Attention-Based Reinforcement Learning for Context-Aware Path Tracking Control скачать в хорошем качестве

Attention-Based Reinforcement Learning for Context-Aware Path Tracking Control

Dokeun Lee, Jeonghwan Song, Hyojae Lee and Jeong hwan Jeon, "Attention-Based Reinforcement Learning for Context-Aware Path Tracking Control," 2026 IEEE Intelligent Vehicles Symposium (IV) Abstract : Effective path tracking control for autonomous vehicles remains challenging, especially in diverse road geometries and varying vehicle dynamics. Although reinforcement learning (RL) has shown promise, typical RL policies still make limited use of temporal context because decisions are conditioned mainly on the current observation, which restricts the modeling of long-range dependencies. We propose an Attention-Based Recurrent Proximal Policy Optimization (AR-PPO) agent for context-aware path tracking. The agent combines Gated Recurrent Unit (GRU) with self-attention to construct a weighted representation of the vehicle’s state history, allowing the controller to emphasize the most informative past states when generating actions. We train and evaluate the proposed agent in the CARLA simulator on custom scenarios with varied driving conditions. The results show that AR-PPO achieves higher performance than both RL-based baselines and a classical model-based controller in terms of path tracking accuracy and driving stability. This work highlights the potential of attention mechanisms to improve the closed-loop stability and context-aware decision making of RL-based controllers in autonomous driving.