Stable Robotic Grasping of Target Object Using Deep Reinforcement Learning

Abstract

Robotic grasping within a cluttered environment has shown its potential for improvement through the integration of pushing. Nevertheless, achieving an effective pushing strategy remains a persistent challenge. Furthermore, the complexity is heightened in highly cluttered environments, particularly when dealing with stacked objects, posing significant difficulties for successful grasping. To address these issues, we propose a method aimed at enhancing pushing mechanisms by incorporating an information-theoretic measure of entropy. This integration facilitates the efficient clearance of obstacle objects surrounding the target object. Additionally, we leverage depth information as a foundational action prior, seamlessly integrating it into Q-value calculations. This integration results in an enhanced exploration strategy, consequently improving the grasping of stacked objects. Furthermore, we ensure the stability of grasping through reward shaping, achieved by accounting for changes in the object's pose. We conducted an evaluation of our approach using three distinct object types within a simulation environment, which revealed a notably higher rate of successful grasping in comparison to the baseline method. To substantiate our findings, we proceeded to test the model within a real-world setting, where the proposed approach showcased a substantial enhancement, reinforcing the efficacy of our method.