democlean 0.1.3

Score robot demonstrations by motion quality

democlean

Quality scoring for robot demonstration datasets.

The Problem

Robot learning datasets often contain bad demonstrations—jerky movements, hesitation, inconsistent timing. Training on these hurts policy performance. Manual review doesn't scale.

democlean automatically scores episodes by motion quality using mutual information (MI) between states and actions. Episodes with smooth, purposeful motion score high. Jerky, inconsistent episodes score low.

Install

pip install democlean

Quick Start

democlean analyze lerobot/pusht

Output:

Dataset lerobot/pusht
Episodes: 50 | Dims: 2→2

Distribution
  ████████████████████  High   30
  ██████████            Medium 15
  █████                 Low     5

Mean   2.55   (typical for human teleop)
Std    0.27

Flagged (lowest MI)
  ep  46  1.897
  ep   6  1.984

What MI Measures

MI quantifies how predictable actions are given states.

• High MI → smooth, purposeful motion
• Low MI → jerky, hesitant, inconsistent

This is useful because motion quality correlates with demonstration quality. But MI is not a direct measure of task success—it measures how the robot moved, not what it achieved.

Score Ranges

MI	Interpretation
>3.0	Very smooth
2.0–3.0	Typical human teleop
1.0–2.0	Moderate
<1.0	Noisy/random

Filtering Episodes

Keep top 80%:

democlean analyze lerobot/pusht --keep 0.8

Drop below threshold:

democlean analyze lerobot/pusht --min-mi 2.0

Save report:

democlean analyze lerobot/pusht --keep 0.8 -r report.json

Limitations

1. Length correlation: MI correlates with episode length (r≈0.8). Longer episodes score higher. Use --normalize-length to adjust.

2. Not task success: MI measures motion smoothness, not whether the task was completed. Use task-specific metrics for that.

3. Sample size: Works best with 50+ episodes. Small datasets may not show meaningful variation.

When to Use

Good fit:

• Human teleoperation data (quality varies by demonstrator)
• Single task, single robot
• 50+ episodes
• Quick triage before training

Not a good fit:

• Scripted simulation data (already uniform)
• Multi-task datasets (MI biases toward easy tasks)
• You need task success metrics

Python API

from democlean import DemoScorer

scorer = DemoScorer(k=3)
scores = scorer.score_dataset("lerobot/pusht")

# Filter
keep = scorer.filter_top_k(scores, percentile=80)
print(f"Keep episodes: {keep}")

CLI Reference

democlean analyze <dataset> [options]

Options:
  --keep R            Keep top R fraction (0-1)
  --top-k K           Keep top K episodes
  --min-mi T          Drop episodes below MI threshold
  --normalize-length  Adjust for episode length
  -k N                KSG neighbors (default: 3)
  --max-dim D         PCA reduce to D dimensions
  --ci                Bootstrap confidence intervals
  -r FILE             Save JSON report
  -q                  Quiet mode (JSON output only)
  --explain           Show interpretation guide

Credits

Inspired by DemInf (Hejna et al., RSS 2025).

Complements score_lerobot_episodes which catches visual issues—democlean catches motion issues.

License

MIT