Remember Before You Explore: Persistent Shared Memory for Zero-Shot Object Navigation

Zhonghua Yi, Xiaohan Zhang, Maochun Luo, Mingxia Chen, Qihao Peng, Kaiwei Wang, Sheng Yang

Keywords: zero-shot object navigation, persistent memory

Abstract:

In practical applications like home robotics, a single agent over a long lifespan or a team of collaborating agents must perform a continuous stream of tasks in the same environment. However, conventional zero-shot object navigation (ZSON) paradigms, which reset memory after each task, are inherently non-collaborative and inefficient for such long-term operations as they lead to redundant exploration.
To bridge this gap, we introduce a Persistent Shared Memory (PSM) mechanism that allows single or multi-agent systems to accumulate and reuse semantic knowledge across tasks and agents. Our approach builds an Temporally Consistent Semantic Map (TCSM), decoupling scene memory from task-specific information and maintaining semantic consistency via weighted confidence updates. On top of this memory, we design a beyond-line-of-sight (BLOS) navigation strategy that propagates stored semantics into nearby navigable areas and performs line-of-sight checks for waypoint selection, enabling reasoning about objects that are currently occluded or distant. Experiments on public benchmarks, including HM3D and MP3D, have shown that our framework avoids redundant scene re-exploration and achieves state-of-the-art performance. Our code will be made available upon acceptance.

Primary Area: applications to robotics, autonomy, planning

Submission Number: 1900