On-line Motion Prediction and Adaptive Control in Human-Robot Handover Tasks

Abstract

Handover tasks are commonly seen in daily life of humans. However, it is challenging in robot applications since continuous motion prediction and adaptation is required by each participant. In this paper we present a system that allows a human operator and a robot to perform seamless and safe handover tasks in a shared work space. The proposed system consists of a human motion predictor, a motion planner and a low-level joint torque controller. The predictor generates multi-step prediction of a reaching motion based on on-line Gaussian Process regression and only requires position measurement of the hand. Depending on the prediction, the motion planner creates smooth and human-like trajectories that synchronize with the human motion. A safety-orient joint torque controller is designed on the basis of impedance control and on-line adapted by monitoring a danger index that takes prediction uncertainties into account. Firstly, performance of the motion predictor is analyzed in a simulation, based on recorded human motion data. Then the whole system is evaluated in a real handover experiment with a FRANKA EMIKA robot. Results show that the proposed system delivers reliable on-line human motion prediction even for a long prediction time and gives consideration of both fluency and safety in the handover process.