A Drosophila connectome prior enables adaptable embodied control

Siyu Yu^1,2,3Kanglei Zhou^1,3Hongwei Yan^3,4Xinyi Wang¹Sen Song^2,3Bo Hong^2,3Liyuan Wang^1,3,*

Tsinghua University

¹ Department of Psychological and Cognitive Sciences² School of Biomedical Engineering³ IDG/McGovern Institute for Brain Research⁴ School of Life Sciences

GitHub Videos

Adult fly brain Anatomical rendering of the adult fly brain connectome, extracted from Figure 1.

138,584neurons in the recurrent core

Connectome-controlled digital fly

The wiring of a fly brain becomes a reusable prior for movement.
We keep the connectome fixed and learn the interfaces that connect it to a body.

Biological structure. Learned interfaces.

Measured synaptic connectivity defines a recurrent network with 3.73 million directed connections. Sensory and motor interfaces learn to use this structure for closed-loop control.

Each simulated body learns its own interfaces around the fixed connectome. In the digital fly, the policy modulates a central pattern generator that drives 42 leg joints.

Original Figure 6b: a learned sensory interface feeds the fixed connectome, followed by a motor interface. State and action close the loop through biped, tripod, quadruped and hexapod bodies, with an additional local-feedback branch.

Terrain, turning and speed.

The fly must negotiate an obstacle, change direction or follow a changing speed command. The connectome prior improves obstacle crossing and speed tracking, and produces smoother turning commands.

Paired illustrative trial. Playback and episode outcomes are labelled in the video; population results appear in Figure 2.

01 / 03

Crossing higher obstacles

The connectome controller maintains higher all-feet crossing success as block height increases. Both controllers reliably traverse flat ground.

0.20–0.525mm block heights evaluated

20 paired training seeds · 64 shared test phases per height

Tracing control through the network.

Neural sensitivity, activity and pathway interventions reveal how the controller uses its biological structure. Mechanosensory and ascending pathways contribute to task-dependent motor outputs.

Turning fly illustration from the manuscript's framework figure

Ascending
neuronsDescending
neurons

Activity during turning

Ascending neurons show 1.46× activity enrichment relative to the input-population average, with enrichment in all 24 training seeds.

A pathway with a behavioral effect

Cutting ascending-to-descending connections increases heading error by 3.52°, compared with 1.16° for matched control cuts.

From walking to navigation.

The same fly-connectome core is adapted to the Ant quadruped. Across seven controller architectures, it achieves the highest mean success in the Medium and Large mazes.

Recorded rollouts with a shared start and goal. Each controller is identified in the video.

Medium-maze success

Adult connectome64.2%

MLP44.1%

Success reaches 50.6% vs 20.4% on routes requiring six or seven turns.

20 paired training seeds · 100 matched evaluation resets per seed

Ant locomotion on flat ground

The connectome controller reaches a mean forward velocity of 1.94 m/s, compared with 0.68 m/s for the MLP. Signed path efficiency improves from 0.701 to 0.885.

This stronger forward drive is accompanied by more episode failures: 38.6% vs 7.0%, across 20 paired training seeds.

The same wiring. Different bodies.

Body-specific interfaces connect the fixed fly brain to a biped, a tripod, a quadruped and a hexapod. Evaluation spans nominal control, pushes, reduced friction, delayed actions, sensor loss and extended trials.

Performance across seven conditions

Mean normalized-performance gain over the matched MLP. Cassie is evaluated by the fraction of each trial spent standing; the other bodies by walking success.

Cassie+0.271

Radial tripod+0.151

Unitree Go1+0.091

Hexapod+0.167

20 paired training seeds and 20 matched evaluation states per condition. Videos show selected paired trials.

Abstract

Embodied intelligence requires effective control as environments, movement demands and body morphologies change. Biological nervous systems support such adaptability through highly structured connectivity, yet whether evolved neural wiring itself confers a computational advantage remains unclear. Here we use the adult Drosophila brain connectome as a fixed recurrent structural prior for closed-loop embodied control. In a neuromechanical digital fly, connectome-based and multilayer-perceptron controllers perform comparably during nominal locomotion but diverge as control demands change. Across vertical terrain negotiation, lateral direction turning and longitudinal speed modulation, the connectome exhibits a consistent advantage, improving obstacle negotiation and speed tracking while supporting smoother directional reconfiguration. Mechanistic analyses reveal spontaneous recruitment of behaviourally relevant sensorimotor pathways and distinct local policy geometry associated with robust adaptation. The same connectome prior improves locomotion after transfer to a quadrupedal body and outperforms representative artificial controllers in maze navigation, with its advantage widening as maze complexity increases, while remaining effective across two-, three-, four- and six-legged embodiments. These results identify adaptability as a functional advantage of connectome structure and establish evolved neural wiring as a transferable prior for adaptive embodied control.