No description available.

Agents with differential constraints, although common in the real world, pose a particular difficulty for motion planning algorithms. Methods for solving such problems are still relatively slow and inefficient. In particular, current motion planners generally can neither "see" the world around them, nor generalize from experience. That is, their reliance on collision tests as the only means of sensing the environment yields a tactile, myopic perception of the world. Such short-sightedness greatly limits any potential for detection, learning, or reasoning about frequently encountered situations. In result these methods solve each problem in exactly the same way, whether the first or the hundredth time they attempt it, each time none the wiser. The key component of this thesis proposes a general approach for motion planning in which local sensory information, in conjunction with prior accumulated experience, are exploited to improve planner performance. The approach relies on learning viability models for the agent's "perceptual space", and the use thereof to direct planning effort. In addition, a method is presented for improving runtimes of the RRT motion planning algorithm in heavily constrained search-spaces, a common feature for agents with differential constraints. Finally, the thesis explores the use of viability models for maintaing safe operation of user-controlled agents, a related application which could be harnessed to yield additional, more "natural" experience data for further improving motion planning.

Agents with differential constraints, although common in the real world, pose a particular difficulty for motion planning algorithms. Methods for solving such problems are still relatively slow and inefficient. In particular, current motion planners generally can neither "see" the world around them, nor generalize from experience. That is, their reliance on collision tests as the only means of sensing the environment yields a tactile, myopic perception of the world. Such short-sightedness greatly limits any potential for detection, learning, or reasoning about frequently encountered situations. In result these methods solve each problem in exactly the same way, whether the first or the hundredth time they attempt it, each time none the wiser. The key component of this thesis proposes a general approach for motion planning in which local sensory information, in conjunction with prior accumulated experience, are exploited to improve planner performance. The approach relies on learning viability models for the agent's "perceptual space", and the use thereof to direct planning effort. In addition, a method is presented for improving runtimes of the RRT motion planning algorithm in heavily constrained search-spaces, a common feature for agents with differential constraints. Finally, the thesis explores the use of viability models for maintaing safe operation of user-controlled agents, a related application which could be harnessed to yield additional, more "natural" experience data for further improving motion planning.