Hierarchical kernelized movement primitives for learning human-robot collaborative trajectories in referred object handover

While research in Learning-by-Demonstration (LbD) methods has made significant progress, learning human-robot collaborative trajectories has been a challenging task. In this paper, a hierarchical kernelized movement primitives (KMPs) model is proposed for learning human-robot handover trajectories. The model learns the non-linear correlations between human hand positions and robotic end-effector positions within a baseline subregion and generalizes them to other subregions. Therefore, the proposed hierarchical KMPs achieve one-shot cross-subregion trajectory skill transfer, which is the major advantage over the classic KMPs. The benefit of such an improvement is the generalization capability of the robot's reactive trajectory in the whole workspace while avoiding collecting a large number of trajectory samples. In addition, we also present a trajectory scaling and modulation method to ensure the adaptation and generalization of joint trajectory scales. This enables the adaptation of the robot's movement toward trajectory scales. Experiments of human-robot handover in a robot-assistive assembly scenario demonstrate the performance gains over the classic KMPs in trajectory prediction accuracy. In addition, the results also validate that the proposed method is capable of one-shot transfer across subregions and adaptation to the trajectory scale variance.