Hugging Face's logo

Datasets:
agibot-world
/
AgiBotWorld2026

Tasks:
Robotics
Modalities:
Image
Text
Languages:
English
Size:
1K<n<10K
Tags:
agibot
imitation-learning
embodied-ai
lerobot
real-world
dual-arm
License:
Dataset card Data Studio Files
xet
 Community
9

Third-party audit: we ran RDA (Robot Data Audit) over all 5 simulation tasks + a real-robot HG-DAgger package

#9
by 375720783jin - opened
Discussion
375720783jin

Hi AGIBOT World team — congratulations on the Theme 3 release. I maintain RDA (Robot Data Audit), an open-source quality auditor for LeRobot-format datasets (pip: robot-data-audit), and wanted to share third-party audit results rather than just say "nice release".

What we ran

  • Simulation: all 5 tasks (tidy_up_food_in_freezer, scoop_popcorn_to_bucket, take_bagged_food_to_cart, take_cup_to_cart, take_drink_to_cart) — 1,102 episodes, ~500MB, every task in the simulation tree
  • Real robot: 1 RL package (ReinforcementLearning/Home/task_12192/HG-DAgger/20857245_20857259, 3.57GB — the smallest; the others don't fit a local disk) — 10 episodes, 23,658 frames, g2a robot

All of it audited with RDA v0.5.5 from PyPI, zero format adaptation (the v2.1 auto-detection just worked).

Results

  • Integrity: 4,448 checks (missing-frame dropout / invalid values / schema consistency / timestamp validity) across 1,112 episodes — 0 failures. Constant 30fps timestamps. The data engineering behind this release is genuinely clean.
  • Verdict distribution: 31 PASS / 1,081 REVIEW / 0 EXCLUDE / 0 FAIL. Nothing that needs cleaning before training; the REVIEW signals are worth a human look, not a pipeline rejection.

The interesting part: RDA's discontinuity flags line up with your intervened labels

The real-robot package ships a frame-level intervened column, which gave us a rare chance to validate whether the tool's action-discontinuity spikes mark real events. We aligned all 1,712 reported spikes (the exact spike_indices from RDA's report) against takeover transitions:

  • 226 spikes (13.2%) fall within ±1 frame of an intervened on/off transition — about 3.1x the random-chance baseline (~4.3%) for these episodes (per-episode ratio ranges 1.6x–5.2x)
  • Spikes inside takeover segments overall are at chance level — most spikes are aggressive-but-normal teleoperation, not takeovers

We read this as: the tool's risk signals genuinely cluster around real events (human takeovers) while correctly not crying wolf about normal teleoperation dynamics — which is exactly why RDA labels these "observational risk signal, not confirmed corruption" instead of failing episodes.

Reproduce 

pip install robot-data-audit
rda audit ./your-dataset          # per-episode JSON report
rda recommend ./your-dataset --offline   # offline pruning suggestions

RDA does not read video files, so the 21GB video volumes aren't needed for auditing.

Scope, honestly stated

Coverage: 100% of simulation tasks, 1 of 50 real-robot packages (disk space, not avoidance). The whole repo uses the same v2.1 format (verified across 6 metadata files), so format-level conclusions extrapolate. Full per-episode audit JSONs are available if anyone wants to double-check.

Happy to run specific packages on request, or adjust anything if you'd prefer this posted differently.

Sign up or log in to comment