In order to solve the problems of low planning efficiency, many polyline segments of the planning path, large polyline angle and unstable operation of agricultural robots in the process of global path planning, a path planning method based on improved A* algorithm and low-order multi-segment Bezier curve splicing (LM-BZS) algorithms was proposed by taking the orchard crawler robot as the kinematic model. To begin with, the orchard environment information was obtained according to the prior map, the fruit trees and the obstacles were regarded as impassable regions, and the impassable regions were expanded and fitted according to the dimensions of the robot body. And then, the improved A* algorithm was used to search for the path, and the tree row nodes were adjusted for the preliminary generation path. In the end, the LM-BZS algorithm was used to optimize the adjusted path points to generate a driving path that meets the operation requirements of the orchard crawler robot. The simulation results manifested that compared with the traditional A* algorithm, the improved algorithm proposed reduced the path planning time by 76.75% and 86.40%, and the number of evaluation nodes by 36.68% and 39.37%, respectively in the barrier-free and obstacle environments. In the barrier-free environment, the average curvature of the path optimized by the LM-BZS algorithm was reduced by 45.81% and 18.94% compared with that of the traditional A* algorithm and the high-order Bezier curve algorithm, respectively, and the average curvature was reduced by 56.98% and 27.81% compared with that of the traditional A* algorithm and the higher-order Bezier curve algorithm in the obstacle environment. The field test results manifested that in the barrier-free and obstacle environment, the maximum lateral error was 0.428m and 0.491m, the average lateral error was 0.232m and 0.276m, and the average course deviation was 11.06° and 13.76° respectively, which was in line with the autonomous driving conditions of the orchard crawler robot.