摘要
针对移动机器人自主导航过程中建图、定位和路径规划系统复杂、误差大的难题,设计了一套未知环境下四足机器人指令式路径规划实验系统。集成VLP-16激光雷达、Realsense深度相机与AGX导航控制器,搭建了四足机器人自主导航硬件系统;采集激光点云数据带入LOAM算法,为机器人自主导航提供定位信息;将相机深度信息输入到端到端指令式路径规划网络,得到自主避障导航期望路径,并通过LCM通信将运动指令信息下发给四足机器人进行运动;指令式路径导航方法将传统“建图—定位—导航”过程简化为训练部署模式的“感知—规划”框架,实现了根据指令直接运动而无需先验地图的端到端自主导航。仿真和物理样机实验表明,四足机器人能够在室内外非建图环境下高动态自主导航到指令目标点,成功率在90%以上,动态环境中的成功率在85%以上,验证了系统的可行性与有效性。
[Objective]The autonomous robot navigation method generally requires first building an environment map and then performing path planning.This autonomous robot navigation method requires a priori maps,its positioning and path planning systems are complex,and problems,such as large cumulative errors,frequently occur.For this reason,this study designs an experimental system for end-to-end imperative path planning of quadruped robots in unknown environments.The robot autonomously determines the non-collision path trajectory using the real-time perceived environmental information and target navigation position and updates and adjusts the path for dynamically changing scenes.[Methods]The hardware system for the autonomous navigation of quadruped robots was built by integrating the VLP-16 laser radar,RealSense depth camera,and AGX navigation controller.The navigation system converts the autonomous motion path obtained by perception decision into motion speed and communicates with the quadruped robot locomotion controller through the network.The locomotion controller calculates the joint torque of the movement of the robot using the optimization algorithm.The autonomous navigation system first collects laser point cloud data and then inputs it into the LOAM algorithm to calculate positioning information for autonomous robot navigation.The depth information of the camera is collected and inputted into the end-to-end imperative path planning network,which is composed of ResNet-18 and MLP.The loss function of the network consists of the cost obtained by mapping the key point path to the cost map and other task-level losses.The imperative path planning network outputs the key points of the planned path,which are learned and trained through the prior trajectory output by the optimal trajectory optimizer.Meanwhile,the imperative planning network is optimized through backpropagation.The motion speed of the robot is calculated by limiting and interpolating the key points of the path and transmitted to the robot motion controller through LCM communication to realize the command-based navigation of the quadruped robot.This method simplifies the traditional“mapping–localization–navigation”process into a“perception–planning”framework that is first trained and then deployed,achieving autonomous movement based on the target location command without the need for a priori maps.[Results]The simulation and physical prototype experiments showed that the quadruped robot can autonomously navigate to the target position with a success rate of more than 90%in indoor and outdoor unmapped environments and a success rate of more than 85%in different scenarios,especially dynamic environments.In the comparative experiment with the A*algorithm,the quadruped robot was unable to adjust to people or obstacles that dynamically entered its field of view,resulting in collisions.However,the quadruped robot based on the method proposed in this study was able to quickly update its path and achieve non-collision autonomous navigation,which indicates the feasibility and effectiveness of the system and the strong generalization and robustness of the method.[Conclusions]Based on the instruction-based learning method,this study designs and builds an experimental system for quadruped robot path planning.The system uses the depth images collected by the depth camera in an end-to-end manner and outputs a smooth collision-free path planned at high frequency through the instruction-based navigation network.The system can cope with complex static and dynamic environments,has high feasibility and effectiveness,and guides the robot to move autonomously to the target point.
作者
陈腾
赵志成
荣学文
李贻斌
荣海林
CHEN Teng;ZHAO Zhicheng;RONG Xuewen;LI Yibin;RONG Hailin(School of Control Science and Engineering,Shandong University,Jinan 250061,China)
出处
《实验技术与管理》
北大核心
2025年第5期74-81,共8页
Experimental Technology and Management
基金
国家自然科学基金项目(62203268)
山东大学教育教学改革研究项目(2024Y183)
山东大学实验室建设与管理研究项目(sy20243302)。
关键词
路径规划
指令式学习
自主导航
四足机器人
path planning
imperative learning
autonomous navigation
quadruped robot
