End-to-End Reinforcement Learning with Large Models for Autonomous Vehicle Control

Engineering notes

Experimental results demonstrate that the proposed framework outperforms traditional RL-based and modular approaches in terms of driving stability, obstacle avoidance, and generalization to unseen traffic conditions.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Autonomous vehicle control requires robust and adaptive decision-making to navigate complex and dynamic environments. Traditional model-based control methods and modular perception-planning pipelines often struggle with generalization and real-time adaptability. End-to-end reinforcement learning (RL) with large models has emerged as a promising approach to directly map sensor inputs to control actions while leveraging the scalability of deep learning. This paper proposes an end-to-end reinforcement learning framework that integrates large-scale deep neural networks with advanced policy optimization techniques to improve decision-making in autonomous driving. The model employs a vision-based transformer backbone for high-dimensional feature extraction and a reinforcement learning agent optimized through Proximal Policy Optimization (PPO) for continuous control adaptation. Additionally, a curriculum learning strategy is employed to improve sample efficiency and enhance safety during training. Experimental results demonstrate that the proposed framework outperforms traditional RL-based and modular approaches in terms of driving stability, obstacle avoidance, and generalization to unseen traffic conditions. These findings highlight the potential of large-model reinforcement learning for robust, real-time autonomous vehicle control.

Links and sources

• Official / arXiv page

Comments