A deep reinforcement learning account of competition and integration of visual and goal vector signals for spatial navigation

Description

Sandhiya Vijayabaskaran and Sen Cheng model in 2026 how an animal might combine what it sees with an internal sense of where its goal lies. Navigation in uncertain surroundings demands exactly that juggling, and engineers usually assume more signals always help. Their deep reinforcement learning agent learned to navigate using both a visual signal and a goal-vector signal, each with adjustable noise.

How noisy each signal was dictated how much the agent leaned on each source and how its internal spatial maps formed. Keeping the ability to use either signal alone, not just the blend, proved crucial when conditions changed. Counter-intuitively, dropping one signal sometimes improved performance when the remaining signal was trustworthy, though the agent became fragile. The work warns biologists that redundant sensing in animals is subtler than sensor fusion in machines.

Catalog record

Source: https://europepmc.org/article/MED/PMC13408606