Abstract

PENA-SOLORZANO, Carlos Andrés; HOYOS-GUTIERREZ, José Gabriel  and  PRIETO-ORTIZ, Flavio Augusto. GENERALIZATION OF THE TRAJECTORIES OF A ROBOTIC ARM USING DYNAMIC MOVEMENT PRIMITIVES AND GAUSSIAN PROCESS REGRESSION. Rev.EIA.Esc.Ing.Antioq [online]. 2018, vol.15, n.29, pp.109-123. ISSN 1794-1237.  https://doi.org/10.24050/reia.v15i29.690.

It is common to find robots performing tasks in areas shared with humans, where they are expected to be able to learn from the actions taken by others and adapt to new situations. In this paper, the trajectories of the arm of an operator are captured while moving to grab an object, making joint tracking with the Microsoft kinect sensor. The technique used for the coding of the training signals is called dynamic movement primitives (DMP), while reconstruction is performed using gaussian process regression (GPR). GPR also allows generalize training movements to new paths when changing both the initial hand position and the location of the object. The technique is compared against a generalization based on Mahalanobis distance and gaussian distribution, which uses the path data to make the estimate. The proposed technique presented low encoding times and small errors regarding the target values when tested with 30 points of consultation for the initial value of the hand, and 30 points for the final position.

Keywords : Robotics; learning by imitation; programming by demonstration; dynamic movement primitives; gaussian process regression.

        · abstract in Spanish | Portuguese     · text in Spanish     · Spanish ( pdf )