无人驾驶铰接式车辆强化学习路径跟踪控制算法  

Reinforcement Learning Algorithm for Path Following Control of Articulated Vehicle

作　　者：邵俊恺[1] 赵翾[1,2] 杨珏[1] 张文明[1] 康翌婷[1] 赵鑫鑫[1] 

机构地区：[1]北京科技大学机械工程学院,北京100083 [2]北京华为数字技术有限公司,北京100085

出　　处：《农业机械学报》2017年第3期376-382,共7页Transactions of the Chinese Society for Agricultural Machinery

基　　金：国家高技术研究发展计划(863计划)项目(2011AA060404);中央高校基本科研业务费专项资金项目(FRF-TP-16-004A1)

摘　　要：针对无人驾驶铰接式运输车辆无人驾驶智能控制问题,提出了一种强化学习自适应PID路径跟踪控制算法。首先推导了铰接车的运动学模型,根据该模型建立实际行驶路径与参考路径偏差的模型,以PID控制算法为基础,设计了基于强化学习的自适应PID路径跟踪控制器,该控制器以横向位置偏差、航向角偏差、曲率偏差为输入,以转角控制量为输出,通过强化学习算法对PID参数进行在线自适应整定。最后在实车道路试验中验证了控制器的路径跟踪质量并与传统PID控制结果进行了对比。结果表明,相比于传统PID控制器,强化学习自适应PID控制器能够有效减小超调和震荡,实现精确跟踪参考路径,可以较好地实现系统动态性能和稳态误差性能的优化。With the industry 4.0 embraced a number of contemporary automation, data exchange and manufacturing technologies, the autonomous driving system is widespread. In order to enable the autonomous driving, path following strategies are essential to maintain the normal work of the vehicles. The articulated flame steering vehicles （ASV） are flexible, efficient and widely implemented in agriculture, mining, construction and forestry sectors due to their high maneuverability. The articulated vehicle usually composes of two units, a tractor and a trailer, which are connected by an articulation joint. However, as the ASV dynamics are significantly different from the conventional vehicles with front wheel steering, the path following controller derived for conventional vehicles is considered not to be applicable for the ASVs. Thus the path following control is challenging the robustness. A path following strategy is proposed for the ASVs on the basis of reinforcement learning adaptive PID algorithm. The kinematic model of the ASV is derived by neglecting the vehicle dynamics. Three measurable errors are defined to indicate the deviation of real path from reference path, i. e. , lateral displacement error, orientation error and curvature error. These errors are served as the inputs in order to synthesize the path following controller and the desired steering angle is served as the output of path following controller. Based on the PID algorithm, the reinforcement learning method is selected for optimizing the parameters of PID online to reduce the overshoot and chattering. Furthermore, the prototype test is conducted to evaluate the performance of the proposed control law. The result shows that compared with the traditional PID, reinforcement learning adaptive PID controller can restrain the overshoot and chattering efficiently and follow the reference path accurately.

关 键 词：铰接式车辆 驾驶 强化学习 路径跟踪 

分 类 号：TP273[自动化与计算机技术—检测技术与自动化装置] U463.325[自动化与计算机技术—控制科学与工程]