3D autonomous dynamic path planning for UUVs based on obstacle motion prediction and reinforcement learning

Abstract

Unmanned Underwater Vehicles (UUVs) are widely applied in marine exploration and other missions due to their low cost and high reliability. Point-to-point path planning serves as the foundation for accomplishing various marine tasks. The limitations of sonar sensor detection pose significant challenges for UUV obstacle avoidance path planning in 3D dynamic environments. To address these issues, this paper proposes a 3D autonomous dynamic path planning approach for UUVs based on obstacle motion prediction and reinforcement learning. This approach consists of three main modules: the observation preprocessing module utilizes dual-sonar data to extend the UUV's environmental observation from 2D to 3D; the prediction module integrates multi-sensor fusion with a Gated Recurrent Unit (GRU) algorithm to predict the motion of dynamic obstacles and conduct risk assessments; the path planning module employs an improved deep reinforcement learning algorithm, incorporating both obstacle motion predictions and current environmental information to make proactive obstacle avoidance decisions, thereby achieving collision-free path planning. In this study, the UUVSimulator platform is used to simulate realistic dynamic underwater environments, verifying the effectiveness of the proposed algorithm across multiple dynamic scenarios. Compared with several deep reinforcement learning algorithms, the proposed approach improves the success rate by approximately 13.4% to 40%.