Improving Reinforcement Learning with Expert Demonstrations and Vision Transformers for Autonomous Vehicle Control
While IL has been successfully applied in RL-based approaches for autonomous driving, significant challenges, such as limited data for RL and poor generalization in IL, still need further investigation. To overcome these limitations, we propose in this paper a novel approach that effectively combine...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
MDPI AG
2024-12-01
|
| Series: | World Electric Vehicle Journal |
| Subjects: | |
| Online Access: | https://www.mdpi.com/2032-6653/15/12/585 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | While IL has been successfully applied in RL-based approaches for autonomous driving, significant challenges, such as limited data for RL and poor generalization in IL, still need further investigation. To overcome these limitations, we propose in this paper a novel approach that effectively combines IL with DRL by incorporating expert demonstration data to control AV in roundabout and right-turn intersection scenarios. Instead of employing CNNs, we integrate a ViT into the perception module of the SAC algorithm to extract key features from environmental images. The ViT algorithm excels in identifying relationships across different parts of an image, thereby enhancing environmental understanding, which leads to more accurate and precise decision making. Consequently, our approach not only boosts the performance of the DRL model but also accelerates its convergence, improving the overall efficiency and effectiveness of AVs in roundabouts and right-turn intersections with dense traffic by a achieving high success rate and low collision compared to RL baseline algorithms. |
|---|---|
| ISSN: | 2032-6653 |