A Direction-Adaptive and Uncertainty-Weighted Pose Dynamic Movement Primitives Framework for Robot Skill Reproduction

Abstract

Dynamic Movement Primitives (DMPs) are widely used for robot skill reproduction from demonstrations, but pose trajectory reproduction for continuous manipulation tasks remains challenging because translational and rotational motions must be represented consistently while maintaining accuracy, disturbance recovery, and terminal smoothness. Existing screw-displacement pose DMPs provide a geometrically consistent formulation on SE(3); however, their isotropic fixed-gain feedback limits direction-dependent correction, and deterministic forcing terms do not provide an explicit estimate of prediction reliability, which may cause over-shaping and high terminal jerk. This paper proposes a direction-adaptive and uncertainty-weighted pose DMP framework for robot skill reproduction from pose trajectories obtained from demonstrations. A Riemannian Motion Policy (RMP)-inspired direction-adaptive feedback mechanism is introduced to adjust recovery and damping according to the current pose error directions and magnitudes, improving trajectory-level correction and disturbance recovery. In addition, a Sparse Spectrum Gaussian Process (SSGP) is used to model the forcing term probabilistically, and its predictive variance is combined with a phase-dependent gate to attenuate low-confidence forcing contributions, particularly near the terminal phase. Simulation studies on RoboMimic trajectories show that the RMP-inspired feedback primarily improves pose reproduction accuracy and disturbance recovery, whereas the SSGP-based weighting substantially reduces terminal translational and rotational jerk, with a slight accuracy compromise relative to RMP-DMP. A papermaking robot case study further demonstrates the deployment feasibility of the generated pose trajectories on a real continuous-operation platform.