FLYNN: Robot navigation net modeled on fruit fly brain beats sensory loss
What if a robot could navigate blindfolded, just like a fruit fly?
Deep learning models excel in controlled settings but crumble under sensory deprivation or unfamiliar environments. Inspired by biology's resilience, researchers Benquan Wang and Jingdao Chen built FLYNN (FLY connectome Neural Network), a recurrent neural network whose wiring is directly derived from the complete synaptic-resolution brain connectome of the fruit fly Drosophila melanogaster. They trained FLYNN to perform vision-based navigation in the MuJoCo physics simulator, achieving performance comparable to modern hand-crafted networks with similar parameter counts.
FLYNN's key breakthrough is its extraordinary robustness. It maintained navigation capability even under total vision loss—no camera input at all—without any additional training. In contrast, conventional hand-crafted networks failed catastrophically, even when explicitly trained with camera dropout. Principal Component Analysis of FLYNN's internal states suggests a high degree of representational modularity, which may explain its tolerance to sensory loss and out-of-distribution data. This work opens a new direction for designing resilient artificial agents by directly mimicking biological brain topology.
- FLYNN’s architecture is copied from the fruit fly’s synaptic-resolution brain connectome.
- Matches performance of hand-crafted networks of similar size on vision-based navigation in MuJoCo.
- Remains functional under total vision loss without retraining, while conventional nets fail even with dropout training.
Why It Matters
Biologically-inspired network topology could make robots far more reliable when sensors fail or environments shift.