Deep neuromorphic controller with dynamic topology for aerial robots

Abstract

Flying robots are at the peak of attention when considering new approaches to search and rescue operations, mapping, and navigation along with inspection and maintenance. With current improvements in neural networks and motor designs, drones can manoeuvre and navigate through routes without human intervention. However, the constant changes in the environment do place a daunt on memory allocation and computational power. Also, being trained in a representative environment tends to overfit to new data or might be unable to adapt. As such a deep neuromorphic controller approach is proposed in the paper which gives it extra computational power. We are using a simulated world in Gazebo where we run the drone for a fixed time and compare the results of its movement with and without adaptive networks. The results clearly display that an adaptive neural network combined with a neuromorphic controller adapt s to the changes in the environment and results in less deviation from the actual path. This paper’s theory is based on the findings of the published paper [1]. The code and associated files for this project can be found on GitHub on this link: https://github.com/ManishDhanetwal/Deep-Neuromorphic-Controller-with-Dynamic-Topology-for-Aerial-Robots.