raw per-frame outputs and manifests are all auditable and we're happy to share.