Robot task planning and trajectory learning based on programming by demonstration

Abstract

This paper presents a method to model and reproduce cyclic trajectories captured from human demonstrations. Heuristic algorithms are used to determine the general type of pattern, its parameters, and its kinematic profile. The pattern is described independently of the shape of the surface on which it is demonstrated. Key pattern points are identified based on changes in direction and velocity, and are then reduced based on their proximity. The results of the analysis are provided are used inside a task planning algorithm, to produce robot trajectories based on the workpiece geometries. The trajectory is output in the form of robot native language code so that it can be readily downloaded on the robot.