On learning task-directed motion plans

Abstract

Robotic motion planning is a hard problem for robots with more than just a few degrees of freedom. Modern probabilistic planners are able to solve many problems very quickly, but for difficult problems, they are still unacceptably slow for many applications. This thesis concerns the use of previous planning experience to allow the agent to generate motion plans very quickly when faced with new but related problems. We first investigate a technique for learning from previous experience by simply remembering past solutions and applying them where relevant to new problems. We find that this approach is useful in environments with very low variability in obstacle placement and task endpoints, and that it is important to keep the set of stored plans small to improve performance. However, we would like to be able to better generalize our previous experience so we next investigate a technique for learning parameterized motion plans. A parameterized motion plan is a function from planning problem parameters to a motion plan. In our approach, we learn a set of parameterized subpaths, which we can use as suggestions for a probabilistic planner, leading to substantially reduced planning times. We find that this technique is successful in several standard motion planning domains. However, as the domains get more complex, the technique produces less of an advantage. We discover that the learning problem as we have posed it is likely to be intractible, and that the complexity of the problem is due to the redundancy of the robotics platform. We suggest several possible approaches for addressing this problem as future work.