DefinePK

Abstract

To achieve autonomy of robotic agents, the estimation or planning of paths is one of the most relevant aspects, since this allows the robots to carry out the movements required to achieve a specific task within a work area. Various increasingly robust techniques have appeared and/or have been applied in this area, such as convolutional neural networks (CNN). This paper presents the implementation of a cascade CNN set, in such a way that, based on an input image that consists of a food dish, a 10-point path is obtained in a three-dimensional space, forming a path that enters to a given area and generates a movement of food collection. The first network, responsible for segmenting the image to obtain the area of the required food is called ResSeg. The second network is a ResNet-50 model modified to be applied in a regression framework, whose function is to estimate the path of the robot to perform the collection task, based on the segmentation obtained from the previous network. An accuracy of 90.33% is obtained in the path estimation, with a general deviation of 12.55 mm, taking into account a deviation threshold of ± 20 mm. Likewise, the network is tested in real-time in a virtual environment, applying the estimated path to a robotic manipulator of 4 degrees of freedom, being able to demonstrate the reliability and smoothness of the estimated path in the execution of the task.