Path Planning for Robotic Training in Virtual Environments using Deep Learning

Abstract

Reducing costs in the acquisition of industrial robots and their benefit in continuous production workdays have increased the number of investigations that expand the robot's action capabilities. This work proposes a trajectory planning system for a UR3 robot that is very usual in academic, research, and industrial applications. The system is presented based on a convolutional network training for regression tasks focused on learning the desired trajectory. A virtual environment has been developed to simulate different trajectories based on the interaction between UR3 robot and object detection and location through the convolutional network employed. This work exposes the network's training and the results of the transport of the object, where the robot can position itself on the desired tool (scissors and screwdriver), which is recognized by training a Faster network R-CNN and the re-localization of the tool in a conveyor band. For the trained trajectory’s, a ResNet-50 model is proposed, and the overall performance achieved was 92.63%, with a mean square error of 24.7 mm in the trained trajectory's repetition. Also, the boxplot of each ax in the trajectory is exposed since they show in a more detailed way the deviation of each of the points in the whole validation set. The average collection time, from when the system takes the workspace capture to its initial positioning after leaving the tool on the belt, was 51.3 seconds, enough for real-time applications.