Memory-MambaNav: Enhancing object-goal navigation through integration of spatial-temporal scanning with state space models

Open Webpage

Leyuan Sun, Yusuke Yoshiyasu

Published: 01 Jan 2025, Last Modified: 15 May 2025Image Vis. Comput. 2025EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Highlights•Introduce Memory-MambaNav, the first Mamba-based architecture for ObjectNav.•Propose spatial difference SSMs and temporal serialized SSMs for efficient memory use.•Reduce GPU memory usage by 67.5% for long-horizon memory processing vs. Transformers.•Develop memory-aggregated egocentric obstacle-awareness tokens and dense rewards.•Achieve competitive results on AI2-Thor, excelling in long-horizon tasks with extended memory.