Learning from the Best: Smoothness-Driven Metrics for Data Quality in Imitation Learning
Researchers cut training data by 83% while improving performance by 16%.
Researchers have introduced RINSE (Ranking and INdexing Smooth Examples), a lightweight framework that scores robot demonstration data quality based on trajectory smoothness, eliminating the need for costly policy training or manual annotation. The method leverages two complementary metrics: Spectral Arc Length (SAL), which measures frequency-domain regularity, and Trajectory-Envelope Distance (TED), which captures contact-aware geometric deviation. Grounded in motor control theory, RINSE identifies smoother trajectories as higher quality, reducing conditional action variance and downstream errors in behavioral cloning.
On RoboMimic benchmarks, SAL filtering achieved 16% higher success rates while using just one-sixth of the original data. In real-world manipulation tasks, TED filtering delivered a 20% improvement with only half the data. As a retrieval-stage filter within STRAP on LIBERO-10, RINSE re-ranking improved mean success by 5.6%. When used as soft weights in Re-Mix domain reweighting, RINSE scores produced allocations highly correlated with learned Re-Mix allocations (Spearman ρ ≥ 0.89). These results demonstrate smoothness as a robust quality signal across filtering, retrieval, and reweighting settings, particularly valuable in noisy or heterogeneous data regimes.
- RINSE uses Spectral Arc Length (SAL) and Trajectory-Envelope Distance (TED) to score trajectory smoothness without policy training
- SAL filtering on RoboMimic achieved 16% higher success using only one-sixth of demonstration data
- TED filtering on real-world manipulation tasks yielded 20% improvement with half the data
Why It Matters
RINSE makes robot learning more data-efficient by automatically filtering low-quality demonstrations, reducing training costs and improving real-world performance.