An Efficient Metric for Data Quality Measurement in Imitation Learning
Authors: Noushad Sojib, Momotaz Begum (University of New Hampshire)
arXiv: 2605.01544v1 [cs.RO] — May 2, 2026
Abstract
The paper proposes using Power Spectral Density (PSD) analysis of robot end-effector trajectories as a fast, training-free quality ranking metric for imitation learning demonstrations. Rather than relying on expensive policy rollouts or expert labeling, PSD captures frequency-domain characteristics: smooth, high-quality demonstrations have low spectral power, while erratic corrective motions introduce high-frequency components. Demonstrations are ranked by ascending PSD and the worst-quality portion is filtered before policy training.
Core Method
- Extract end-effector 3D Cartesian coordinates from each demonstration
- Compute Discrete Fourier Transform per dimension
- Calculate total spectral power as quality score (lower = smoother = better)
- Rank demonstrations by PSD; discard the lowest-quality fraction
No policy training, environment interaction, or expert labeling required.
Key Results
Offline Policy Learning (50% data filtering)
| Dataset | Unfiltered | PSD-Filtered | Delta |
|---|---|---|---|
| Robomimic-Can | 91% | 100% | +9pp |
| Layman2-Can | 31% | 86% | +55pp |
Outperforms competing methods DemInf and CUPID on both datasets.
Computational Speed
| Method | Time per dataset |
|---|---|
| PSD ranking | ~10 seconds |
| DemInf / CUPID | ~300–360 minutes |
~1800× faster than learning-based approaches.
Real-World User Study (older adults fine-tuning π0.5)
- Unfiltered: 15% mean success rate
- PSD-filtered: 38% mean success rate (+23pp)
Limitations
- Task-outcome agnostic: ranks by motion smoothness, not task success; may retain smooth but failed demos
- Kinematic focus: doesn’t capture semantic quality (e.g., wrong object grasped smoothly)
- Limited to kinematic trajectories; applicability to image-based policies is limited
- Assumes erratic motion is the primary degradation source — may not hold universally