Paper
Code (Coming soon)
Structure of the fly-connectomic Graph Model.
(a) Aggregated synapse graph.
(b) Force-directed graph layout.
In this work, we introduce the Fly-connectomic Graph Model (FlyGM), a reinforcement learning controller whose architecture is instantiated directly from a complete neuronal wiring diagram of an adult Drosophila.
We show that FlyGM can achieve diverse embodied locomotion behaviors, demonstrating that static brain connectomes can be transformed to instantiate effective neural policy for embodied learning of movement control.
Overview of the FlyGM enabled whole-body locomotion control framework.
Inductive Bias from Connectome Topology
We investigate whether the biological connectome offers a structural advantage beyond its parameter count.
By comparing our model against (1) an Erdős–Rényi random graph (2) a degree-preserving rewiring graph and (3) a standard MLP architecture. By comparing the connectome against these non-connectome models, we could isolate the contribution of the brain wiring to the efficiency and stability of locomotion control.
In the figure on the right side, we evaluate (1) Total training loss, reflecting sample efficiency; Mean Squared Error (MSE) of the action (2) means and (3) standard deviations relative to the expert trajectories, reflecting the fidelity of trajectory tracking.
The results demonstrate that the connectome-based architecture consistently outperforms all other graph topologies across all metrics. Notably, the connectome exhibits a markedly superior sample efficiency.
Learning efficiency and metrics across different graph topologies and architectures.
We also evaluated trajectory tracking across various topologies in a path-following task. The biological connectome-based architecture achieves near-optimal precision in both position (Pos Err) and angular (Angle Err) accuracy. While rewiring graph remains competitive, it is overall inferior to the connectome, and ER-Random Graph exhibits the highest errors. These results confirm that evolved brain wiring provides a critical structural inductive bias for high-fidelity embodied control.
Locomotion Demonstration
1. Gait Initation ( X0.08 Speed )
The video shows the representation intensity dynamics (left) and simulated fly (right) during locomotion onset (0-100 ms), before the first complete gait cycle. It illustrates the agent’s transition from rest to stepping, where irregular, asymmetric leg movements gradually converge into a coordinated pattern.
Eye camera views are captured by MuJoCo sensors, producing two 32×32 RGB images with a 150° field of view and an inter-ommatidial angle of 4.6° to approximate Drosophila vision. These are downsampled to 16×16 grayscale images and combined with other proprioceptive and exteroceptive inputs to form the input of FlyGM.
2. Straight-Line Walking ( X0.08 Speed )
The video shows the representation intensity dynamics (left) and simulated fly (right) during the task of straight-line walking (0-400 ms) at a velocity of 3 cm/s. The model maintains stable stepping sequences with tripod coordination emerging naturally from the FlyGM.
3. Turning ( X0.08 Speed )
The video shows the representation intensity dynamics (left) and the simulated fly (right) during a high-speed left turn (0–400 ms) at 3 cm/s and 10 rad/s. Over the course of the trajectory, stride lengths on the turning side decrease while those on the contralateral side increase, producing a smooth curve. This asymmetry demonstrates that FlyGM can generalize beyond straight walking to directed maneuvers.
4. Flight ( X0.008 Speed )
The video shows the representation intensity dynamics (left) and the simulated fly (right) during straight flight (0.0–30.0 ms) at a velocity of 20 cm/s. The trained controller maintains stable forward flight at a constant speed and keeps body orientation aligned with the target direction, demonstrating that the connectome-based network extends from walking to flight locomotion.
Neural Representations
Neural and behavioral dynamics across a continuous walking sequence.
We adopted a visualization that aggregates representations by annotated flow and superclass labels. The figure demonstrates how activity evolves over different behavioral phases. Distinct activation patterns emerge within each class, and transitions between behavioral phases imply how sensory observations and motor intentions are dynamically transmitted through the network.
Method
We model the connectome as an unweighted, directed graph, partitioning nodes into afferent, intrinsic, and efferent sets according to the direction of information flow.
The FlyGM architecture instantiates the full Drosophila connectome as a graph neural network, in which each of the 139,246 neurons (comprising 19,262 afferent, 118,496 intrinsic, and 1,488 efferent neurons) is represented by a node embedding.
At each control step, FlyGM receives sensory inputs through afferent pathways, propagates information across the brain-wide graph via message passing, and produces motor outputs through efferent pathways to actuate the body.
Conclusion
Our FlyGM demonstrates that whole-brain connectomes can be used for embodied motor control. By grounding policy architectures in neuronal wiring diagrams, this approach suggests a possible towards more human-aligned AI systems, where the inductive biases that shape adaptive behavior in animals can be systematically transferred to artificial agents.