Eliminating sensor ambiguities via recurrent neural networks in sensor-based learning

This paper presents a state identification approach which eliminates ambiguities caused by sensing and uncertainty. In manipulation tasks, the identification of contact states based on force and position sensing is affected by ambiguities. The approach uses recurrent neural networks to learn an internal representation of the finite state automata defined by the sensor patterns. The technique is demonstrated using a simulated learning task. State identification is combined with other techniques in a sensor-based learning architecture for robotic manipulation. Results are presented for simulated tasks. The system is able to manage ambiguous states which previously were impossible to learn