Robot Data Audit (RDA)
Audit robot dataset quality. Local only.
RDA checks LeRobot-format datasets for data integrity, temporal consistency, and motion anomalies. It flags problematic episodes and reports what's wrong. RDA is a diagnostic tool only — it does not guarantee training success rate improvements.
Quick Start
pip install robot-data-audit # install
rda audit /path/to/dataset # audit → text + JSON report in <dataset>/rda_report.json
Sample Output
$ rda audit ~/datasets/my_robot_data
── Verdict ──
PASS: 180 (90.0%)
REVIEW: 15 ( 7.5%)
EXCLUDE: 5 ( 2.5%)
── Top Issues ──
1. [HIGH ★] Action discontinuity: 3421 spikes across 195 episodes
2. [MEDIUM] High idle ratio: median 72% idle, 28% effective motion
3. [LOW] Extreme acceleration spikes: 2847 across 180 episodes
Output options: --format json for scripting, -o FILE to save elsewhere.
Real Audit Results
Audited on ArmnetBench (SO-101 arm, 2,499 episodes) and DROID (100 episodes):
| Metric | ArmnetBench (200 ep subset) | DROID (100 ep) |
|---|---|---|
| Action spikes detected | 5,229 | 1,428 |
| Median idle ratio | 68.6% | 70.7% |
| Median episode duration | 22.4s | 15.0s |
| Verdict | All PASS | All PASS |
Both are curated benchmark datasets — all episodes pass. RDA also runs on noisier, real-world collections where REVIEW/EXCLUDE verdicts appear more frequently.
Full calibration analysis (ArmnetBench, comparing successful vs failure episodes by label): docs/ARMNETBENCH_CALIBRATION_REPORT.md
Design
- Local only — No data leaves your machine. Runs entirely offline.
- Diagnostic, not predictive — RDA identifies data issues. Whether fixing them improves training is a separate question and depends on your task, model, and setup.
- Statistical anomaly detection — Uses MAD on reference distributions instead of fixed thresholds (no hardcoded 3σ rules). Adapts to each dataset's characteristics.
- Universal core metrics — Primary ranking uses 3 platform-independent metrics (duration, spike_count, effective_motion_ratio). Platform-specific signals (velocity, path_length) are optional diagnostics.
See docs/MVP_PRODUCT_SPEC.md for full metric definitions.
Metrics
| Tier | Metric | Detects | Cross-platform? |
|---|---|---|---|
| L1 | Timestamp monotonicity | Clock resets, duplicates | ✅ |
| L1 | Frame interval consistency | Irregular sampling | ✅ |
| L1 | Schema compliance | Missing/extra fields | ✅ |
| L1 | Numeric sanity | Values exceeding loose physical bounds (angle/force/torque/velocity) | ✅ |
| L2 | Temporal gap detection | Time discontinuities | ✅ |
| L2 | Sensor synchronization | Multi-sensor drift | ⚠️ |
| L2 | Temporal sufficiency | Idle vs active structure | ✅ |
| L3 | Velocity spikes | Implausible jumps | ️ |
| L3 | Motion discontinuities | Jerky trajectories | ⚠️ |
| L3 | Idle frame detection | Paused segments | ✅ |
| L4 | Duration outliers | Too short / too long | ✅ |
| L4 | Spike count outliers | Unusual jerk profiles | ✅ |
| L4 | Effective motion ratio | Low-activity episodes | ✅ |
| SP | Privacy/PII exposure scan (pii_scan) |
Identifiers in text surfaces, precise geolocation features, faces in sampled frames (--pii-video) |
✅ |
pii_scan(v0.9.22+) covers YD/T 6771-2026's 安全性 dimension — the only P0 admission gate among the eight quality dimensions. It reports evidence a human must review, not compliance: WARNING-level by default (never changes the verdict), promoted to a gate only with--pii-strict. No OCR (licence plates, badges), no voice biometrics, no consent review. Findings carry a masked snippet plus a SHA-256 fingerprint — raw identifiers never reach the report.
CLI Reference
rda audit
rda audit /path/to/dataset [OPTIONS]
| Option | Description |
|---|---|
-o, --output FILE |
Save JSON report (default: <path>/rda_report.json) |
--format [json|text] |
Output format (default: text) |
--platform TEXT |
Robot platform name for Tier 3 normalization |
--blind |
Redact identifying paths and operator-authored free text for externally shareable reports |
--pii-video |
Extend the privacy scan to sampled video frames (needs [pii] extra) |
--pii-strict |
Treat privacy findings as an acceptance gate instead of a warning |
-v, --verbose |
Verbose output |
Exit codes: 0 = no EXCLUDE, 1 = error, 2 = at least one EXCLUDE,
3 = audit completed but nothing was effectively audited (zero episodes, or
every episode lacked auditable content) — never report this as a pass.
Python API
from rda.audit.dataset_audit import DatasetAuditor
from rda.io.lerobot_loader import iter_episodes, load_lerobot_dataset
dataset_info = load_lerobot_dataset("/path/to/dataset")
auditor = DatasetAuditor()
result = auditor.audit_dataset(dataset_info, iter_episodes("/path/to/dataset"))
print(f"PASS: {result.verdict_counts['PASS']}")
Development
git clone https://github.com/liesliy/rda.git
cd rda
pip install -e ".[dev]"
pytest
Citation
@software{robot_data_audit,
title = {Robot Data Audit: Quality Auditing for Robot Manipulation Datasets},
author = {Niu Su Tech},
year = {2026},
url = {https://github.com/liesliy/rda}
}
License
MIT
Metadata
Release files for robot-data-audit 0.10.0
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| robot_data_audit-0.10.0.tar.gz | 275.5 kB | Without provenance | Details for robot_data_audit-0.10.0.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| robot_data_audit-0.10.0-py3-none-any.whl | Python 3 | none | any | Details for robot_data_audit-0.10.0-py3-none-any.whl |
Total release size: 589.9 kB
Release files / robot_data_audit-0.10.0.tar.gz
| Download URL | robot_data_audit-0.10.0.tar.gz |
|---|---|
| Size | 275.5 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
7588796fda3771294e097381edbf21465b72468ff57f22913b07e8218603ad4b
|
|
BLAKE2b-256 checksum How to use checksums |
e4ab0c45712f384ba1704a523282873517ededad2bf14270fba4fbece1159379
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.2.0 CPython/3.11.9
|
Release files / robot_data_audit-0.10.0-py3-none-any.whl
| Download URL | robot_data_audit-0.10.0-py3-none-any.whl |
|---|---|
| Size | 314.3 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
05d0f8766567a78833df682a20426f9e043cbb2e88f99e0910549d603a2ce09d
|
|
BLAKE2b-256 checksum How to use checksums |
3a867c2489fcb2a31dd528f75ab109e5976e29feee58f908b277e19453f0d2a4
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.2.0 CPython/3.11.9
|
Security for robot-data-audit 0.10.0
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
No provenance found
None of the files in this release were uploaded with provenance metadata.
A note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.