Learning Human-robot Handover Trajectory by Imitation Based on Multi-level Kernelized Movement Primitives

Aiming at the problem of robot trajectories reproduction and generalization for human-robot handover tasks, a one-shot imitation learning method for human-robot joint trajectories is proposed based on multi-level kernelized movement primitives. By transferring the non-linear relationship between the positions of the human hand and the robot end-effector, which is learned in a spatial sub-region, to other sub-regions, the human-robot handover skills can cover the overall interaction workspace. By modeling and modulating the length of robot end-effector motion trajectories, the algorithm can adapt to the length variance of robot end-effector trajectory during trajectory reproduction. Experiments on a UR5 robot show that, compared with classical kernelized movement primitives algorithm, the proposed multi-level kernelized movement primitives algorithm avoids the uniform sampling of the complete human hand position space and eliminates the interference of sampling points from other sub-regions during inference procedure of hand position in a certain sub-region. Therefore, the average inference error is reduced from 7.11 cm to 1.85 cm, and the average inference time consumption is shortened from 0.138 s to 0.015 s, which achieves higher accuracy and better real-time performance. The proposed algorithm improves handover success rate from 13.3% to 100% on the hand-robot joint trajectories with significant scale variance, which outperforms classical kernelized movement primitives algorithm in the learning ability and adaptability.