Sim-to-real transfer for end-to-end monocular autonomous driving using reinforcement learning | |
| Author | Nattabude Tanasansurapong |
| Call Number | AIT Thesis no.ISE-24-27 |
| Subject(s) | Deep learning (Machine learning) Reinforcement learning Automotive engineering |
| Note | A thesis submitted in partial fulfillment of the requirements for the degree of Master of Engineering in Mechatronics and Machine Intelligence |
| Publisher | Asian Institute of Technology |
| Abstract | This thesis investigates sim-to-real transfer for end-to-end autonomous driving using deep reinforcement learning. A CARLA simulation is used to train an autonomous driving agent, which is subsequently deployed on a real golf car. The focus of this research is to enable vehicle turning in response to high-level commands with input from only a single camera. Pre-trained semantic segmentation models are employed to minimize the reality-simulation gap. To expedite training, a variational autoencoder (VAE) encodes the semantic map, providing simplified data for the agent.The study compares the performance of SAC and TQC. Experimental results show that TQC achieves 17.24% higher scores and 10.2% longer survival times on average compared to SAC, demonstrating superior adaptability to unseen paths and input instability. In real-world testing, agents encountered challenges with turning tasks, requiring adjustments. Nonetheless, this research demonstrates that the proposed approach successfully enables turning and lane-keeping on both straight and curved paths, with TQC showing robustness under diverse conditions. |
| Year | 2024 |
| Type | Thesis |
| School | School of Engineering and Technology |
| Department | Department of Industrial Systems Engineering (DISE) |
| Academic Program/FoS | Mechatronics and Machine Intelligence (MMI) |
| Chairperson(s) | Mongkol Ekpanyapong |
| Examination Committee(s) | Chaklam Silpasuwanchai;Huynh, Trung Luong |
| Scholarship Donor(s) | His Majesty the King's Scholarships (Thailand) |
| Degree | Thesis (M. Eng.) - Asian Institute of Technology, 2024 |