Using Programming by Demonstration and Dynamical Systems for Robot Learning

This work presents the methodology used for implementing programming by demonstration methods on a Universal Robot UR10. We used non-linear time-invariant Dynamical Systems that can imitate unconventional forms of motion. Gaussian Mixture Models were used to model input data and parameters of DS were optimized based on MSE and Log-likelihood cost functions. We incorporated a Kinect to detect users and ensure safety. Our system learns any non-circular path within three manual demonstrations and was spatially asymptotically stable.