An Improved Dual Quaternion Dynamic Movement Primitives-Based Algorithm for Robot-Agnostic Learning and Execution of Throwing Tasks

Abstract Inspired by human nature, roboticists have conceived robots as tools meant to be flexible, capable of performing a wide variety of tasks. Learning from demonstration methods allow us to “teach” robots the way we would perform tasks, in a versatile and adaptive manner. Dynamic movement primitives (DMP) aims for learning complex behaviors in such a way, representing tasks as stable, well-understood dynamical systems. By modeling movements over the SE(3) group, modeled primitives can be generalized for any robotic manipulator capable of full end-effector 3D movement. In this article, we present a robot-agnostic formulation of discrete DMP based on the dual quaternion algebra, oriented to modeling throwing movements. We consider adapted initial and final poses and velocities, all computed from a projectile kinematic model and from the goal at which the projectile is aimed. Experimental demonstrations are carried out in both a simulated and a real environment. Results support the effectiveness of the improved method formulation.