Robot Data Audit (RDA)

Quality auditing + optimization recommendations for robot datasets. Diagnose data quality issues. Get actionable, confidence-graded suggestions. RDA is a diagnostic tool — it does NOT guarantee training success rate improvements.

RDA audits robot manipulation datasets (LeRobot format) for integrity, temporal consistency, motion quality, and distribution coverage. It then generates optimization recommendations calibrated to your target model architecture.

Features

• 13 quality metrics across 3 tiers: integrity, temporal, motion, and distribution
• rda recommend — Data optimization suggestions calibrated to your model type
  • Frame-wise models (MLP/BC): mild idle trimming suggestions
  • Temporal models (ACT/DP/Transformer): conservative "do not prune" guidance
  • All suggestions include confidence levels (HIGH / EXPERIMENTAL / NOT_RECOMMENDED)
• LeRobot v2.1 + v3.0 dual-format auto-detection
• Three-tier verdicts: PASS / REVIEW / EXCLUDE
• CLI-first design: JSON + text output, pipe-friendly
• Temporal sufficiency analysis: idle detection, active run distribution, valid window ratios

Tested on public data

RDA has been run over 12 local LeRobot-format datasets — 4,959 episodes — with the same default thresholds everywhere, zero per-dataset tuning. A few things it found:

• One xArm dataset is genuinely clean (767/800 episodes PASS, 20.8% median idle); another from the same robot platform runs 83.3% idle — task difficulty, not collection sloppiness
• A community SO-100 dataset has action spikes in 100% of episodes and 86.7% median idle
• A local LIBERO copy had 773 episodes missing from the downloaded data files. RDA 0.5.3 fixes a LeRobot v3.0 metadata/data-file mapping issue so missing local files are distinguished from episodes that are actually present

Full table, per-dataset numbers, and the five recurring patterns: docs/benchmark.md

Methodology and other experiment write-ups (i18n smoke test, server deploy verify, the spike/verdict bug regression pin, wheel leak guard): experiments/

Installation

pip install robot-data-audit

With LeRobot dependency (for .parquet dataset loading):

pip install robot-data-audit[lerobot]

With the Streamlit web UI:

pip install robot-data-audit[ui]

Quick Start

1. Audit a dataset

rda audit /path/to/lerobot/dataset

Runs all 13 metrics, prints a text summary. JSON report saved to <dataset>/rda_report.json.

2. Get optimization recommendations

# For frame-wise models (MLP, BC, etc.)
rda recommend /path/to/dataset --policy frame-wise

# For temporal models (ACT, Diffusion Policy, Transformer)
rda recommend /path/to/dataset --policy temporal

# JSON output for scripting
rda recommend /path/to/dataset --policy frame-wise --format json

# English output (default is Chinese)
rda recommend /path/to/dataset --policy temporal --lang en

What recommend tells you:

• Whether your dataset has excessive idle frames
• Whether trimming is advisable (and how aggressively)
• Model-specific warnings (e.g., "DO NOT prune for temporal models")
• Confidence levels and experimental caveats for every suggestion

Privacy (since v0.5.0): rda recommend computes all metrics locally and sends only aggregated statistics (<1KB, no raw episode data) to the RDA rules API (https://rda.niusu2026.cn) for evaluation. Results are cached locally for offline reuse. rda audit remains fully offline. For private/air-gapped deployments, point RDA_API_URL at your own server.

3. Launch the web UI

pip install robot-data-audit[ui]
rda ui

Opens a Streamlit dashboard at http://localhost:8501 — upload a dataset, run the audit with live progress, explore per-episode results, generate recommendations (via the same privacy-preserving API path), and export reports.

Bilingual UI (since v0.5.2): the entire dashboard — including backend recommendation copy and exported reports — switches cleanly between English and 中文 via the language selector in the sidebar. No mixed text.

4. JSON output & piping

# JSON to stdout
rda audit /path/to/dataset --format json

# Blind report for external sharing (paths are hashed)
rda audit /path/to/dataset --blind --format json

# Save report to custom path
rda audit /path/to/dataset -o /tmp/my_report.json

# Verbose mode with platform info
rda audit /path/to/dataset --platform so101 -v

5. 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"DHI: {result.quality['dhi']} / 100")

CLI Reference

rda audit

rda audit [OPTIONS] PATH

Option	Description
-o, --output FILE	Save JSON report (default: <path>/rda_report.json)
--format [json|text]	Output format (default: text)
--platform TEXT	Robot platform (e.g. so101, droid) for Tier 3 metrics
-v, --verbose	Verbose output
--blind	Redact identifying paths for externally shareable reports

rda recommend

rda recommend [OPTIONS] PATH

Option	Description
--policy [frame-wise|temporal]	Target model architecture type (required)
-o, --output FILE	Save JSON recommendation report
--format [json|text]	Output format (default: text)
--lang [zh|en]	Language of the recommendation text (default: zh)
-v, --verbose	Verbose output

rda example

Show example usage and sample dataset paths.

Exit Codes

Code	Meaning
0	Completed, no EXCLUDE verdicts
1	Error (invalid path, load failure, etc.)
2	Completed, at least one EXCLUDE verdict

Understanding Recommendations

RDA recommendations follow a conservative, evidence-graded approach:

Confidence	Meaning
HIGH	Well-supported by optimization experiments; low risk
EXPERIMENTAL	Directionally consistent but not yet validated for your setup
NOT_RECOMMENDED	Likely harmful for your model type; proceed with caution

Key principles:

• All suggestions are hypotheses, not guarantees
• Effects vary by task domain and model architecture
• Always validate on a held-out set before applying to training data
• Temporal models (ACT, DP) are generally more sensitive to data trimming

Metrics Overview

Tier	Metric	What it detects
L1	Timestamp monotonicity	Clock resets, duplicate timestamps
L1	Frame interval consistency	Jittery or irregular sampling
L1	Schema compliance	Missing/extra fields, type mismatches
L2	Temporal gap detection	Large time discontinuities
L2	Sensor synchronization	Cross-sensor timestamp drift
L2	Temporal sufficiency	Idle/active structure, valid window analysis
L3	Joint limit violations	Actuators driven beyond safe range
L3	Velocity spikes	Sudden implausible jumps
L3	Motion discontinuities	Non-smooth trajectory segments
L3	Idle frame detection	Stationary/paused segments
L4	Duration outliers	Episodes too short/long vs. cohort
L4	Spike count outliers	Episodes with unusual jerk profiles
L4	Effective motion ratio	Low-activity episodes

Project Structure

rda/
├── cli/          # Click CLI entry points
├── io/           # Data loading and schema definitions (LeRobot v2.1/v3.0)
├── metrics/      # 13 audit metric implementations
├── recommend/    # Optimization recommendation engine
├── audit/        # Dataset and episode-level audit orchestration
└── report/       # Report generation and summary

Development

git clone https://github.com/liesliy/rda.git
cd robot-data-audit
pip install -e ".[dev]"
pytest

Testing the gate itself

The suite in tests/ exists for a reason worth explaining. During the 0.4.x era, a bug let anomalous episodes walk away with a PASS badge: the behavior layer correctly computed action spikes and idle ratios, but the verdict aggregator ignored those signals entirely. The gate wasn't consuming its own evidence.

tests/test_negative_control.py pins that exact failure mode: metric results that pass every rule but contain known anomalies (150 spikes, a frozen arm) must come back REVIEW — and hard corruption must stay EXCLUDE (the gate can't fail open in either direction). The rest of the suite covers the i18n catalog (zh/en key alignment) and boundary behavior of the two core behavioral metrics.

CI runs the tests on every push and additionally verifies that the closed-source recommendation layer never leaks into the published wheel.

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