Prehensile Robotic pick-and-place in clutter with Deep Reinforcement Learning

In this paper, we present a self-learning deep reinforcement learning-based framework for industrial pick-and-place tasks in a cluttered environment through intelligent prehensile robotic grasping. This approach aims to enable agents learn and perform pick and place regular and irregular objects in clutter through robotic grasping in order to enhance both quantity and quality in various industries. In order to do so, we design a Markov decision process (MDP) and deploy a model-free off-policy temporal difference algorithm Q-learning. We utilize end-to-end DenseNet-121 architecture fully convolutional network (FCN) in extended format for Q-function approximation. A pixelwise parameterization scheme is designed to calculate the pixelwise maps of action values. Rewards are allocated according to the success of the action performed. The proposed approach doesn't require any domain specifications, geometrical knowledge of objects or any extraordinary resources such as huge datasets or memory requirements. We have presented the training and testing results of our approach compared to its different variants and random density clutter sizes.