Several key issues involved with the planning and executing of optimally generated paths for autonomous vehicles are addressed. Two new path planning algorithms are developed, and examined, which effectively minimize replanning as unmapped hazards are encountered. The individual algorithms are compared via extensive simulation. The search strategy results are implemented and tested using the University of Colorado's autonomous vehicle test-bed, RoboCar, and results show the advantages of solving the single-destination all-paths problem for autonomous vehicle path planning.;Both path planners implement a graph search methodology incorporating dynamic programming that solves the single-destination shortest-paths problem. Algorithm 1, termed DP for dynamic programming, searches a state space where each state represents a potential vehicle location in a breadth-first fashion expanding from the goal to all potential start locations in the state space. Algorithm 2, termed DP*, couples the heuristic search power of the well-known A* search procedure (Nilsson-80) with the dynamic programming principle applied to graph searching to efficiently make use of overlapping subproblems. DP* is the primary research contribution of the work contained within this thesis. The advantage of solving the single-destination shortest-paths problem is that the entire terrain map is solved in terms of reaching a specified goal. Therefore, if the robot is diverted from the pre-planned path, an alternative path is already computed.;The search algorithms are extended to include a probabilistic approach using empirical loss functions to incorporate terrain map uncertainties into the path considering terrain planning process. The results show the importance of considering terrain uncertainty. If the map representation ignores uncertainty by marking any area with less than perfect confidence as unpassable or assigns it the worst case rating, then the paths are longer than intuitively necessary.;A hierarchical software control architecture is introduced that uses as the main guidance function an arbitration-based scheme which is able to efficiently and robustly integrate disparate sensor data. The flexibility provided by such an architecture allows for very easy integration of any type of environmental sensing device into the path planning algorithm.
展开▼
