Connectome-derived structural support identifies a compact scaffold for robust embodied control in BIOSEED
Year: 2026
Type: preprint
Abstract: Biological nervous systems combine sparse connectivity with structured recurrent organization, but it remains unclear whether such structure can serve as a functional prior for artificial embodied control. Here we use BIOSEED, a bounded food/toxin foraging framework, to test whether connectome-derived structural support has behaviorally consequential value under controlled perturbations and learned-controller controls. A staged audit demotes historical shuffled/random controls to fixed-support perturbations, shows that corrected density-matched binary rewiring collapses performance, verifies transfer across implemented environmental shifts, and shows that causal lesions degrade the TRUE structure. A compact top-in 10% scaffold is statistically non-inferior to the full topology in locked validation and transfer, so 85% of the connectome-derived edges can be removed at no behavioral cost. Matched non-biological reductions of equal sparsity reach at most about two thirds of the full-topology score, and connectome-derived selections based on out-degree or synaptic weight also fail, so the sufficiency is specific to incoming-connectivity structure rather than to sparsity or connectome provenance alone. Learned MLP/RNN/GRU controls show that the task is learnable and that intake can be increased, including by a compact GRU approaching the conservative band; however, they do not reduce the scaffold result to a trivial optimization artifact. Live and rover-style demonstrations provide secondary closed-loop diagnostics. BIOSEED therefore identifies a connectome-derived structural prior for conservative embodied control within the tested framework.
Source: Research Square
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225798561