PacMouse: Mice solve hidden abstract graph traversal problems in freely-moving virtual reality

Biological agents such as mice solve navigation problems in uncertain and partially observable environments. We introduce PacMouse, a rodent behavioral paradigm that elicits exploration of latent graphs while minimizing confounding internal and external cues that could trivialize state estimation. The task is implemented through a freely-moving virtual reality mouse experiment paradigm, in which all decisions occur at a single physical junction with motorized gates controlling access to four radial arms. Proximity sensors in the radial arms trigger virtual transitions between adjacent nodes in a hidden latent graph. This design precludes reliance on path integration or landmark signals, enabling direct comparison between animal behavior and simulated agents. We show that mice frequently outperform a range of baseline policies. The framework provides open-source tools for linking mouse behavioral data to standard RL implementations, establishing a common task in which to benchmark head-to-head biological and artificial decision making.