Reinforcement Learning Based Suction Grasping Using Soft Actor-Critic in Simulation Environment

Lee, Yong Yi (2026) Reinforcement Learning Based Suction Grasping Using Soft Actor-Critic in Simulation Environment. Final Year Project (Bachelor), Tunku Abdul Rahman University of Management and Technology.

Text LEE YONG YI_Full Text.pdf Restricted to Registered users only Download (4MB)

Abstract

Industry 4.0 has driven the integration of artificial intelligence (AI), robotics, and advanced automation in industrial systems, increasing the need for more intelligent and adaptive robotic manipulation. Robotic arms are commonly used in tasks such as material handling and assembly, where both precision and flexibility are required to deal with variations in object position and interaction conditions. Traditional grasping methods, especially those using parallel jaw grippers, often depend on predefined rules and are less effective in dynamic environments. Suction-based grasping offers a simpler and more flexible alternative, particularly for handling objects with different shapes and surface properties. This research presents a reinforcement learning-based framework for suction grasping using the Soft Actor-Critic (SAC) algorithm implemented in a PyBullet simulation environment. SAC is chosen due to its stability and suitability for continuous control tasks. The study focuses on developing a structured learning process for sequential manipulation, including approach, alignment, contact, lifting, and placement. A multi-stage reward function is designed to guide the agent through these steps, while action constraints are applied to ensure smooth and stable motion during training. In addition, environment randomisation is introduced to improve robustness by exposing the agent to variations in object position and interaction conditions. The results showthat the proposed framework is able to learn consistent suction grasping behaviour within the simulation environment. The agent demonstrates improved training stability and successfully completes the pick-and-place task through a clear sequence of actions. This study highlights the importance of reward design and training stability in reinforcement learning for robotic manipulation, and provides a practical framework for developing reliable suction-based grasping systems in simulation