Binocular Vision-Based Motion Planning of an AUV: A Deep Reinforcement Learning Approach

Vision-based motion planning of autonomous underwater vehicles (AUVs) is regarded as a critical requirement for marine intelligent transportation systems. However, the limited vision range and the uncertain model parameters of an AUV make it difficult to fulfill this requirement. This study focuses on a binocular-vision-based motion planning issue for an AUV. First, we develop an intelligent AUV system that mainly comprises binocular cameras for patrolling the target, a localization unit for acquiring the position information, and an acoustic modem for communicating with buoys. Accordingly, the parallax angles from the AUV to target are used to construct an optimal motion planning problem. To solve the aforementioned problem, we develop a deep reinforcement learning method called the improved twin delayed deep deterministic (TD3) policy gradient algorithm in order to minimize the reward function, such that the AUV can perpendicularly patrol the target with a fixed distance. The advantages of our solution are as follows: 1) the binocular-vision-based motion planning method can achieve a trade-off between motion stability and observation effectiveness; 2) the improved TD3 algorithm can accelerate the convergence compared to other algorithms, while it can simultaneously overcome the dependency on the model parameters of the AUV. Finally, simulation and experimental studies are conducted to verify the effectiveness.