Deep Reinforcement Learning based Robotic Grasping for Manufacturing

Robots in manufacturing are programmed to achieve specific tasks in a deterministic environment. However, we see recently new manufacturing needs where robot controlling is required to be flexible. It became necessary to develop technologies to enable robot programming in this context. The challenge is to enable robots to accomplish assigned tasks within non-deterministic environments, where positions of machinery, robots, and objects are not static. Our goal is to investigate the elementary manufacturing task of robot grasping. In particular, we are interested in vision-based bin picking techniques which consist of robot arm grasping random objects from a bin using visual inputs. In this work, we investigate the possibility to apply Deep Reinforcement Learning to accomplish this task in a simulated environment. We study several variants of the Off-policy and On-policy Deep Q-learning algorithms. In particular, we study the importance of stochastic optimization for Deep Q-learning and Neural Network initialization. We propose a procedure to make consistent tests in order to compare rigorously Agents performance and generalization. We designed several AutoEncoder models in order to encode depth maps visual features into a latent space representation. We trained a Reinforcement Learning agent that utilizes this latent space representation instead of raw visual inputs. We analyze benefits and drawbacks of proposed techniques and finally, we discuss possible future directions of the following work.