2026-10-10T20:40:29Z https://www.scidb.cn/oai

10.57760/sciencedb.41512 2026-07-03T11:31:01Z

2026-07-03 Simulation Code, Trained Models, and Evaluation Dataset for Three-Dimensional Obstacle Avoidance of Autonomous Underwater Vehicles doi:10.57760/sciencedb.41512 en This dataset is derived from the study entitled “Research on Autonomous Obstacle Avoidance for Autonomous Underwater Vehicles Based on Sonar Perception and SAC-LADRC Hierarchical Control.” The dataset contains simulation code for three-dimensional obstacle avoidance of autonomous underwater vehicles, reinforcement learning training configurations, trained models, evaluation results, ablation experiment data, and trajectory visualization files. It can be used to reproduce autonomous underwater vehicle obstacle avoidance experiments based on forward-looking sonar perception and hierarchical reinforcement learning control.This study focuses on safe autonomous navigation of autonomous underwater vehicles in complex underwater environments. A five-degree-of-freedom motion simulation model of an autonomous underwater vehicle, a forward-looking sonar local perception model, and a hierarchical obstacle avoidance framework based on Soft Actor-Critic and linear active disturbance rejection control are developed. In the upper layer, the Soft Actor-Critic policy generates the reference surge speed, reference pitch angle, and reference yaw angle according to the vehicle motion state, goal-relative state, and forward-looking sonar risk features. In the lower layer, a linear active disturbance rejection controller based on a linear extended state observer tracks the reference commands and outputs the surge force, pitch moment, and yaw moment. The forward-looking sonar observations are encoded into low-dimensional risk features, including the left and right obstacle intensity features, the nearest obstacle distance in the front sector, the nearest obstacle distance in the left sector, and the nearest obstacle distance in the right sector, which are used to support local obstacle avoidance decision-making.The simulation data and evaluation results in this dataset were generated using a Python-based simulation environment. The training process follows a staged curriculum learning strategy, including obstacle-free goal-reaching training, single-obstacle and simple composite-obstacle training, complex static multi-obstacle avoidance training, and extensible scenarios involving ocean current disturbances, actuator faults, and dynamic obstacles. The current main results focus on three-dimensional autonomous obstacle avoidance in complex static multi-obstacle and composite-obstacle environments, and include comparative results between the complete method and multiple ablation variants. The evaluation metrics include task success rate, collision rate, timeout rate, minimum safety distance, path length, trajectory smoothness, control tracking error, sonar response, and obstacle-avoidance mode activation. experimental; data; AUV Liu Wensuo Huang Aigen PUBLIC https://api.github.com/licenses/agpl-3.0 dataset Science Data Bank