Abstract: All pineapples are harvested manually at present in China. But manual harvesting cannot fully meet the large-scale production against the ever-increasing greying workforce. Fortunately, automatic navigation can be expected to develop the pineapple harvester. In this study, a path planning algorithm was proposed as the navigation scheme to improve the mechanization and automation level of pineapple harvesting. An improved RRT* algorithm was also used for global path planning. Firstly, the self-heuristic idea was used to constrain the generation range of sampling points. Then, the bias probability pbias was introduced to generate the random sampling points. Specifically, the sampling points were randomly generated with the probability p in the space, when p>pbias. Otherwise, the target point was used as the sampling point, in order to decrease the blindness of sampling point generation. Thirdly, the gravitational field of the artificial potential field, and the concept of direction weight were introduced in the new node expansion. The weights wg and wk were assigned to the directions of the sampling and target point, respectively, where the direction was constrained in the expansion of the new node. Fourthly, the bidirectional expansion was used to speed up the iteration speed in the double-tree expansion. Finally, the greedy algorithm was applied to prune the redundant nodes of the path. The Cantmull-Rom interpolation function was also used to smooth the path corners. Three environments (including multiple obstacles, mazes, and narrow passages) were created to simulate the path planning process, in order to evaluate the performance among the improved navigation path planning, RRT* and bidirectional RRT* algorithm. Planning time, node number, and path length were selected as the indicators. Each algorithm experimented with 30 times in every single environment. The average, maximum, minimum, and standard deviation were calculated for the simulation data of the three indicators, respectively. The simulation results showed that the average planning time of the path planning algorithm of this paper in the three environments was 18%, and 46.12% higher than that of the RRT* and bidirectional RRT* algorithms, respectively, while the average programming speed was 5.7, and 2.3 times as rapid as that of the RRT*, and bidirectional RRT* algorithm, respectively. Furthermore, the average node number was 87.23% and 52.52% less than that of the RRT* and bidirectional RRT* algorithms, respectively. The average path length was 3.81% and 6.08% less than the RRT* and bidirectional RRT* algorithms, respectively. The field test showed that the planning time was only 14.12% and 20.34% of the RRT* and bidirectional RRT*, respectively. The iteration number was 80.89% and 69.70% less than that of the RRT* and bidirectional RRT*, respectively. In addition, the rotation angles larger than 60° on the path planned by RRT* and bidirectional RRT* algorithms were 1.56 and 2.06 times as much as that of the improved, respectively, and the rotation angles larger than 100° on the path were 1.55 and 2.18 times. The improved RRT* algorithm can fully meet the path navigation requirements of agricultural machinery in the field. The pineapple harvester can run along the planned path to the target point as moving with the speed of 0.2, 0.4, and 0.6 m/s, but the position and heading deviation increase with the moving speed. This finding can provide a sound reference for the navigation development in agricultural machines.