Using incremental learning for transition from rule based behavior to motion sequences for motor control of a flexible robot system

In the context of service robotics very often no model of the environment exists. The actual situation can partially be recognized by sensors as a snapshot, which is then used for planning an action; the optimal trajectory, e.g., in manipulation tasks, to reach a desired destination is often very difficult to compute because not all aspects were captured by the sensors. Therefore, systems with rule based behavior can be used to autonomously plan the motion online using mainly tactile sensor information. For repeating the same or a similar task, it would be nice to have motion sequences allowing much faster motions with less collisions. The problem we address in this paper is the transition from rule based behavior towards motion sequences without losing the flexibility of the system, which is necessary to adapt a motion sequence to the actual situation and to deal with unforeseen events during execution.