XIPER: Cross-domain Video Prediction Reward for RL

Arxiv CS.CV (computer vision)

XIPER (Cross-domain Video Prediction Reward) addresses the challenge of reinforcement learning from expert videos across visually distinct domains, where reward signals are absent and domain gaps exist. It uses a video prediction model to generate reward signals for imitation learning without requiring domain adaptation efforts. The method is validated on multiple cross-domain transfer tasks.

reinforcement-learningvideo-prediction

Arxiv CS.CV (computer vision)·Zhao YangXinrui ZuJacob E. KooiThomas DelliauxHe LiuShujian YuKevin Sebastian LuckVincent Fran\c{c}ois-Lavet

8 days ago