A region filled with stationary and non-stationary obstacles always needs a path,based on clustering all the obstacles,for moving of autonomous robots.This paper provides a method,called "two-phase stepwise clustering" to solve this problem.Phase I uses Dendrogram as an approach for the preliminary clustering in the beginning to avoid misleading of the path start.Phase II links piecewise segments from centers of triangles deployed by neural discrimination.In the simulation,two paths constructed from longitudinal deployment and lateral deployment links each other and thus forms a robot path and also an envelope as clustering obstacles.