ABSTRACT
Building a world model for manipulators is computationally expensive. If the environment changes over time and no further information is provided, other than that a local misfit has occurred, classical algorithms have to start the model’s computation anew. In order to plan complex trajectories, hierarchical planning shows many advantages. In this article we report on an approach, that hierarchically combines sub-trajectories using symbolic graph search techniques and a sub-symbolic neural classification and generalization preprocessing system. We combine a graph of possible subgoals in configuration space with a mapping from configuration space on to this graph.
