Research On Velocity Estimation Of Legged Robots Based On Extended Kalman Filter

Legged robot is an important part of mobile robot. Compared with wheeled andtracked robot, legged robot could select its footholds itself and therefore surmountobstacles, which has great potential in mountain transport, rescue, military fields, etc.Currently, fast dynamic walking of legged robot has become the research focus at homeand abroad, while the realtime motion parameters, such as body attitude and velocity, arethe feedback information needed for robot’s stable motion control. However,there usuallyexists bias and random errors in the inertial measurement unit (IMU), therefore thevelocity integration drifts heavily. Moreover, the legged robot’s impact and vibrationcaused by interaction of feet and ground, makes the velocity estimation more difficult.How to estimate the velocity of legged robot at lower cost by utilizing its structure,proprietary sensors and IMU, has become an important research direction in robotsnavigation technology. The thesis compared several methods of robot’s state estimation,and fused the strapdown inertial navigation system (SINS) and forward kinematics byextended kalman filter (EKF) to acquire stable velocity estimation for the biped andquadruped.Firstly, the processing of MEMS-SINS was researched for the recognition of biasand random errors, and then its navigation solution was also studied. Both the static andon-line calibration of accelerometers were studied and compared, realizing the dynamicrecognition and compensation of bias. The random errors of accelerometer wererecognized by means of Allan variance, which laid the foundation for data fusionalgorithm. Furthermore, the thesis researched and simplified the method of velocityintegration according to the SINS mechanization and the actual robot locomotion. Theslider experiment was conducted and validated the SINS performance, processingmethod and navigation solution preliminarily.Secondly, according to the model of the biped and quadruped needed in theexperiments, the forward kinematics was calculated to acquire the measurement ofvelocity estimation by building the link frame of legs, which was testified by simulation.Furthermore, the thesis researched the data fusion algorithm-EKF, and the errors ofmotion parameters were selected as the state variables. Then the state equation andmeasurement equation were deduced and the optimal estimation of errors were acquiredby tuning the variances. On the basis, the optimal estimation of errors were compensatedthrough feedback correction. Then the motion simulation of quadruped was carried outand validated the effectiveness of EKF for walking forward, walking laterally and toughterrains.At last, experiments were conducted using biped and quadruped platform to validatethe processing of SINS, navigation solution and data fusion algorithm. The motioncapture system was built to obtain the real velocity of robot and experiments of walking forward and laterally, turning and surmounting obstacles were carried out for the biped.Then experiment of walking on plat road was conducted for the quadruped， usingaverage velocity calculated from forward kinematics as the reference. According toexperiment results, the velocity estimation algorithm was validated and could be referredby locomotion control of robots.