Multi-sensor Fusion Of Quadruped Robot Navigation System

With the continuous expansion of the application range of quadruped robots,their working environment is becoming more and more complex.Many tasks such as exploration and inspection in dangerous and remote areas are often undertaken by robots,so the autonomous navigation of quadruped robots has good research value and significance.However,complex working conditions such as sensor data distortion caused by body and sensor shaking,feature mismatch in a monotonous environment,and leg slipping caused by uneven roads will greatly reduce the robustness of the navigation system.There are still many problems in the state estimation and map generation of the footed robot navigation system.This paper introduces the leg odometer into the traditional SLAM framework,and proposes an extended Kalman filter localization algorithm based on dynamic errors,which improves the accuracy of the robot’s odometer;at the same time,the extraction accuracy of the ground point cloud is improved by likelihood estimation,and the The ground point cloud is used as a constraint to generate a higher-precision grid map,and finally a complete quadruped robot navigation system is realized.The specific work content is as follows:(1)The multi-sensor fusion positioning method based on radar,gyroscope and encoder is realized by using the extended Kalman filter.Dynamically adjust the covariance of the radar odometer to reduce the influence of the error of the radar odometer when the field of view is limited.At the same time,the gyroscope is used to improve the instantaneous accuracy of the odometer.The final fusion odometer travels about 50 m in a monotonous environment,and the positioning accuracy remains at Within 0.2m,travel200 m in an outdoor environment,the accuracy is kept within 2.0m,and the output frequency is above 10 Hz.(2)Improve the traditional ground detection algorithm based on the working conditions of the quadruped robot,improve the uniform segmentation of the radar scanning area to an adaptive area segmentation based on the point cloud density,and introduce vertical Degree,height,and flatness are three evaluation criteria,and the ground point cloud is filtered twice through likelihood estimation,and the cost map is constrained based on the ground point cloud and the pose of the quadruped robot to obtain a higherprecision navigation map.Compared with the traditional octree map projection method,the misjudgment of obstacles is reduced obviously compared with the traditional fixed threshold projection.(3)Build and implement a complete quadruped robot navigation system,including hardware platform construction and software implementation.Based on the D* algorithm and the DWA algorithm as the global and local path planners of the navigation system,point-to-point navigation based on prior maps is realized,and it maintains a good ability to avoid static and dynamic obstacles during the travel process,within 500 square meters The navigation error is kept within 1.5m.