Egocentric Head Direction Encoding and Perception Model Based on an Adaptive Ring Attractor Network.

Abstract

Head direction perception is a fundamental ability that enables humans and animals to navigate and orient themselves effectively in natural environments. In the era of embodied intelligence and agent research, head direction perception has broad applications in robots, drones, autonomous vehicles, spacecraft, and so on. However, to the best of our knowledge, no models for head direction sensing, encoding, and perception have been developed so far. In this article, we proposed a new neural network system with a sensor-encoder-processor framework for the head direction perception task based on the findings of neuroscience research. The system contains a new signal perception (sensor) component which is an adaptive ring attractor network (RAN) that updates its synaptic weights via Hebbian learning rules, a new signal encoder that converts nerve impulse to sparse distributed representation vectors, and an central information processor which is a brain-inspired and low-energy consumption neural network-hierarchical temporal memory (HTM) model to receive the encoding and predict the head direction of the agent. At last, a simulated robot is created which is equipped with our sensing, encoding, and perception systems. After training the HTM network, our system achieves a prediction accuracy of 94.4% with a mean error of just 0.062-an order-of-magnitude improvement over existing models. The codes will be available from the website: https://github.com/czgbjy/HeadDirectionEncoding.git.