Brain-inspired robot navigation using reinforcement learning with hippocampal and striatal contributions

This study investigates the dual-system theory of navigation by developing brain-inspired "digital behavioral twins" of a mobile robot. We implemented two reinforcement learning (RL) agents in a simulated double H-maze: a Deep Q-Network (DQN) to emulate the striatal model-free (egocentric) system, and a Proximal Policy Optimization (PPO) agent for the hippocampal model-based (allocentric) system. We systematically tuned hyperparameters analogous to neuromodulators (e.g., discount factor, exploration) and evaluated navigation strategies using a probe trial paradigm where the robot's starting position was changed after training. Our results show that the DQN agent defaults to a rigid, habitual strategy, failing the probe trial. In contrast, the PPO agent develops a flexible, allocentric cognitive map, successfully navigating from the new starting point. Crucially, specific hyperparameter tuning reproduced behaviors analogous to clinical profiles: a myopic, hyper-exploratory DQN mirrored traits of ADHD, while a PPO with a less robust map showed cue-dependency akin to ASD. This work establishes a powerful computational testbed for linking RL mechanisms to the neural underpinnings of navigation and offers a novel yet minimal framework for investigating navigational deficits in neuropsychiatric disorders.