arXiv Robotics — research abstracts· John L. Zhou, Yuxuan Dong, Jonathan C. Kao·· 2 days agoEditorial score42
Goal-Conditioned Policy Learning Faces Horizon-Dependent Performance Degradation
An Informational Curse of Horizon in Goal-Conditioned Policy Learning
Summary
Goal relabeling horizon affects policy generalization. Longer horizons reduce performance in goal-conditioned BC policies, which RL mitigates. Jacobian distillation improves long-horizon performance.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
Source：arXiv Robotics — research abstracts · arxiv.org