Master thesis : Deep Reinforcement Learning for Robotic Grasping

Abstract

The development and deployment of robotic grasping systems in the industry help to improve the efficiency and productivity of one’s production lines. Even though interesting for any industrial actor, those robotic systems require a significant upfront investment. This significant investment is composed of two primary types of costs: hardware and software. Thanks to recent developments in Deep Reinforcement Learning applied to robotic grasping through vision-based systems, IntegrIA is researching solutions that could reduce the software costs of robotic grasping applications focused on pick-and-place tasks.

Thus, this master’s thesis implements a state-of-the-art reinforcement learning algorithm named QT-Opt and aims to compare it with IntegrIA’s one. Both online and offline learning versions of QT-Opt are developed, resulting in three training algorithms to compare across three training datasets. Performances of resulting agents are quantitatively evaluated and qualitatively compared through metrics such as the normalised area under the success rate curve.

In the end, it is observed that this master thesis best agent trained on a dataset composed of 1,800 objects achieves a grasping success rate of 96.67% on previously unseen objects, against 97.32% for IntegrIA’s agent. Even though it cannot outperform their implementation, it is interesting to observe that the best agent trained for this master’s thesis achieves the 96% success rate from the original paper while being powered with a fraction of its resources.