Curriculum Learning for Pre-Grasp Skill Acquisition Toward Sim-to-Real Robotic Manipulation Under Variable Object Positions

Ding, Abel Ze Quan (2026) Curriculum Learning for Pre-Grasp Skill Acquisition Toward Sim-to-Real Robotic Manipulation Under Variable Object Positions. Final Year Project (Bachelor), Tunku Abdul Rahman University of Management and Technology.

Text Abel Ding Ze Quan_Full Text.pdf Restricted to Registered users only Download (2MB)

Abstract

Simulation-based training has become a widely adopted approach in robotic manipulation due to its safety, scalability, and cost efficiency. However, transferring learned policies from simulation to real-world systems remains a major challenge due to the sim-to-real gap and limited generalization capabilities. In particular, reinforcement learning (RL) policies often struggle to learn stable and meaningful behaviours when trained directly on complex grasping tasks. This study proposes a curriculum learning-based framework for pre-grasp skill acquisition in robotic manipulation under variable object positions. The approach focuses on improving foundational behaviours, including reaching accuracy, contact consistency, and controlled interaction, before progressing to more complex task components. A SAC algorithm is implemented within a PyBullet simulation environment using a KUKA-based robotic arm model. The training process is structured into progressive stages to enhance learning stability and behavioural consistency. To evaluate the effectiveness of the proposed method, a comparison is conducted between curriculum-based training and a non-curriculum baseline under identical conditions. Performance is assessed using intermediate metrics such as end-effector distance to object, contact occurrence, grasp ratio, entropy, and episodic return. The results demonstrate that while the baseline approach achieves higher reward values, it exhibits unstable and unstructured behaviour. In contrast, the curriculum-based approach produces more stable, consistent, and interpretable behaviours, particularly in early-stage skill acquisition. Although full task success is not achieved, the findings highlight the importance of structured learning in RL-based robotic manipulation. The proposed approach provides a practical and resource-efficient methodology for improving behaviour learning and supporting future sim-toreal transfer. This work contributes to the understanding of how curriculum learning can enhance training stability and generalization in contact-rich robotic tasks