Grasping objects in cluttered environments

Abstract

One of the core challenges in todays robotic manipulation is the grasping problem. It consists in designing a mathematical model of the environment so that the robot is able to compute a hand and finger trajectories that yields a grasping configuration. In order for autonomous robots to interact with their environment, we need them to be able to grasp objects by understanding the environment surrounding them. We will thus present in this work a new solution for the grasping problem in cluttered environments. More specifically, we consider in this work the problem of detecting, planning and executing a grasp in a cluttered complex and uncontrolled environment. We consider here only a single RGB-D camera as input to our model. To solve this problem we use a hybrid model composed of a context segmentation and geometric model. The segmentation model, based on deep learning, captures information about where, and the geometric model, using a dictionary of grasping prototypes alongside a searching algorithm, tells us how to grasp. The segmentation network achieve a jiccard index value of 76.61\% on the training set and 56.93\% on the validation set. We achieve an overall grasp succes rate of 2 out of 3 grasps.