Publikation
Comparison of Sensor-Feedback Prediction Methods for Robust Behavior Execution
Christian Rauch; Elmar Berghöfer; Tim Köhler; Frank Kirchner
In: KI 2013: From Research to Innovation and Practical Applications. German Conference on Artificial Intelligence (KI-13), September 16-20, Koblenz, Germany, Pages 200-211, ISBN 978-3-642-40941-7, Springer, Berlin - Heidelberg, 9/2013.
Zusammenfassung
Robotic applications in inaccessible environments like in space strongly depend on detailed planning in advance as there are only short communication windows, a high latency in communication, and as there is often no way of recovering the system when it gets into a fault state. Furthermore, unknown terrain requires continuous monitoring of behavior execution by a human operator. The effort on detailed planning and especially the delay through remote monitoring can be decreased by supporting the autonomy of the robot by predicting and self-monitoring behavior consequences. Presented are three approaches for creating prediction models. The models are used to generate expectations on sensor feedback caused by given actions. The expected sensor feedback is compared with the actual sensor feedback through a monitoring stage that will trigger a change of the robot behavior in case of unexpected sensor output. Two function fitting approaches (analytic model and generic function approximation) and a vector quantization method are compared with each other. The
evaluation of the triggering mechanism in real scenarios will show that the execution of emergency actions in unexpected situations is possible.