Learning model based control of lightweight robotic arms

Robots are slowly entering into the unstructured environments populated with humans where traditional control techniques, applied in industrial settings, are becoming inadequate. Reasons for this are that we can not pre-progam tasks in advance reliably and the usual assumption of rigid body dynamics is violated for most lightweight robotic arms. Learning techniques allow for learning models form data only, with no assumptions on structure which allows for learning of nonlinearities and other dynamics that are hard to take into account explicitly. As the data in robots are usually available through sensors readings supervised techniques are in the focus of this paper with application to robotic manipulator modeling and control.