An innovative trajectory tracking and obstacle avoidance framework for autonomous underwater vehicles using adaptive and sparse attention-based DRL

Autonomous Underwater Vehicles (AUV) are widely used for civil and military purposes, especially in complex underwater tasks. Several approaches have been proposed for accurately determining the complex underwater environment in AUV. The existing models are not efficiently utilised, which impacts the simulation performance. They do not efficiently predict complex external disturbances to minimise the model accuracy. This paper implements an Adaptive and Sparse Attention-based Deep Reinforcement Learning (ASA-DRL) approach to resolve the 3D path tracking and the obstacle avoidance problems. By employing the safety factors-aided reward functions, unwanted developments are eliminated. In addition, the ASA-DRL approach's hyperparameters are optimised by Randomly Revised Quokka Swarm Optimisation (RRQSO). The step number and number of episodes parameters are quickly tuned in the optimisation phase. The proposed approach attained an optimal energy consumption rate of 156%, 196% and 116% in terms of 32, 64 and 128 state sizes.