Research On State Estimation For Quadruped Robots

Legged robot is a kind of representative mobile robot.Compared with common wheeled robot and tracked robot,legged robot has innate advantages in humancomputer interaction and complex terrain adaptation.Since it has ability to choose its own landing point in the process of movement,it can travel freely on the unstructured road,which possesses potential application value in the fields of rescue and relief,mountain transportation,field investigation,military application,planetary exploration and so on.Quadruped robot can be regarded as an indispensable member of the family of legged robots.Taking quadruped as the bionic object,quadruped robot not only has better stability and bearing capacity than biped robot,but also has better flexibility and simpler structure than hexapod and octapod robot.More importantly,quadruped robot presents great development potential in motion ability,so it has become a research hotspot of scholars at home and abroad.In order to achieve the expected autonomous motion ability of quadruped robot,it is necessary to provide the position,velocity,attitude and other real-time motion state information of the main control system accurately.However,the accuracy of any sensor is limited,and there is a certain degree of uncertainty in the observation value,which leads to state estimation.The goal of state estimation is to obtain the global optimal estimation of state information.For an application like quadruped robot,it often brings new challenges to the state estimation.The impact vibration caused by the foot bottom interaction and the possible sliding between the foot end and the ground further increase the difficulty of motion state estimation.It is required that the designed motion state estimation method should not only take the high precision but also the robustness into account.Besides the above estimation performance,the motion state estimation method is also needed to meet the real-time requirements of quadruped robot.Therefore,how to estimate the motion state of quadruped robot stably and effectively is a challenging and profound subject.In this paper,the state estimation method of the quadruped robot is studied.Two motion state estimation algorithms,a laser infrared acquisition system and a fusion estimation method,are proposed to complete the overall construction of the motion state estimation system,and finally realize the global optimal estimation of the motion state of the quadruped robot,which provides a stable,accurate and real-time motion for the main control system of the robot control parameters.First,the SINS is preprocessed to identify the deviation and random error,and the accelerometer and gyroscope will be disturbed by multiplicative noise in the working process.In order to improve the estimation accuracy of quadruped robot's motion state,this paper uses forward kinematics to provide internal reference information to SINS for auxiliary correction of accelerometer.On the basis of the quadruped robot model needed in the experiment,the leg bar coordinate system is established,and the forward kinematics analysis is carried out.In the ideal environment where the system is only disturbed by additive noise,the Strong Tracking Mixed order Cubature Kalman Filter(STMCKF)algorithm proposed in this paper can be used to effectively fuse the information of the forward kinematics and inertial navigation system under the condition that the system is only disturbed by additive noise and multiplicative noise.The improved STMCKF can effectively fuse the two noises.The fusion result compensates the accelerometer deviation by feedback correction and improves the output accuracy of INS.Due to the introduction of strong tracking,the algorithm can overcome the problem that the quadruped robot's forward kinematics calculation error increases due to the impact and vibration of the foot,the foot tip slip,the maneuvering action or being interfered by external forces.At the same time,the traditional strong tracking cubature Kalman filter algorithm is reduced from three volume point sampling to one,which greatly reduces the amount of calculation.The correctness and validity of the algorithm are verified by simulation and experiment.Then,in order to calibrate the gyroscope in the inertial navigation system on line and improve the accuracy of attitude angle calculation,this paper introduces external reference information.At the same time,in order to avoid the problem of large calculation and poor real-time performance of the traditional visual algorithm,the acquisition system of the visual odometer is improved,and the laser infrared acquisition system is designed.The system is installed in the chassis of the robot,which aims to make the laser emission end perpendicular to the main axis of the camera to the ground.Due to the strong robustness of the scheme,it does not need precise debugging.In order to facilitate the operation,the laser spot can be photographed by the monocular camera as long as it nears the image center.This system only needs to shoot the local ground near the spot to obtain the infrared thermal image which can be used for robot motion state estimation.By adopting the active marking of feature points,the traditional vision algorithm can avoid the problem of no matching caused by repeated texture,non obvious features and poor gray scale differentiation.For example,in the face of white wall,desert,or longdistance feature points caused by the matching failure,as well as in the camera range,there are problems such as difficult positioning and large errors caused by moving targets.Because only the ground area is photographed,the probability of moving targets in the field of view can be reduced,especially indoors,which is more obvious.The problem of no feature points to follow can be avoided by using the active marking method.The prior range of gray difference can avoid the strong assumption that the gray level between frames of optical flow method is constant,and also avoid the interference caused by light.Then through the center direction traversal fast feature point matching algorithm or center circle traversal fast feature point matching algorithm proposed in this paper,the feature point matching between frames can be carried out quickly and accurately,thus greatly reducing the amount of calculation.Due to the inherent characteristics of the sampling system,the interference between feature points can be avoided according to the prior topological structure,and the whole image does not need to be traversed in every frame matching like the feature point method.Only a small range of traversal around the gray image of hot spot can complete the matching according to the direction matching strategy or the circle matching strategy.In view of the problem that the multiplicative noise interference caused by the visual odometer is related to the process noise and the observation noise,the Multiplicative Noises and Additive Correlated Noises Cubature Kalman Filter(MACNCKF)is proposed.The visual odometer and the inertial navigation system are fused and estimated by the loose coupling method.Taking the error of attitude angle and gyroscope deviation as the newly added state variable,the state equation and observation equation are set up to filter,and the obtained error estimate value is output to the robot's motion state information obtained by the visual odometer for correction,at the same time,the deviation of gyroscope in the IMU is online compensated,and the attitude angle estimation information can be directly submitted to the main control system for use.The correctness of the above method is verified by simulation.Next,a robust federated filter is proposed to fuse the local optimal estimation information of the position and velocity gained by the two local sub filters to obtain the global optimal estimation.The process noise of sub filters is identified online by Sage-Husa method.By improving the calculation method of information allocation factor of non feedback federated filter,the above two local sub filters are improved.The local optimal estimation information of position error and velocity error is fused to get the global optimal estimation.This method has the function of adaptive fault tolerance,which can quickly and correctly check the fault of the sub filter and isolate in time to avoid the pollution of the global data.The method can improve the fusion estimation accuracy and robustness of the whole motion state estimation system,and ultimately send the global optimal state estimation of the position and velocity of the quadruped robot to its main control system for closed-loop control.The correctness of the above method is verified by simulation.Finally,taking LS3 and Spot Mini as reference,we design the heavy and light quadruped robot models with obvious differences in shape and size are used to simulate the rough road surface and flat road surface,and the actual quadruped robot walking experiment is carried out.The experimental results show the adaptability and feasibility of this method,which can provide guarantee for the smooth motion control of quadruped robot.