Research On Terrain Perception And Static Gait Planning For The Quadruped Robot Based On TOF Camera

Compared to the wheeled and tracked robot,legged robot has discontinuous motioncharacteristics,which possesses a wide application prospect in the domain of military ac-tion, emergence rescue etc. Since the quadruped robot Big Dog and Little Dog came outin Boston Dynamics, the research upsurge on the quadruped robot has been raised in allof the world. It is very challenging for the quadruped robot walking in the rough terrainenvironment, because of its feature of structure and motion style. In this paper, accordingto the need of autonomous navigation for the quadruped robot in rough terrain conditions,both the terrain perception algorithm based on Time Of Flight(TOF) camera and the mo-tion gait planning algorithm for the quadruped robot based on the map from TOF camerawill be thoroughly researched, which will provide the solution for improving the motionperformance of the quadruped robot in the complex environment.Three dimensional point cloud map from TOF camera is necessary for planning therobot, but the cloud from only one frame can not meet the navigation need because ofits smaller view. So this paper starts with registering two point clouds captured at thetwo positions,respectively, after fully analyzing the advantage of the TOF camera, thealgorithm for registering two clouds based on the gray scaled image will be researched.As soon as the contrast of the image is improved by preprocessing, the algorithm fordetecting and matching scale invariant features from gray scaled image and the algorithmfor filtering invalid pairs of 3D features will be researched. Iterative Closest Point(ICP)with initial value will be researched in the basis of estimating the initial relative posebetween the two frames.The 6 DOF pose of the quadruped robot will be located by capturing TOF frameswhen the robot is walking, so multiple frames will be captured continuously, if only theadjacent frames are applied to location, the error accumulation will come out. Pose opti-mization by multiple frames based on graph structure will be researched, whose premiseis generating the pairs of 3D points through frame to frame registration. The effectivenessof pose optimization is directly determined by the accuracy of frame to frame registra-tion, so the paper proposes a hierarchical index structure for increasing the precision ande?ciency of frame to frame registration. The precision for estimating relative pose be-tween two frames is also needed for pose optimization, so the closed form algorithm forestimating the precision based on implicit function is researched after frame to frameregistration.In order to walk in the rough terrain environment, the quadruped robot must havethe ability of adjusting the foothold, or it will yield foot sliding etc. 2.5D elevation mapbased on grids will be created after processing the point cloud from TOF camera, and thegrid is the smallest unit for foothold selection. The default foothold is computed afterthe robot model is created. After computing the normal and curvature, the cost functionfor the foothold is built by getting foot terrain characteristics and the default locationtogether. The foothold in the rectangle whose center is the default point is representedby 11-dimension vector, and its cost value is acquired by computing the inner productbetween the weight vector and the feature vector. The algorithm for computing the weightvector based on machine learning will be researched, and the expert guide the footholdsorting, then the support vector machine(svm) will be used to learn the weight vector onthe training data.In addition to the ability of foothold selection, the quadruped robot should be ableto select movement path and to plan the gait for walking in the complex terrain. 2-dimension robot center path selection will be researched, three terrain features includingstep height,slope and roughness will be detected from the coverage area by the robot base,the whole cost map for selecting path is generated by considering with the average cost ofthe default footholds, and 2D body center path will be searched based on the whole costmap by designing the heuristic function. The algorithm for searching the series of staticgait is proposed based on 2D center path, using recursive backtracking strategy to multiplesteps ahead. The robot state for each gait is determined by foothold,roll angle,pitch angleand the distance between the robot center point and the ground, and then the reachabilityof the next gait is estimated, including collision detection and The kinematic constraintsatisfaction measurement etc.Finally, the vision perception algorithm based on TOF camera and the static gaitplanning strategy for the quadruped robot are integrated into the real quadruped robot.After the relative pose transformation between TOF camera and robot center is calibrat-ed, the contrast experiment of linear walking for the robot with and without TOF camerais carried out, which proves the effectiveness of the location algorithm through TOF cam-era. Then the experiment that the quadruped robot stride over the wooden step is carriedout,which illustrates the feasibility of terrain perception,location,foothold selection andstatic gait planning algorithms applied to the real quadruped robot,and these algorithmsshow good stability and adaptability.