FLYNN: Fly Brain Topology Inspires Robust Robot Navigation Neural Network

While deep learning models excel in complex tasks, they often exhibit fragility when confronted with unfamiliar environments or sensory deprivation. Biological systems, in contrast, demonstrate remarkable resilience to such challenges. This research addresses the vulnerability of artificial neural networks by developing a recurrent neural network (RNN) named FLYNN. FLYNN's architecture is directly inspired by the synaptic-resolution brain connectome of the fruit fly, Drosophila melanogaster. The study successfully trained FLYNN to perform vision-based navigation in a MuJoCo environment, achieving performance comparable to modern, hand-crafted networks with similar parameter counts. Crucially, FLYNN exhibited superior resistance to out-of-distribution data and maintained functionality even under significant sensory loss, such as total vision loss, without requiring further training. This contrasts sharply with hand-crafted networks, which largely failed even when specifically trained with camera dropout. Principal Component Analysis of FLYNN's internal states suggests a high degree of representational modularity, which is hypothesized to be a key factor in its observed robustness. This work opens new avenues for designing resilient artificial agents by drawing inspiration from biological brain topologies.