Modeling and simulation of a double DQN algorithm for dynamic obstacle avoidance in autonomous vehicle navigation
Sivayazi Kappagantula, Giriraj Mannayee
Abstract
Developing an autonomous vehicle in safe navigated open environments through self-learning is a formidable challenge. So, in this paper, an autonomous navigation implementation for urban environments with dynamic obstacles using a double DQN algorithm is proposed. Furthermore, the vehicle is trained in varied training environments, explored neural network architectures, and fine-tuned action and reward functions to optimize its performance. In conclusion, the vehicle performance is steadily improved with each iteration of training, as evident from the upward trend in the average reward graph. These graphs depict the vehicle's learning progress in the scenario where it is trained to avoid dynamic obstacles. Finally, the proposed algorithm shows better performance relative to the conventional algorithms. The entire work is modelled and simulated in CARLA software.