Learning strategies for sensor-based manipulation tasks

An architecture that incorporates a seamless integration of different learning paradigms is presented. Sensor processing, recurrent neural networks, learning from experience and qualitative knowledge are the key elements of the system. The goal applications are those tasks which cannot be fully programmed due to uncertainties and incomplete knowledge. The proposed sensor-based architecture combines several learning paradigms as well as pre-programmed modules, since experimental evidence suggests that some paradigms are more convenient for learning certain skills. The correspondence between qualitative states and actions is learnt. The qualitative treatment of information makes it suitable for the analysis of system behavior, knowledge extraction and generalization to other more complex tasks. Programming is used to decrease the complexity of the learning process. This general approach is a suitable scheme for a wide range of robot situations. Results are provided for the simulation of a sensor-based goal-finding task as well as for a real application of the architecture in a robotic insertion process in three dimensions