Intelligent robot manipulation with deep learning

• Main Author: Tan, Jun Aun
• Other Authors: Lin Zhiping
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Engineering; Reinforcement learning; Imitation learning; Robotics
• Online Access: https://hdl.handle.net/10356/176388
• Institution: Nanyang Technological University

The application of reinforcement learning (RL) in robotics has seen significant advancements across various sectors, yet a critical challenge persists: the simulation-to-reality (sim-to-real) gap. In real-world scenarios, robots frequently underperform due to their inability to access or accurately interpret all observable states, such as the precise state of objects in their environment. This gap often results in discrepancies between expected and actual robot behavior, hindering the effective translation of learned skills from simulated environments to practical applications. This paper presents an innovative approach to bridge this gap through a combination of reinforcement learning (RL) and imitation learning (IL). We introduce a novel teacher-student framework designed to enhance the performance of robotic systems. In this framework, the teacher possesses complete access to all states, including both the environment and the robot, whereas the student is limited to observing only the robot's state and visual input. By implementing an imitation loss in conjunction with the Proximal Policy Optimization (PPO) policy loss during the training of the student's policy, and leveraging expert knowledge transferred from the teacher, we demonstrate a significant improvement in the student's learning efficiency and performance. Our results reveal that the student model, when trained under this hybrid learning paradigm, converges more swiftly, and outperforms models trained solely on RL. The experiments were conducted across multiple environments, each selected for their unique challenges and the ability to test the algorithm's adaptability and performance in varied conditions. This study not only showcases the potential of combining RL with imitation learning to mitigate the knowledge gap but also establishes a foundational framework for future research in enhancing the adaptability and efficiency of robotic systems in real-world applications.