Towards Optimal Grasping Of Unknown Objects Using Deep Reinforcement Learning

Smart manufacturing, which uses advanced technologies to optimize processes, has become increasingly popular. Improving the agility of robotic systems, including manipulation skills, is crucial in this field. Grasp synthesis, the process of developing grasping plans while manipulating objects, can be approached empirically. Deep reinforcement learning (DRL) is an empirical method that does not require a dataset and learns tasks by interacting with an environment. This study aims to utilize DRL algorithms to perform grasping by creating a novel grasping environment for a simulation model of a three-finger gripper and then validating this model to achieve optimal grasp on unknown objects. The results showed that the trained DRL model in the validated simulation environment successfully grasped unknown objects placed randomly. The agent identified optimal grasps using a grasp quality score in the DRL model’s reward function.