Real-time Mapping on a Neuromorphic Processor

Abstract

Mapping is a critical component for developing a simultaneous localization and mapping (SLAM) system in mobile robots. We draw from the brain's dedicated network that solves the spatial navigation problem by learning a cognitive map of the surrounding environment using networks of specialized neurons, such as place cells, grid cells, head direction cells, and border cells. We further integrated our neuro-inspired network into a neuromorphic processor, namely Intel's Loihi chip. Here, we proposed an SNN that used Winner-Take-ALL (WTA) structure and heterosynaptic competitive learning for place field generation and dendritic trees for reference frame transformation. The network learned distributed sub-maps on place cells, that, when combined, they encode accurately a unified map of the environment. By using an efficient interaction framework between the Robot Operating System (ROS) and Loihi, we showcase how our SNN may run in real-time interacting with a mobile robot equipped with a 360-degree LiDAR sensor. These results pave the way for an efficient neuromorphic SLAM solution on Loihi for robots operating in unknown environments.