Brain-Inspired Memory Architecture Improves Long-Term Robotic Object Navigation Benchmarks

A research team has developed Human-like Memory Navigation (HM-Nav), a brain-inspired architecture designed to bridge the cognitive gap in long-term object navigation for embodied robots. Operating under egocentric perception and partial observability, embodied robotic agents frequently encounter difficulties when tracking dynamic spatial relations and maintaining long-term consistency, which often leads to inefficient, repetitive search patterns. By integrating transient sensory inputs with an evolving internal model, the HM-Nav framework allows robotic agents to maintain cohesive environmental awareness over extended periods, enabling more resilient and efficient search behaviors.

The architecture is built upon three synergistic pillars designed to replicate the effortless object-finding abilities of humans, who seamlessly integrate perception, memory, and contextual inference even within cluttered or unseen environments. The first pillar is a perception module featuring multi-view fusion, which reconciles inconsistencies across varying viewpoints into unified spatial and visual representations.

The second and third pillars govern the system’s memory formation and spatial decision-making processes. For its memory component, HM-Nav deploys an adaptive dynamic knowledge graph that continuously accumulates semantic-spatial associations over extended timescales, preserving critical contextual relationships as the robot explores. Supporting this memory structure is an inference pillar powered by an experience-driven trajectory optimization mechanism. This mechanism actively learns from past navigation failures to suppress cyclic and suboptimal search paths, ensuring that the robot does not repeat erroneous exploration routines.

In validation experiments conducted across both simulations and physical real-world trials, HM-Nav demonstrated superior navigation performance and achieved robust sim-to-real transfer, significantly outperforming existing benchmark models. The researchers, reporting in Communications Engineering, noted that emulating human-like memory structures is essential for achieving resilient, long-term autonomy in complex, open-ended environments.

Subject of Research: Brain-inspired memory architecture for robotic object navigation

Article Title: Human-like memory empowers embodied robots for long-term object navigation

Article References: Zhang, Y., Song, R., Ren, H., Yan, H., Liu, J., Hua, C., & Zhang, C.-H. (2026). Human-like memory empowers embodied robots for long-term object navigation. Communications Engineering. https://doi.org/10.1038/s44172-026-00767-5

Image Credits: AI Generated

DOI: 10.1038/s44172-026-00767-5

Keywords: embodied robotics, object navigation, brain-inspired architecture, HM-Nav, human-like memory, dynamic knowledge graph, multi-view fusion, trajectory optimization, egocentric perception, partial observability, sim-to-real transfer, semantic-spatial associations, contextual inference, spatial relations, autonomous navigation, cyclic search suppression