부산대학교 도서관

Martin Marietta's multi-year Independent Research and Development effort to develop intelligent, high-level AUV control capabilities for demonstration aboard the Mobile Undersea Test (MUST) Laboratory has produced a robust approach for planning and executing waypoint transiting in a complex and dynamic ocean scene. This paper presents the architecture and techniques implemented in the high-level control system and preliminary results obtained during evaluations conducted in the MUST Dry Lab. Missions anticipated for a MUST-like AUV will require the vehicle to arrive at specified points at designated times, and the execution of various activities at certain times or places, or under certain conditions. To accomplish such behaviors, the high-level controller must pilot the AUV through potentially complex and uncharted obstacle fields and terrain features. In the course of traversing an area, the high-level controller should be able to "learn" its environment for future use should it be required to re-traverse the same area. Our approach to intelligent waypoint transiting begins with a set of desired waypoints scattered within a complex natural terrain. Given this top level requirement, our high-level controller is capable of autonomously piloting the AUV to each waypoint, while adhering to order and time requirements specified by a user or an activity scheduling program. Range and bearing information required by the controller is provided by a multi-beam sonar system that is itself under the direction of the controller. While planning optimal transits to various waypoints, the controller reasons about the AUVs current state to conserve fuel, avoid counter detection, and avoid obstacles; while maintaining pre-specified keep-out distances. The controller is constructed around a blackboard architecture and utilizes forward-chaining rules under "meta-level" supervision. Obstacle avoidance is performed using algorithms that analyze the geometry of sensed obstacles. Specific algorithms are invoked by the rules. Knowledge about the environment that is "learned" in the process of achieving waypoints is used to find the "best" paths to subsequent waypoints. Since the sonar range is finite, an A* algorithm is employed to operate on a Learned Visibility Graph (LVG). A path relaxation technique is then utilized to optimize the path. Timely computations and an ability to commence transiting towards a waypoint without undue delays is assured through inferencing performed by the meta-level supervisor.