Enhancing stable grasping of containers with dynamic center of mass using reinforcement learning

• Main Author: Gao, Yuan
• Other Authors: Lin Zhiping
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2025
• Subjects: Engineering; Robotic grasping; Center of mass; Reinforcement learning
• Online Access: https://hdl.handle.net/10356/182184
• Institution: Nanyang Technological University

This research focuses on advancing robotic grasping capabilities, specifically targeting dynamic containers with varying center of mass and irregular internal shapes. Leveraging computer vision and reinforcement learning techniques, we aimed to optimize the grasping process. Our approach involved the integration of vision-based Gelsight and GelSlim tactile sensors to capture tactile information, particularly emphasizing tactile shear forces for enhanced perception of the container. A reinforcement learning-based controller was developed to determine the optimal grasping location. Utilizing the Soft Actor-Critic (SAC) algorithm, we trained a control policy that synthesized sensor readings and visual data as inputs. The policy outputted delta changes in the grasping location, facilitating adaptive grasping in diverse environments. Finally, we applied the reinforcement learning policy trained in simulation to the real UR5 robot, demonstrating the feasibility and effectiveness of our approach in real-world scenarios.