Bridging Neuroscience and AI: A Grid-Cell-Inspired Deep Learning Model for Spatial Navigation

Abstract

Spatial navigation is a fundamental cognitive function in both biological and artificial systems. Neuroscientific studies have shown that place and grid cells within the hippocampal-entorhinal complex play a critical role in encoding spatial environments. Inspired by these mechanisms, this study proposes and evaluates an enhanced Spatial Transformer with Learned Grid-Like Coding (e-STL) model for artificial spatial navigation tasks. Using publicly available maze-based simulation environments, we compared the performance of the e-STL model to established deep reinforcement learning models, including Deep Q-Networks (DQN) and Asynchronous Advantage Actor-Critic (A3C). The e-STL model demonstrated superior performance in success rate, path efficiency, and learning speed across multiple navigation tasks. Our findings align with existing literature on grid cell modeling in AI, such as work by Banino et al. (2018), and further demonstrate that incorporating biologically inspired spatial priors significantly enhances navigation capabilities in artificial agents. These results highlight the potential of interdisciplinary approaches that integrate neuroscience insights into machine learning systems.