【作者】 吴红涛；

【导师】 何顶新；

【作者基本信息】 华中科技大学 ， 电子信息（专业学位）， 2024， 硕士

【摘要】 多机器人运输系统在室内建造中的应用可以提高工程的安全性和效率。避障和定位技术是机器人实现自主导航的关键。然而在动态的室内建造场景中,环境的不确定性使得机器人的避障和定位面临许多挑战。针对课题中机器人的自主导航需求,提出了一种多机器人避障算法并设计了一个经济的机器人室内定位方案。室内建造场景中存在环境信息部分可观测、障碍物复杂等挑战,而传统的避障算法大多要求对环境的精确感知且存在大量的参数调整工作。针对传统方法的缺陷,提出了基于深度强化学习的分布式多机器人避障算法。将多机器人避障问题建模为一个部分可观测的马尔可夫决策过程,采用深度神经网络拟合避障策略。策略网络以局部栅格地图等信息作为输入,直接输出机器人的速度指令。在机器人操作系统中搭建了深度强化学习的训练框架,构建了观测、动作空间和奖励函数。基于Gazebo物理仿真引擎搭建了多样的仿真场景,使用多机器人分布式近端策略优化算法来优化策略参数。采用集中学习、分布执行的框架,通过添加控制器引导并使用多场景两阶段的训练方法来加速模型的收敛。高效精准的定位是机器人自主导航的前提,工程中还需要考虑定位方案的经济性。目前主流的室内定位方案存在定位精度较低或者成本高昂等问题。针对课题需求,设计了基于AprilTag视觉基准系统的组合定位方法。搭建了基于AprilTag的位姿估计系统,机器人通过标签可以获取精确的全局位姿信息。建立了机器人定位模型,利用IMU和轮速计获取机器人的实时位姿估计。设计了基于拓展卡尔曼滤波的数据融合方案,基于AprilTag位姿估计系统的定位信息对机器人的实时位姿估计进行校正。通过仿真确定了AprilTag定位信息可用的距离阈值,通过轨迹对比实验验证了组合定位方法的有效性。通过仿真和物理实验证明了提出的避障算法和设计的定位方法的有效性和一定的可行性。在搭建的室内建造模拟场景中进行了综合性的多机器人导航实验,验证了所提算法和方案在实物平台上的整体有效性。更多还原

【Abstract】 The application of multi-robot transportation systems in indoor construction can improve the safety and efficiency of the project.Obstacle avoidance and localization techniques are crucial for robots to achieve autonomous navigation.However,in dynamic indoor construction scenarios,the uncertainty of the environment makes robot obstacle avoidance and localization face many challenges.To meet the autonomous navigation needs of robots in the project,a multi-robot obstacle avoidance algorithm is proposed and an economical indoor robot localization scheme is designed.Indoor construction scenarios present challenges such as partially observable environmental information and complex obstacles.Most traditional obstacle avoidance algorithms require accurate perception of the environment and extensive parameter tuning.To address these shortcomings,a distributed multi-robot obstacle avoidance algorithm based on deep reinforcement learning is proposed.The problem of multi-robot obstacle avoidance is modeled as a partially observable Markov decision process.A deep neural network is utilized to fit the obstacle avoidance strategy.The strategy network takes inputs such as local raster maps and outputs speed commands for the robots.A training framework for deep reinforcement learning is built in the robot operating system,and the observation,action space,and reward functions are constructed.Simulation scenarios are created based on the the Gazebo physical simulation engine.A multi-robot distributed proximal policy optimization algorithm is then used to optimize the policy parameters.To accelerate model convergence,a centralized learning and distributed execution framework is employed,which includes controller guidance and a multi-scene two-stage training method.Efficient and accurate localization is essential for autonomous robot navigation,and economic considerations must also be taken into account.The current mainstream indoor localization solutions suffer from problems such as low accuracy or high cost.To meet the project’s requirements,a combined localization method based on the AprilTag visual reference system is designed.A position estimation system utilizing AprilTag is implemented,allowing the robot to obtain precise global position information.Additionally,a robot localization model is developed to provide real-time position estimation using IMU and wheel speedometer data.To further refine the real-time position estimation,a data fusion scheme based on expanded Kalman filtering is employed,utilizing the localization information obtained from the AprilTag position estimation system.The available distance thresholds for AprilTag positioning information are determined through simulation.The effectiveness of the combined positioning method is verified through trajectory comparison experiments.The effectiveness and feasibility of the proposed obstacle avoidance algorithm and the designed localization method are demonstrated through simulation and physical experiments.Comprehensive multi-robot navigation experiments are conducted in an indoor construction simulation scenario to verify the overall effectiveness of the proposed algorithms and schemes on a physical platform.更多还原