本文由 AI 分析生成
建立時間： 2026-08-18 來源： https://arxiv.org/html/2605.01544v1
Summary
This University of New Hampshire paper proposes Power Spectral Density (PSD) of end-effector trajectories as a training-free, sub-10-second quality metric for ranking imitation learning demonstrations. The core insight: smooth expert demonstrations have low spectral power, while erratic corrective motions inject high-frequency components detectable without policy rollouts. Filtering the bottom 50% by PSD improves success rates from 31% to 86% on lay user datasets and is ~1800× faster than learning-based competitors.
本文以末端執行器軌跡的功率頻譜密度（PSD）作為免訓練、約 10 秒完成的示範品質排序指標。核心洞見：平滑的專家示範頻譜功率低，抖動修正動作引入高頻成分；過濾最差 50% 示範後，非專家資料集成功率從 31% 提升至 86%，速度比學習型方法快約 1800 倍。
Prerequisites
- Imitation Learning / Behavior Cloning — the paper’s setting: policies trained on demonstration datasets; data quality directly affects learned policy performance.
- Discrete Fourier Transform / Power Spectral Density — the quality metric is computed in the frequency domain; reader needs to understand that high-frequency power corresponds to rapid, irregular changes in a signal.
Core Idea
The observation driving the method: human corrective motions during teleoperation produce characteristic high-frequency jitter in the trajectory signal, while smooth, purposeful expert execution concentrates energy in low frequencies. This is detectable without knowing what task success looks like — PSD is a pure signal-quality measure, not a task-completion measure. Because it requires only trajectory coordinates and a DFT, the entire ranking step completes in seconds, making it practical to apply before every training run as a preprocessing step. The method is deliberately agnostic to policy architecture and task semantics, which is both its strength (universally applicable) and its weakness (cannot detect smooth failures).
Results
| Task / Benchmark | Unfiltered | PSD-Filtered (50%) | Competitors |
|---|---|---|---|
| Robomimic-Can | 91% | 100% | DemInf / CUPID: slower, marginal gain |
| Layman2-Can (lay users) | 31% | 86% | Outperforms both |
| Older adults (π0.5 fine-tune) | 15% | 38% | — |
| Compute time | — | ~10 sec | ~300–360 min |
Limitations
- Author-stated: Task-outcome agnostic (smooth but failed demos may pass the filter); kinematic focus misses semantic quality; untested beyond manipulation; image-based policy applicability limited.
- Unstated: The 50% filtering threshold is empirically chosen, not principled — a dataset with many high-quality lay-user demos could over-filter good data. Also, the method is redundant for datasets already collected from expert operators (where frequency noise is low across the board).
Reproducibility
- Code: Not mentioned
- Datasets: Robomimic-Can, Layman2-Can, user study with 5 older adults + π0.5 model
- Compute: Trivial — DFT only, no GPU needed
Insights
PSD filtering is a natural preprocessing complement to DQAF (the other data-quality paper in this vault): DQAF provides rich post-episode feedback to the operator during collection; PSD filtering provides silent post-hoc curation before training. They operate at different points in the pipeline and can stack: collect with DQAF feedback → filter residual noise with PSD → train.
The older-adult user study angle is notable: this isn’t just about teleoperation experts — it targets the democratization use case where end-users with varying motor abilities collect data for their own assistive robots. PSD filtering makes IL more robust to the natural motion variability of non-expert populations.
The gap between Layman2-Can results (31% → 86%, +55pp) vs. Robomimic (91% → 100%, +9pp) reveals the metric’s value is primarily in correcting lay-user data, not marginal gains on already-decent datasets.
Connections
- closing-the-loop-teleoperation-dqaf — DQAF: online operator feedback during collection; PSD: offline post-hoc filtering; complementary in the data pipeline
- 2026-05-07-human-demo-methods-il-taxonomy — taxonomy of demonstration collection methods
- robot-manipulation-unified-survey-2510-10903 — broader manipulation + data survey context
- robotic-teleoperation-manipulation — teleoperation as the data source these methods filter
Raw Excerpt
“PSD ranking: ~10 seconds per dataset. Competitors (DemInf/CUPID): ~300–360 minutes.”