消防机器人的移动能力是其在非结构化动态场景下进行作业任务的关键。本文根据消防机器人人工遥控技术和移动机器人自主导航理论，提出了一种消防机器人人机协同式导航系统，并针对系统的平台搭建、技术实现和改进优化进行了研究讨论，具体内容如下：首先，提出了一种融合人工和自主导航的人机协同式导航系统。该系统充分考虑了消防机器人在实际火灾现场执行灭火救援任务时对导航和控制技术的功能需求以及火灾现场非结构化和动态化的特点，实现了人工遥控技术和自主导航方案的有机融合；结合分布式控制系统和多传感器融合等技术原理，搭建了由上位机笔记本电脑、下位机TX2/STM32和通讯系统组成的硬件平台以及基于机器人操作系统和PyTorch的仿真模型与深度学习模型训练的软件环境。其次，设计了基于经典分层式导航的人机协同式导航系统算法部分的基本运行流程。结合人机结合思想和移动机器人坐标系变换基本原理，建立了算法前端隐目标点计算模块的数学模型；根据消防机器人和传感器实际物理特性，对人机协同式导航系统后端自主导航模块的具体算法选型和参数优化调整；进行了仿真环境和真实场景的定性和定量实验，实验结果证明了分层式人机协同式导航系统在多种结构化和非结构化场景中均具备导航有效性和动态避障能力，实际场景中该系统相比人工遥控方案的平均导航速度要快约20.16%，具备更高的导航效率。最后，根据深度强化学习原理对分层式人机协同式导航系统的后端模块进行了优化。采用TD3算法搭建避障网络模型，端到端地建立从激光雷达输入数据到实时速度指令的映射关系；针对训练环境和真实情况中激光雷达的数据格式不一致的问题，提出了两种解决方案，将网络感知接口的二维激光距离数据转化为三维点云数据；将训练好的网络模型无需人工调参直接部署在人机协同式导航系统中进行定性和定量测试，真实实验结果证明了基于TD3避障网络的人机协同导航系统比人工遥控和分层式算法的平均导航速度要快约73.91%和44.76%，实际CPU资源的消耗降低约1.7%，成功解决了分层式算法的平均导航速度慢、计算资源消耗高和人工调参困难等问题，具备更好的实际应用落地潜力。
The mobility of fire-fighting robots is the key to their tasks in unstructured dynamic scenarios. This thesis proposes a novel human-machine cooperative navigation system for fire-fighting robots combining the artificial remote control techniques and the theory of autonomous navigation and discusses the platform construction, technical implementation and optimization of the system. The details are as follows:First, this thesis proposes a human-machine cooperative navigation system that integrates manual and autonomous navigation. This system takes full account of functional requirements of navigation and control technology for fire-fighting robots performing disaster relief missions at the actual fire scene, as well as the unstructured and dynamic characteristics of the fire scene and achieves the organic integration of manual remote control and autonomous navigation techniques. Based on principles of distributed control systems and Multi-sensor fusion, a hardware platform consisting of upper computer laptop, lower computer TX2/STM32 and communication system and a software environment for robot model simulation, algorithm tuning and deep learning training.Secondly, this thesis designs the basic operation flow of the algorithmic part of the human-machine cooperative navigation system based on classical hierarchical navigation. Combining the idea of human-computer integration and the basic principle of mobile robot coordinate system transformation, the mathematical model of the hidden target point calculation module at the front end of the algorithm is proposed. According to the actual physical characteristics of firefighting robots and sensors, algorithm selection and parameter optimization adjustments are made to the back-end autonomous navigation module of the human-machine cooperative navigation system. This navigation system is deployed in Gazebo virtual environments and real scenarios for qualitative and quantitative experiments. The experimental results show that the hierarchical human-machine cooperative navigation system has navigation effectiveness and dynamic obstacle avoidance capability in a series of different structured and unstructured scenarios, and the average navigation speed of the system is about 20.16% faster than that of the manual remote control solution in real scenarios, which has higher navigation efficiency Finally, this thesis proposes an optimization method of the back-end module of the hierarchical human-machine cooperative navigation using deep reinforcement learning algorithms. The TD3 obstacle avoidance network replaces the map construction and planning decision algorithms in the classical system, eliminating the need to build a map of the environment and establishing a direct end-to-end mapping relationship from sensor data to velocity commands. To address the problem of inconsistent data formats of LIDAR in the training environment and the real situation, two solutions are proposed to convert the 2D laser distance data from the network sensing interface into 3D point cloud data. The trained network models are deployed directly in the human-machine cooperative navigation system for qualitative and quantitative testing without manual tuning. Simulation and real experimental results show that the human-machine cooperative navigation system based on TD3 obstacle avoidance network is about 73.91% and 44.76% faster than the average navigation speed of the manual remote control and the classical algorithm, and the actual CPU resource consumption is reduced by about 1.7%, which successfully solves the problems of slow average navigation speed, high computational resource consumption and difficulty of manual tuning of the hierarchical human-machine cooperative navigation algorithm, and has better potential for practical application.
