RESUMO
In this paper, we present a strategy for a legged robot to stably cross cinder blocks with a limited area acquired from a camera. First, we used the point cloud acquired from the camera to detect the planes and calculate their centroids and directions. This information was used to determine the position and direction of the foot to which the robot should go. Existing A*-based footstep planners require a global map to reach the goal from the start and do not generate a path if there is no solution to the goal due to completeness of A*. In addition, if the map is not updated while moving the path, it is vulnerable to changes in the object position. Our strategy calculates the footsteps that the robot can walk in a limited camera area without securing a global map. In addition, it updates the local map information every walking step so that it quickly recognizes nearby objects and finds a path that can move. While the robot is walking, objects may not be detected due to the narrow camera field of view. In addition, even if an area for the robot to land is found, a situation in which the robot's legs collide may occur. We present a strategy to solve this problem using previous landing data. In the experimental environment composed of several patterns, the performance was verified by stably walking on the blocks without collision between the robot's legs.
