Emulating insect brains for neuromorphic navigation

TL;DR

This work demonstrates a purely spike-based path integration network implemented on the accelerated BrainScaleS-2 neuromorphic chip to guide a virtual bee back to its home. By grounding the model in insect physiology and introducing axo-axonic short-term memory, the authors achieve robust two-dimensional navigation on hardware while integrating a virtual body via a co-processor. The approach runs orders of magnitude faster than biology, enabling 4800 bee journeys across 320 generations in about 0.5 hours and achieving substantial performance gains through CMA-ES optimization that mitigates hardware variability. This study highlights the potential of accelerated neuromorphic systems for rapid neurorobotic prototyping and memory-enabled control, with applicability to broader insect-inspired or memory-centric computing tasks. All mathematical expressions are presented with proper notation to ensure precise interpretation of the models and results.

Abstract

Bees display the remarkable ability to return home in a straight line after meandering excursions to their environment. Neurobiological imaging studies have revealed that this capability emerges from a path integration mechanism implemented within the insect's brain. In the present work, we emulate this neural network on the neuromorphic mixed-signal processor BrainScaleS-2 to guide bees, virtually embodied on a digital co-processor, back to their home location after randomly exploring their environment. To realize the underlying neural integrators, we introduce single-neuron spike-based short-term memory cells with axo-axonic synapses. All entities, including environment, sensory organs, brain, actuators, and the virtual body, run autonomously on a single BrainScaleS-2 microchip. The functioning network is fine-tuned for better precision and reliability through an evolution strategy. As BrainScaleS-2 emulates neural processes 1000 times faster than biology, 4800 consecutive bee journeys distributed over 320 generations occur within only half an hour on a single neuromorphic core.

Figures (7)

• Figure 1: BrainScaleS-2 prototype system. A) Entire system with , BrainScaleS-2 chip and periphery. B) BrainScaleS-2 prototype chip of the second generation. C) System schematic depicting the and the chip. Events coming from the are distributed to the synapse rows. The synaptic columns are connected to the triangular-shaped neurons that send spikes to the spike counters at the bottom and then back to the .
• Figure 2: Network architecture. The black and blue arrows correspond to excitatory and inhibitory connections, respectively. The and populations are each divided into a left and right subpopulation, resulting in two logical hemispheres. The network topology relies on a physiological model proposed in stone2017anatomically.
• Figure 3: Network activity and trajectory. The top left plot shows the entire trajectory consisting of the outbound (black) and return phase (gray). The top right plot shows a zoom into a region measuring 5000 $\times$ 5000 steps around the origin where the agent loops around after his return. Four particular moments are highlighted as blue dots and connected to the network activity plot below: the moment of return, reaching the home location, and two situations during the looping phase. The neuron activities are measured as output spike counts in 1ms intervals (corresponding to 1s intervals in the biological time domain) and are displayed color-coded over time. The maximum count is approximately 100 which corresponds to an instantaneous spike rate of 100k (equivalent to 100 when translated into biological time scales).
• Figure 4: Statistical performance of the evolutionary optimization. Three sample trajectories generated by the primitive (left) and the evolved network (right). Below is a $6000\times 6000$ zoom into the center region that shows a histogram over the data points of the looping phase of 1000 trajectories. The primitive network's center of looping is shifted to the lower left with a broad and elliptically deformed looping area. The evolved network is more centered and exhibits tighter and more symmetric looping behavior. The lower left plot shows the population fitness (gray) and the fitness of the best three individuals (blue). For each individual, the faint line is the actual fitness whereas the strong line provides a moving average for better visibility. The optimization converges after $\sim$ 200 steps.
• Figure 5: Coordinate system of the insectoid agent. $\Theta$ is the direction of flight and $\phi^{(t)}$ and $\phi^{(t-1)}$ the head direction of the current and the previous time step, respectively. $\Delta\phi$ is the difference between both and indicates the head rotation. As the time interval is discrete and defined to be 1, the traveled distance is equal to the velocity vector $\bm{v}$.