【作者】 王宁；

【导师】 封锡盛； 王永；

【作者基本信息】 中国科学技术大学 ， 控制科学与工程， 2022， 硕士

【摘要】 无人水面航行器作为一种自主水上运载平台,在海洋监测、环境保护以及应急救援中发挥着至关重要的作用,而不同的任务需要部署水面航行器到不同环境,因此需要准确、可靠的导航技术引导水面航行器顺利执行任务。目前一个完整的自主导航系统通常划分为建图、规划和控制,需要针对每一环节设计相应的算法,并建立各模块之间的联系。其中局部路径规划算法和运动控制算法是自主导航系统的重要组成部分,直接决定了导航系统的性能。为进一步提高水面航行器的自主作业能力,本文围绕无人水面航行器的局部路径规划和运动控制算法展开研究,并进行了仿真和实物实验。具体来说,本文的主要研究内容和研究成果如下:1.基于深度强化学习构建了水面航行器的局部路径规划算法,设计了无动力学环境对深度强化学习模型进行训练,减少模型收敛时间的同时,也可以避免水上环境建模的繁琐。另外,针对深度强化学习应用于机器人导航存在的模拟到现实迁移问题,对环境信息引入域随机化,提高模型的泛化性和可迁移性。但在复杂多变的场景中训练会使得深度强化学习收敛困难,因此本文精心设计了自适应课程学习方案来加速神经网络的收敛。同时,也提出了一致性策略,通过考虑稳定性和执行器约束来预测控制指令,提高导航指令的可行性和轨迹平滑性。2.对本文所用水面航行器进行了运动学和动力学建模,并设计了两种运动控制器,分别是基于自适应控制的传统算法以及基于强化学习的智能算法,对比分析了两种控制算法在跟随不同输入信号以及干扰下的表现。3.为了评估所提出方法的有效性,本文在仿真环境下进行了广泛的实验,展示了所提出算法在收敛速度、泛化性、鲁棒性和轨迹平滑度等方面的优越性。更进一步的,在物理平台上开发了局部规划控制系统并投入到实际水上环境进行测试,通过定性、定量的分析,验证了所提出方法可以直接部署于水面航行器中,而不需要额外的调整和训练,并且相比于人工势场法,具有更短的导航路径和更好的轨迹平滑性。更多还原

【Abstract】 As an autonomous waterborne platform,unmanned surface vehicles play a vital role in marine monitoring,environmental protection,and emergency rescue.Since different missions will deploy unmanned surface vehicles to different environments,effective and reliable navigation technology is required to guide surface vehicles.A complete navigation system is usually divided into mapping,planning,and control,and it is necessary to design corresponding algorithms for each module and establish the connection between the modules.Among them,the local path planning algorithm and motion control algorithm are important components of the autonomous navigation system,which directly determine the navigation performance.In order to further improve the autonomous operation capability of surface vehicles,this dissertation focuses on the local path planning and motion control algorithms of unmanned surface vehicles,and conducts simulation and physical experiments.Specifically,the main research contents of this dissertation are as follows.1.This dissertation proposed a local path planning algorithm for surface vehicles based on deep reinforcement learning,and designs a dynamics-free environment for training deep reinforcement learning models,which can reduce the convergence time while avoiding the tediousness of modeling waterborne environments.In addition,for the sim-to-real problem,domain randomization is introduced to environmental information to improve the generalizability and transferability.However,training in complex and variable scenarios will cause convergence difficult.Therefore,this dissertation designs adaptive curriculum learning to accelerate the convergence of the neural network.Also,a consistency strategy is proposed to improve the feasibility and trajectory smoothness of navigation commands by considering stability and actuator constraints.2.In this dissertation,we model the kinematics and dynamics of the surface vehicle,and design two controllers for surface vehicles,the conventional algorithm based on adaptive control and the intelligent algorithm based on reinforcement learning,respectively.Then we compare the performance of the two controllers in following different input signals.3.In order to evaluate the effectiveness of the proposed method,extensive experiments are conducted in simulation environments to demonstrate the superiority of the proposed method in terms of convergence speed,generalization performance,robustness,and trajectory smoothness.Further,the autonomous navigation system is developed on a physical platform for testing,and through qualitative and quantitative analysis,it is verified that our method can be directly deployed in a real surface vehicle without additional tuning as well as training and has a shorter navigation path as well as better trajectory smoothness compared to the artificial potential field method.更多还原