Neuromorphic Computing for Autonomous Mobility in Natural Environments

Abstract

Certain bat species (horseshoe bats, Rhinolophidae) can navigate, avoid obstacles, and pursue prey in dense vegetation with exceptionally capable and efficient biosonar systems that operate in complete darkness and process biosonar echoes with masses as low as 0.2 g. Hence, they provide an opportunity to develop a novel integrated sensor system to enable autonomous navigation capabilities in small drones operating in complex natural environments. To explore this, we have investigated the biosonar system of Horseshoe bats from engineering, neuroscience, and signal processing perspectives, with the goal of developing a biomimetic system for small drones so that they can navigate in complex natural environments. To achieve this, a biomimetic sonar head capable of mimicking the peripheral dynamics seen in the biosonar of horseshoe bats has been used to collect a large data set of echoes (220,000) from natural forests. Then, these echoes were passed through a prototype brain-inspired signal processing chain that simulates the complete auditory process from input sound stimulus to the neural spike response. It should be possible to implement an integrated biomimetic/brain-inspired prototype system on a single chip (e.g., like prior neuromorphic implementations of auditory and vision systems). Such a single-chip implementation would make it possible to miniaturize the brain of an autonomous drone and serve to expand the use of autonomous small drones to numerous applications.