Finding the Goal: Insect-Inspired Spiking Neural Network for Heading Error Estimation

Insects have extraordinary navigational abilities. Monarch butterflies migrate every year to the same forest over hundreds of kilometers, desert ants find their way back to the nest tens of meters away and dung beetles maintain the same heading direction over meters. The performance of these agents has been optimized by evolution over the last 500 million years leading to power-efficient, low-latency and precise sensorimotor systems. Research efforts in the field of neuroscience, biology and robotics are instrumental for uncovering the neural substrate of insect navigation abilities. The development of models of insect navigation tightly coupled with the insect connectome and neurophysiology and their embedding in closed loop systems support the understanding of embodied animal cognition and can advance robotic systems. In this work, we focus on insect navigation because of the efficient insect navigational apparatus. Furthermore, the recent discovery of the central complex, the neuronal center of insect navigation, facilitates the development of new hypotheses about insect navigation. All navigating insects need to perform some kind of goal-directed behavior during which they have to reach a specific goal location or maintain the same movement direction over long distances. Such behavior requires the agent to be aware of its current heading direction, desired heading direction, and the error between them. Building on previous research in the field, we propose a novel model for this error estimation that can in principle be generalized for all navigating insect species. We implement the model in a spiking neural network and test its capabilities on a simulated robotic platform. The precision of the network is comparable to or even better than the biological role model. Thus, our implementation serves as a working hypothesis for how the heading error might be computed in the insect brain. Our model will help to explain navigational behavior in fruit flies, orchid bees, bumble bees and some less researched insect species. Furthermore, its simplicity in comparison to other models and implementation in a spiking neural network makes it very suitable for neuromorphic systems, an emerging field of brain-inspired hardware.