Deep Reinforcement Learning for Robotic Grasping: Insights into Learning from Raw Visual Data

This paper addresses the challenges of applying deep reinforcement learning (RL) to robotic grasping tasks, focusing on bridging the sim-to-real gap to ensure seamless transfer from simulated to real-world environments. We propose a two-phase modeling approach: (1) an image-free model designed to validate core functionalities and (2) an image-based model that incorporates visual data to enhance grasping precision. Using QT-Opt as a foundation, the models are trained to learn optimal grasping actions based on state-action pairs in both simulation and real settings.To stabilize training and improve learning efficiency, we employ multiple neural networks enhanced with Polyak averaging and clipped double Q-learning to address gradient instability. For action optimization in continuous control tasks, the Cross Entropy Method (CEM) is integrated to ensure robust policy learning. Additionally, we leverage the Jump-Start Reinforcement Learning (JSRL) method to improve convergence, providing initial guidance from a non-image-based agent to enhance exploration efficiency. Experimental results show the effectiveness of different reward functions and the impact of camera positioning on model performance. We also explore the use of generative frameworks to mitigate visual discrepancies between simulated and real images, improving robustness in real-world grasping tasks. We conducted comprehensive experiments to evaluate the effects of various reward functions and the influence of camera positioning on model performance.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.