Geometric Entropy: When Trajectory Diversity Helps and Hurts in Imitation Learning
Overview
This research paper introduces Geometric Entropy (H_G), a metric that quantifies the intrinsic diversity of transit trajectories in imitation learning by normalizing for extrinsic variations like goal pose. The study finds an inverted-U relationship between H_G and task success: while increasing geometric diversity improves robustness in low-diversity regimes, it degrades performance when diversity induces strategy ambiguity. Furthermore, the optimal entropy level shifts toward lower values as task mastery increases through more data, easier tasks, or stronger priors. The authors demonstrate that H_G can be used for fast pre-training auditing of demonstration datasets and provide guidelines for calibrating demonstrations to the learnable regime.
Generated from attributed reports · Updated 17 minutes ago
Event evidence and corrections
0 attributed source owners. Ownership does not establish independent confirmation. Quantities are reported separately and are never added together.
No current evidence-backed claims. Missing information remains not reported.
Report timeline
Follow attributed reports and material updates.
- arXiv Robotics — research abstractsSignalGeometric Entropy: When Trajectory Diversity Helps and Hurts in Imitation Learning
This research paper introduces Geometric Entropy (H_G), a metric that quantifies the intrinsic diversity of transit trajectories in imitation learning by normalizing for extrinsic variations like goal pose. The study finds an inverted-U relationship between H_G and task success: while increasing geometric diversity improves robustness in low-diversity regimes, it degrades performance when diversity induces strategy ambiguity. Furthermore, the optimal entropy level shifts toward lower values as task mastery increases through more data, easier tasks, or stronger priors. The authors demonstrate that H_G can be used for fast pre-training auditing of demonstration datasets and provide guidelines for calibrating demonstrations to the learnable regime.
Event coverage history
There is not enough continuous observation data to show a trend.
Timezone · UTC
Article dates follow your selected timezone. Briefing editions use Hong Kong time (UTC+8).