基于激光雷达和神经网络的移动机器人综合局部路径规划

【作者】 赵慧；

【导师】 蔡自兴；

【作者基本信息】 中南大学 ， 模式识别与智能系统， 2004， 硕士

【摘要】 路径规划是移动机器人研究中的重要问题之一，本文主要研究在不确定环境下移动机器人的局部路径规划。 人工势场法模仿物理学中势场的概念，假想环境对机器人有一定的作用力，力的方向就是机器人前进的方向。这种方法计算简单，但是存在局部最小、相邻障碍之间找不到通道、走廊环境下存在振荡等问题。对于局部最小，可以增加一个指向自由区域的力作为机器人的前进方向。 Q学习提供智能系统在马尔可夫环境中利用经历的动作序列选择最优动作的一种学习能力。采用模拟退火算法进行随机动作选择，根据动作之间的相似性调整每个动作的Q值，提高了机器人对环境的适应能力。但是这种方法计算复杂，规划周期是人工势场法的1.5倍，所占用的存储空间是人工势场法的4倍。 本文提出了一种综合局部路径规划方法，可以实现二者的优势互补。引入BP网络对环境进行划分，把环境分成四类：相邻障碍、走廊环境、U型区域和其它环境。其它环境采用人工势场法规划，以发挥其方便灵活的特点；相邻障碍和走廊环境等人工势场法不能正常工作的情况下采用Q学习方法；对于U型区，局部路径规划方法不能保证越过障碍，因此可以使机器人沿着较近的U型边界走出障碍。 激光测距范围广、精度高、传输速度快，适合机器人的实时避障，本文采用激光雷达实时采集环境信息，作为分类BP网络、人工势场和Q学习网络的输入。 试验和仿真结果证明这种方法可以实现人工势场法和Q学习之间的优势互补。通过这两种方法的综合应用，机器人在不确定环境下可以找到接近最优的路径，避免与静态障碍或者动态物体的碰撞，安全到达目的地。更多还原

【Abstract】 Path planning is one of the most important issues of mobile robot. This paper concentrates on the local path planning of the mobile robot in uncertain environment.Artificial potential field imitates the concept of the tendency field in physics. Supposed the environment exert some strength on the robot, the direction of the strength is the advancing direction of the robot. This algorithm is convenient to be realized, but there are inherent limitations, such as trap situations due to local minima (cyclic behavior), no passage between closely spaced obstacles, oscillations in narrow passages. For the local minima, an additional strength towards the free area is used in the direction of the robot forwarding.Q-learning offers the intelligence system a kind of learning ability by utilizing the experienced movement array to choose the optimum movement in the environment of Malkov. The robot action is selected by simulated annealing algorithm. Each action’s value is adjusted based on resembling among actions to improve the adaptive capacity of robot in environment. However, the Q-learning has some disadvantages such as calculation inconveniently, longer planning cycle (1.5 times compared to the artificial potential field) and lager memory space (4 times to the artificial potential field).A synthesized method of the local path planning, which can realize the mutual supplement with advantages of these two algorithms, is proposed in this thesis. BP neural network is introduced to classify the robot environment. This environment is divided into four kinds: closely spaced obstacles, corridor, U-shape area and other. In the last kind of environment, the algorithm of artificial potential field is used and is shown computation simpleness. The Q-learning works in the environment of closely spaced obstacles and corridor in which the artificial potential field can’t work well enough. The local path planning can’t guaranteepassing obstacles for U-shape, thus the robot avoid the obstacles along the relatively near U-shape border.The laser range finder has very long range, high precision and quick transmission speed, and is suitable for avoiding obstacles in real time environment. Real-time environment information is obtained through laser radar, and acts as the input for the classified BP network, artificial potential field and Q-learning.It is proved that the method can realize mutual supplement with artificial potential field and Q-learning each other by test and simulation. Through the integrated application of these two algorithms, the robot can find the almost optimum route under the uncertain environment, prevent collisions with fixed obstacles or dynamic ones, and reach the destination safely.更多还原