LAT: A simple Learning from Demonstration method

Abstract

Learning from Demonstration (LfD) is a powerful method for training robots to solve tasks involving low level motion skills, thus avoiding human programming effort. We present Learning from Demonstration by Averaging Trajectories (LAT) which is a new, simple and computationally fast method and provide an implementation on a service robot. We compare LAT theoretically as well as empirically to LfD with Gaussian processes (GP) and to LfD with dynamic movement primitives (DMP). It turns out that LAT is as powerful as Gaussian processes, computationally faster than ordinary GPs and comparable to local GPs. The comparison of LAT to DMPs shows that LAT is able to detect constraints and thus can learn abstract concepts which DMPs can not. DMPs on the other hand can dynamically react to changing object positions which LAT and GPs can not. This gives rise for future work on a combination of LAT and DMPs.