Robot Orientation Learning Based on Interaction Primitives for Human-Robot Collaboration

In some scenarios of Human-Robot Collaboration, it is more convenient to express the synergies in task space, where the robot positions and orientations are taken into consideration. Interaction Primitives (IP) stand out as an effective approach that encodes human and robot movement in collaboration. However, IP can not be directly exploited to learn the robot orientation trajectories represented by unit quaternions because the unit quaternions lie on the manifold ${\mathcal{S}^3}$, which is not Euclidean. In this paper, we propose an approach built on Interaction Primitives and Riemannian manifold to realize robot orientation trajectory learning and retrieving in collaboration. This work is validated in the real human-robot collaborative pouring task and the experiment that compared with the original Interaction Primitives and Gaussian Mixture Model on the quaternions manifold.