Learning Fused State Representations for Control from Multi-View Observations
Keywords: multi-view learning, reinforcement learning
TL;DR: We proposed a method to learn fused state representations for multi-view RL.
Abstract: In visual control tasks, leveraging observations from multiple views enables Reinforcement Learning (RL) agents to perceive the environment more effectively. However, while multi-view observations enrich decision-making information, they also increase the dimension of observation space and introduce more redundant information. Thus, how to learn compact and task-relevant representations from multi-view observations for downstream RL tasks remains a challenge. In this paper, we propose a Multi-view Fusion State for Control (MFSC), which integrates a self-attention mechanism with bisimulation metric learning to fuse task-relevant representations from multi-view observations. To foster more compact fused representations, we also incorporate a mask-based latent reconstruction auxiliary task to learn cross-view information. Additionly, this mechanism of mask and reconstruction can enpower the model with the ability to handle missing views by learning an additional mask tokens. We conducted extensive experiments on the Meta-World and Pybullet benchmarks, and the results demonstrate that our proposed method outperforms other multi-view RL algorithms and effectively aggregates task-relevant details from multi-view observations, coordinating attention across different views.
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.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
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: 14220
Loading