FAVEN: Fast Audio-Visual Embodied Navigation in 3D Environments

Shentong Mo, Kilian Q Weinberger, Marco Pavone, Boyi Li

Keywords: audio-visual learning, audio-visual navigation

TL;DR: In this paper, we introduce FavEN, a novel transformer and mamba architecture that combines audio and visual data into early fusion tokens for embodied navigation..

Abstract: Achieving fast audio-visual embodied navigation in 3D environments is still a challenging problem. Existing methods typically rely on separate audio and visual data processing merged in late stages, leading to suboptimal path planning and increased time to locate targets. In this paper, we introduce FavEN, a novel transformer and mamba architecture that combines audio and visual data into $\textit{early fusion}$ tokens. These tokens are passed through the entire network from the initial layer on and cross-attend to both data modalities. The effect of our early fusion approach is that the network can correlate information from the two data modalities from the get-go, which vastly improves its downstream navigation performance. We demonstrate this empirically through experimental results on the Replica and Matterport3D benchmarks. Furthermore, for the first time, we demonstrate the effectiveness of early fusion in improving the path search speed of audio-visual embodied navigation systems in real-world settings. Across various benchmarks, in comparison to previous approaches, FavEN reduces the search time by 93.6\% and improves the SPL metrics by 10.4 and 6.5 on heard and unheard sounds.

Primary Area: applications to computer vision, audio, language, and other modalities

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: 4803