Research on Local End-to-End Learning Model for Autonomous Driving

Engineering notes

Key topics: autonomous driving system, autonomous driving, end-to-end, reinforcement learning, carla, perception, planning, control. See the paper for implementation details and experimental results.

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

To address the problems of low inter-module coordination efficiency, information transmission delay, and insufficient environmental adaptability in traditional autonomous driving systems, this paper proposes an integrated decision-planning-control end-to-end learning method based on a Deep Reinforcement Learning (DRL) framework. This method uses a Variational Autoencoder (VAE) to extract compact latent representations from high-dimensional visual inputs and integrates it with vehicle state data and navigation instructions to construct a unified state representation space. On this basis, a Proximal Policy Optimization (PPO) algorithm is used to achieve a direct end-to-end mapping from raw perception input to low-level vehicle control commands. A set of test tasks covering various complex traffic scenarios was designed on the CARLA simulation platform for system evaluation. Experimental results show that the proposed model demonstrates excellent performance in core driving tasks such as lane keeping and dynamic obstacle avoidance. This paper proposes an end-to-end learning model that integrates VAE-based multi-source state representation with a PPO optimization framework. Compared with existing approaches that rely solely on single-visual input or modular architectures, the proposed model demonstrates significant advantages in stability, convergence speed, and adaptability to complex scenarios.

Links and sources

• Official / arXiv page

Comments