Human Motion Prediction in Human-Robot Handovers based on Dynamic Movement Primitives

Human-robot handovers can be made more seamless by predicting handover place and time on-line as soon as the human agent initiates a handover process. We consider the prediction problem as a model-based parameter estimation problem where the point attractor and the timescale of human hand motion are estimated on-line. Using dynamic movement primitives as a parameterization of human motion, its point attractor and timescale are successfully estimated on-line using an extended Kalman filter. Convergence of the parameter estimates is shown and the performance of the proposed predictor is evaluated using generated trajectories as well as experimental data of human-human handovers. Thanks to the good prediction of the handover place, the presented algorithm can be used to improve human-robot collaboration.