Goal-Conditioned Reinforcement Learning with Virtual Experiences
Keywords: Subgoal Planning, Curriculum Learning, Imitating Learning
TL;DR: We train goal-conditioned policies with virtual experiences by self-supervised subgoal planning.
Abstract: Goal-conditioned reinforcement learning often employs a technique known as Hindsight Experience Replay (HER) for data augmentation by relabeling goals. However, HER limits goal relabeling to a single trajectory, which hinders the utilization of experiences from diverse trajectories. To address this issue, we present a curriculum learning method to construct virtual experiences, incorporating actual state transitions and virtual goals selected from the replay buffer. Considering that virtual experiences may contain a lot of noise, we also propose a self-supervised subgoal planning method that guides the learning of virtual experiences by imitating the subgoal-conditioned policy. Our intuition is that achieving a virtual goal may be challenging for the goal-conditioned policy, whereas simplified subgoals can provide effective guidance. We empirically show that the virtual experiences from diverse historical trajectories significantly boost the sample-efficiency compared to the existing goal-conditioned reinforcement learning and hierarchical reinforcement learning methods, even enabling the agent to learn tasks it has never experienced.
Primary Area: reinforcement learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 5840
Loading