Robot motion planning using real-time heuristic search

Autonomous mobile robots must be able to plan quickly and stay reactive to the world around them. Currently, navigating in the presence of dynamic obstacles is a problem that modern techniques struggle to handle in a real-time manner, even when the environment is known. The solutions range from using: 1) sampling-based algorithms which cut down on the shear size of these state spaces, 2) algorithms which quickly try to plan complete paths to the goal (to avoid local minima) and 3) using real-time search techniques designed for static worlds. Each of these methods have fundamental flaws that prevent it from being used in practice. In this thesis I offer three proposed techniques to help improve planning among dynamic obstacles. First, I present a new partitioned learning technique for splitting the costs estimates used by heuristic search techniques into those caused by the static environment and those caused by the dynamic obstacles in the world. This allows for much more accurate learning. Second, I introduce a novel decaying heuristic technique for generalizing cost-to-go over states of the same pose (x. y.theta.v) in the world. Third, I show a garbage collection mechanism for removing useless states from our search to cut down on the overall memory usage. Finally, I present a new algorithm called Partitioned Learning Real-time A*. PLRTA* uses all three of these new enhancements to navigate through worlds with dynamic obstacles in a real-time manner while handling the complex situations in which other algorithms fail. I empirically compare our algorithm to other competing algorithms in a number of random instances as well as hand crafted scenarios designed to highlight desirable behavior in specific situations. I show that PLRTA* outperforms the current state-of-the-art algorithms in terms of minimizing cost over a large number of robot motion planning problems, even when planning in fairly confined environments with up to ten dynamic obstacles