Model identification and model analysis in robot training

Robot training is a fast and efficient method of obtaining robot control code. Many current machine learning paradigms used for this purpose, however, result in opaque models that are difficult, if not impossible to analyse, which is an impediment in safety-critical applications or application scenarios where humans and robots occupy the same workspace. In experiments with a Magellan Pro mobile robot we demonstrate that it is possible to obtain transparent models of sensor-motor couplings that are amenable to subsequent analysis, and how such analysis can be used to refine and tune the models post hoc.