Leveraging Skills from Unlabeled Prior Data for Efficient Online Exploration
Keywords: reinforcement learning, exploration, skills, unsupervised pretraining, offline to online rl
Abstract: Unsupervised pretraining has been transformative in many supervised domains. However, applying such ideas to reinforcement learning (RL) presents a unique challenge in that fine-tuning does not involve mimicking task-specific data, but rather exploring and locating the solution through iterative self-improvement. In this work, we showcase how unlabeled prior trajectory data can be leveraged to learn efficient exploration strategies. The key insight is to use unlabelled trajectories twice, 1) to extract a set of low-level skills offline, and 2) as additional data for a high-level policy that composes these skills to explore. We utilize a simple strategy of learning an optimistic reward model from online samples, and relabeling past trajectories into high-level, task-relevant examples. We instantiate these insights as SUPE (Skills from Unlabeled Prior data for Exploration), and empirically show that SUPE reliably outperforms prior strategies, successfully solving a suite of long-horizon, sparse-reward tasks.
Supplementary Material: zip
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: 12140
Loading