Learning Robot Motion from Creative Human Demonstration

This thesis presents a learning from demonstration framework that enables a robot to learn and perform creative motions from human demonstrations in real-time. In order to satisfy all of the functional requirements for the framework, the developed technique is comprised of two modular components, which integrate together to provide the desired functionality. The first component, called Dancing from Demonstration (DfD), is a kinesthetic learning from demonstration technique. This technique is capable of playing back newly learned motions in real-time, as well as combining multiple learned motions together in a configurable way, either to reduce trajectory error or to generate entirely novel motions based on user specified parameters. DfD was utilized to enable a cooperative robot-human dance performance, and that performance has been evaluated to demonstrate that DfD achieved its design goals. The second component of this newly developed robot control framework, called Pose Energy Correspondence Mapping (PECM), is a passive-observation based learning from demonstration technique which is used to convert human pose trajectories into corresponding robot joint trajectories. This conversion process makes use of an energy based neural network model in order to attempt to achieve high quality results with a minimally sized training set. These two components have been combined together and the resulting framework has been evaluated by means of a comparative survey between human-generated robot motions, PECM-generated robot motions, and robot motions generated by a baseline neural network technique. These survey results are analyzed and discussed in order to identify the strengths and limitations of the newly developed framework