Learning Through Adverse Event for Collision Avoidance: A Self-Learning Approach

Abstract

We introduce a deep learning based collision avoidance based on learning events accompanied by an online, semi-supervised learning algorithm that allows the learning agent to gain experiences and learn by itself without any preacquired training dataset through online trial-and-error approach. Using distance sequences as inputs, two procedures are performed in the proposed algorithm; data gathering procedure and learning procedure. Simulation results show that our system can achieve minimum of 99.86% up to 99.99% accuracy in classifying collision event from all possible events, allowing autonomous agent to navigate within simulated environments without collision.