Terrain Traversability Analysis Of Quadruped Robot Based On Terrain Features And Leader Teaching And Learning

In the field of mobile robots focused on the outdoor environment,the quadruped robot has been highly concerned by researchers because of their better motion flexibility and environmental adaptability.There have been some achievement in the aspects of robot's structural design,power system and control strategies.The basic performance of the quadruped robot continues to be improved,and the motion scene has also begun to expand from the laboratory to the complex and diverse outdoor environment.It is necessary for the quadruped robot not only to improve the motion stability,but also have better performance in autonomy.Terrain traversability analysis is one of the most important content of the reaserch about robot autonomous navigation.The quadruped robot can obtain the ability of terrain recognition through the perception and analysis of complex terrain,which is of great significance for the robot to select the best motion strategy and improve the robot's autonomy.This paper focused on the terrain traversability analysis for the quadruped robot by using 3D laser radar(VLP-16)as the sensor for terrain perception.After analyzing and summarizing the status of domestic and foreign research on robot environment perception and terrain traversability,the following research work has been done:(1)Calibration of the laser radar was achieved by using an external parameter calibration method based on surface normal vectors.The acquisition method and characteristics of VLP-16 radar data are analyzed.The external parameter calibration algorithm based on surface normal vector is used and combined with the plane fitting method(RANSAC)and singular value decomposition algorithm(SVD)to obtain the rotation matrix and translation vector between the laser coordinate system and the robot center coordinate system.As a result,three-dimensional point cloud in robot center coordinate system can be obtained.(2)Terrain traversability is preliminarily analyzed using terrain hazard assessment functions that combine terrain features with robot structural parameters.A 2.5D grid map was established using three-dimensional laser points,and terrain features such as slope,roughness,and stepheight for the traversability analysis were extracted.Combined with the structural design and performance of the quadruped robot,the parameters about robot motion stability such as maximum step height,maximum trunk inclination,and upper and lower body vibration margins were learned to establish the terrain hazard asscessment function for the preliminary determination of terrain traversability.As the same time,terrain feature parameters space is avalible.(3)A online learning method using SVDD based on leader teaching for terrain traversability is proposed.In order to sovle the problem of terrain analysis of quadruped robots,an on-line learning method based on leader teaching is proposed.Terrain feature parameters are extracted based on terrain marker information provided by the robot motion process of following the leader and used as a passable positive sample to online learning,using the SVDD algorithm to train quadruped robot's terrain traversability model,which enabled the robot to have the ability of terrain cognition and learning.At the same time,aiming at the efficiency problems caused by the accumulation of the positive samples in the process of SVDD online training,the DBSCAN density clustering method is used to cluster and denoise the samples,and effective samples are extracted to improve the efficiency of online learning.