RMAL-ODMPs: Orientation Skills Learning and Generalization Method for Robotic Manipulation

Orientation skills learning and generalization in robotic manipulation tasks play a pivotal role in efficient task planning. In order to provide a sensible region of exploration and the ability to adapt to changes in the robot environment, multiple demonstrations are typically required. However, human operators naturally exhibit variations in angular velocities during each demonstration, which leads to noisy parameter estimation. In this paper, a method called Riemannian manifold arc-length based orientation dynamic movement primitives (RMAL-ODMPs) is proposed. In comparison to orientation dynamic movement primitives (ODMPs), the RMAL-ODMPs generalized orientations are parameterized by arc length instead of time. The arc length is expressed as the geodesic in quaternion manifold space. Therefore, temporal variations in angular velocities do not affect spatial modelling. In addition, a theoretical framework for the accurate learning of multi-orientation trajectories by RMAL-ODMPs is proposed. Subsequently, the weights of RMAL-ODMPs are computed by locally weighted regression (LWR) to solve the optimization problem of orientations. The effectiveness of RMAL-ODMPs is validated through simulations and experiments on the orientation rotation of the robot's end-effector, and the results show that RMAL-ODMPs outperform ODMPs in terms of orientation learning accuracy when learning multiple orientation trajectories under varying angular velocities.