An interactive motion planning framework that can learn from experience

Our framework consists of two main processes: a learning process and a motion generation process. In the former process, the motion segment (a part of motion path) and its related scenario segment (a part of workspace passed through by the object) are gathered, after an interactive motion planning process finds a collision-free motion path or reaches the conclusion of inaccessibility. According to the similarity between the skeletons of scenario segments, the gathered scenario segments and motion segments are organized by a hierarchical structure in the motion library. The latter process permits users to control only one point in the workspace for the selection of a new scenario, and then the similar scenarios are retrieved from the motion library, to help quickly detect the connectivity of the new scenario and generate a repaired motion path to guide users with feasible manipulations. We highlight the performance of our framework on a challenging problem in 2D, in which a non-convex object passes through a cluttered environment filled with randomly shaped and located non-convex obstacles.