FLYNN: Fly-brain RNN for robust robot navigation in MuJoCo

Benquan Wang and Jingdao Chen submitted FLYNN (arXiv:2607.00025) on 21 Jun 2026: a recurrent neural network whose architecture is directly derived from the synaptic-resolution connectome of the fruit fly Drosophila melanogaster, trained for vision-based robot navigation in MuJoCo.

What is FLYNN and how was it constructed?

FLYNN is an RNN whose wiring follows the fly brain connectome at synaptic resolution, rather than using a hand-designed architecture. The authors mapped the synaptic connections from the Drosophila connectome into a recurrent network topology, then trained that network to perform vision-based navigation tasks inside the MuJoCo simulator. The submission appears on arXiv as arXiv:2607.00025, uploaded 21 Jun 2026 (submission file 5,145 KB).

The paper positions the connectome-derived topology as the central design choice. Training occurred in a simulated navigation environment, using visual inputs and standard reinforcement-style objectives to learn control policies. The authors report studying the network's internal dynamics with Principal Component Analysis to probe representational structure.

How does FLYNN perform compared with hand-crafted networks?

FLYNN achieves performance comparable to modern hand-crafted networks that have similar parameter counts, while showing stronger out-of-distribution robustness and resilience to sensory loss. The authors state that FLYNN "remained functional even under total vision loss" whereas hand-crafted networks largely failed, including those that had been trained with camera dropout meant to simulate vision degradation.

The paper highlights two outcome classes: nominal task performance and robustness under degraded or OOD conditions. On nominal navigation in MuJoCo, FLYNN matched hand-crafted baselines of similar size. Under OOD inputs and complete vision removal, FLYNN continued to operate effectively without further training; the hand-crafted controllers did not, even after being trained with camera dropout. The internal PCA analysis suggests a higher degree of representational modularity in FLYNN, which the authors hypothesize may relate to its robustness.

Why does the fly-brain topology change robustness?

The paper links FLYNN's robustness to its connectome-derived wiring and to the network's internal representational modularity, as revealed by PCA. That analysis indicates FLYNN organizes internal activity into more modular components than the hand-crafted networks the authors compared it against. The authors propose that this modularity could underlie tolerance to sensor loss and OOD inputs, letting different subcircuits maintain function when visual streams fail.

This is an explicit attempt to translate a biological wiring diagram into an engineering design for resilient agents. The work does not claim superiority across every metric; it reports parity in performance under normal conditions and clear gains in robustness metrics during vision degradation and OOD exposures.

What to watch

Look for follow-up evaluations that publish exact benchmark numbers and training details for both FLYNN and the hand-crafted baselines, and for real-robot trials beyond MuJoCo. Additional analyses that connect specific connectome motifs to the PCA-derived modular components would test the authors' hypothesis about topology causing robustness. The arXiv submission identifier is arXiv:2607.00025; it was submitted on 21 Jun 2026.