Loading...

Deep Reinforcement Learning for Autonomous Vehicle Control by Integrating Longitudinal and Lateral Dynamics

Rastegar, Nima | 2025

48 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 58602 (08)
  4. University: Sharif University of Technology
  5. Department: Mechanical Engineering
  6. Advisor(s): Saadat Foumani, Mahmoud
  7. Abstract:
  8. With the increasing complexity of tasks in autonomous driving, designing controllers that exhibit both strategic intelligence and high dynamic precision has become a significant challenge. This study focuses on the stability and realism of autonomous driving by designing and comparing two control architectures based on deep reinforcement learning. Initially, an end-to-end controller was developed using a soft actor-critic algorithm to address a highway exit scenario, relying on a 10-dimensional engineered state vector. To overcome limitations such as prolonged training time steps and simplified vehicle dynamics, a two-layer hierarchical control architecture was proposed. In this architecture, a high-level controller, based on the soft actor-critic algorithm, handles strategic decision-making. The commands from this layer are executed by a data-driven low-level controller, which consists of two separate longitudinal and lateral inverse dynamic models trained on simulated vehicle dynamic data using machine learning techniques. These models are responsible for mapping commands to physical actions, including throttle/brake and steering. The longitudinal dynamic model achieved an accuracy of 97%, while the lateral dynamic model reached 98%. To evaluate performance, the two developed architectures were compared in the CARLA simulator. Test results demonstrated a 100% success rate and stability for both methods. The hierarchical architecture significantly accelerated the learning process by reducing the required time steps for convergence by 78%. Additionally, a 50% increase in average speed and a 23% reduction in average jerk were observed in the hierarchical architecture, indicating a smoother control policy closer to human-like behavior. This research demonstrates that both studied approaches are implementable in autonomous vehicles. However, the hierarchical architecture, leveraging a low-level controller trained on vehicle dynamic data, provides a more effective solution for developing safe, efficient, and realistic autonomous driving systems
  9. Keywords:
  10. Autonomous Vehicles (AVs) ; Deep Reinforcement Learning ; Vehicle Dynamics ; Machine Learning ; Hierarchical Control ; Longitudinal Dynamic

 Digital Object List

 Bookmark

No TOC