Due to the aging of skilled farm-workers and the difficulties of finding the successors, it is now a concern whether the agricultural industry in Japan can continue to exist. Due to these situations efforts are being made to solve this problem. One of the solutions is the robotization of labor. With the progress of sensor and computer technology, this is a practical approach. Recently, studies on the robotization of agricultural tractors have been conducted in Japan as well as in other developed countries.

In Japan, most of the farm lands are on the mountainous and hilly areas. However, present researches on tractor navigation concentrate mainly on flat land. Therefore, the automatic tractor guidance on sloped terrain should be investigated. Torisu et al. (2002) developed a NN vehicle model for slope-land and designed a controller for the tractor navigation along contour line. However, a number of things must be studied to improve the automatic tractor guidance on slope.

The main objective of this thesis is to develop a wheeled tractor-like robot for the sloped terrain. The structure of the study is broken down into the following components and procedures:

　Chapter II of this study deals exclusively with the study of the dynamic characteristics of vehicle motion on sloped terrain so as to justify the rationality of NN vehicle modeling for the slope. Different types of open-loop vehicle motions were applied on a prototype vehicle in the real environment and the behavior of each vehicle motion was observed. It was found that in all cases of the rectilinear, steady state circular turning and slalom motions, the vehicle motion on slope was non-linear. Therefore, conventional vehicle modeling for sloped terrain is impractical. A NN vehicle model was developed for sloped terrain and trained along a circular path on 10° sloping ground. Simulations were done for those tested motions with the NN vehicle model. The simulated results were in good agreement with all the actual tested motions and validated the NN vehicle model for sloped terrain.

　Chapter III discussed about the designing of a rectangular path planner for the vehicle navigation on sloped terrain. The steps taken for this purpose were (a) designing of a coordinate transformation system, (b) developing a feed forward control method for the quarter circular turns, (c) compounding the feedback control method of the contour line navigation and feed forward control method for turns, (d) applying the compound control method for vehicle navigation along different directions of the rectangular path, and (e) implementing an autonomous travel test with the developed navigation planner on 18° sloping ground.

　Chapter IV deals with the generalization of the individual controller of each slope land into a single controller for all slope lands. In the studies discussed up to last chapters, the look-up tables were being used to navigate the tractor on slope. An individual table was needed to navigate the vehicle on each land-inclination. Finding suitable test field to prepare a look-up table for each degree of land-inclination is difficult. Therefore, only four look-up tables of optimal steering angle for 0°, 5°, 10° and 15° sloped terrains were prepared by doing field tests, and those tables were then generalized to a single table by developing a cubic polynomial matrix. Field test was done with the developed generalized matrix controller on 20° sloping land and found that the method was effective.

　Chapter V deals with (a) finding an optimal turning path on slope, (b) developing a feedback controller for the turns, and (c) determining the minimum turning radius for each turn of the rectangular path on sloped terrain. To find the optimal turning path (1) a cost function was designed to optimize the final vehicle posture, and (2) simulations were done by genetic algorithm (GA) using the NN vehicle model and the cost function. In developing the feedback controller for the turns, a geometrical relationship was developed for determining the lateral and heading deviations of the vehicle with respect to the circular turning path. According to the deviations, the generalized or the separate look-up tables developed for the rectilinear motion can be used to navigate the vehicle along the curved path. Simulations with the NN vehicle model and the feedback control method for the turns were conducted to find the minimum turning radius for each turn of the rectangular path.

　Chapter VI was devoted to develop a NN based steering controller to obtain the optimal steering for each micro-level deviation of vehicle on slope. A multi-layered NN controller was designed for this purpose, where the input vector consists of the land-inclination, lateral deviation and heading deviation of the vehicle, and the output vector consists of only the optimal steering. The back-propagation algorithm was used to train the controller. The training data was taken from the look-up tables of optimal steering angle for 0°, 5°, 10° and 15° sloped terrains. The controller performance was validated by simulation and field tests. In the field tests, both feedback and feed forward control methods for the quarter circular-turns were applied alternatively and the performance of the feedback control method was compared with that of the feed forward control method.