Robotic Grasping using Deep Reinforcement Learning

Abstract

The evolution of technology has led to a growing demand for robotic systems in industrial sectors, including autonomous robotic grasping. Traditional analytical methods, although efficient, are impractical in scenarios that require regular changes in programming. This dissertation presents a methodology based on deep reinforcement learning to address this limitation in a simulated environment. The objective is to have the capacity to generalize to novel objects and ensure the learned grasp behavior is stable. The task consists of using a robotic manipulator to grasp an object, lift it, and transport it to a predefined target position. Training is done using Proximal Policy Optimization in the Isaac Sim simulator. Knowledge distillation is also used to transfer the learned behavior to a more efficient model with reduced and more realistic observation requirements. The results show that the proposed methodology generalizes to objects not seen during training while maintaining an adequate and stable grasping behavior.