Session: 08-04-02 Machine Learning and Optimization for Marine Design and Operation

Paper Number: 122872

122872 - Static Obstacle Avoidance With Path Following of Asvs by Using Twin Delayed Deep Deterministic Policy Gradients 

In the realm of autonomous surface vehicles (ASV), the presence of an intelligent collision avoidance system is imperative to guarantee safety during navigation. Within the scope of this study, a Deep Reinforcement Learning (DRL) algorithm based on Twin Delayed Deep Deterministic Policy Gradients (TD3) that operates with a continuous action space is used for path following and collision avoidance of a ship. The study also introduces an epsilon-greedy based exploration approach for TD3, offering an alternative to the noise-based method. The incorporation of delayed updates in TD3 serves to diminish per-update errors, leading to a further enhancement in the performance of neural networks. Additionally, a decaying epsilon-greedy algorithm is employed to facilitate the transition between exploration and exploitation phases. The design of the reward function is specifically tailored to minimize both the cross-track error and the heading error of the agent. In this particular case, the agent represents the scaled model of a KVLCC2 tanker. The central objective of this research is to achieve path following and static collision avoidance using data-driven control. Error calculations are based on the Line of Sight (LOS) algorithm. By constantly refining its policy using reinforcement learning, the ASV can respond to diverse scenarios and make instantaneous choices that prioritize safety while maintaining optimal navigation. The data-driven controller is tested for different paths and collision scenarios. Within this research, the neural network architecture for avoiding static obstacles has been effectively developed, trained, and tested.

Presenting Author: Vishnu K T INDIAN INSTITUE OF TECHNOLOGY MADRAS

Presenting Author Biography: Vishnu is MS student in the Dept. of Ocean Engg., IIT Madras, India. His research is in the field of development of the data driven controls for ship navigation.

Authors:

Vishnu K T INDIAN INSTITUE OF TECHNOLOGY MADRAS
Suresh Rajendran Indian Institute of Technology Madras
Abhilash S Somayajula Indian Institute of Technology Madras