Obstacle Evasion Algorithm Using Convolutional Neural Networks and Kinect-V1

Abstract

The following paper presents the development of an algorithm for the evasion of static obstacles during the process of gripping the desired object, using an anthropomorphic robot, artificial intelligence, and machine vision systems. The algorithm has developed to detect a variable number of obstacles (between 1 and 15) and the grip desired element, using a robot with 3 degrees of freedom (DoF). A Kinect V1 was used to capture the RGB-D information of the environment and Convolutional Neural Networks for the detection and classification of each element. The capture of the three-dimensional information of the detected objects allows comparing the distance between the obstacles and the robot, to make decisions regarding the movement of the gripper to evade elements present in the path and hold the desired object without colliding. Obstacles of less than 18 cm in height were avoided, concerning the ground, with a probability of collision of 0% under specific environmental conditions, moving the robot since initial path in a straight line to the desired object, which is prone to changes according to the obstacles present in its. Function tests have been according to the manipulator's ability to evade possible obstacles of different heights located between the robot and the desired object