Learning to improve path planning performance

In robotics, path planning refers to finding a short. collision-free path from an initial robot configuration to a desired configuratioin. It has to be fast to support real-time task-level robot programming. Unfortunately, current planning techniques are still too slow to be effective, as they often require several minutes, if not hours of computation. To remedy this situation, we present and analyze a learning algorithm that uses past experience to increase future performance. The algorithm relies on an existing path planner to provide solutions to difficult tasks. From these solutions, an evolving sparse network of useful robot configurations is learned to support faster planning. More generally, the algorithm provides a speedup-learning framework in which a slow but capable planner may be improved both cost-wise and capability-wise by a faster but less capable planner coupled with experience. The basic algorithm is suitable for stationary environments, and can be extended to accommodate changing environments with on-demand experience repair and object-attached experience abstraction. To analyze the algorithm, we characterize the situations in which the adaptive planner is useful, provide quantitative bounds to predict its behavior, and confirm our theoretical results with experiments in path planning of manipulators. Our algorithm and analysis are sufficiently, general that they may also be applied to other planning domains in which experience is useful