On the Benefits of Pixel-Based Hierarchical Policies for Task Generalization
Reinforcement learning practitioners often avoid hierarchical policies, especially in image-based observation spaces. Typically, the single-task performance improvement over flat-policy counterparts…