Scholarlink

<p>To address the problem of insufficient control continuity and driving stability in conventional end-to-end autonomous driving models under complex urban scenarios, this paper proposes an end-to-end autonomous driving approach based on a VAE-MLPPO parallel network. The method integrates a Variational Autoencoder (VAE), Long Short-Term Memory (LSTM) networks, and Proximal Policy Optimization (PPO). Specifically, the VAE compresses input images to remove redundant information and improve training efficiency, while fused camera perception and vehicle state data are processed by the MLPPO network to extract spatiotemporal features and generate continuous control commands. Experimental validation in CARLA complex urban scenarios, including the CoRL2017 and No Crash benchmarks, demonstrates that the proposed method achieves 6.3% and 9.3% higher success rates than the BEV + PPO baseline in the most challenging tasks. In addition, roundabout trajectory tests show reductions of 0.11&#xa0;m in average centerline deviation and 0.03&#xa0;rad in angular deviation. Ablation studies further confirm the effectiveness of the VAE in feature compression and the LSTM in enhancing control continuity and driving stability. These results indicate that the VAE-MLPPO method can significantly improve task success rate, driving stability, and model training efficiency in complex urban driving scenarios.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

VAE-MLPPO-Based End-to-End Control for Autonomous Driving in Complex Urban Traffic

  • 期刊论文

  • 发表日期：

    2026年4月17日

查看全文
  • Lixin Wang,
  • Xiaoci Huang

摘要

To address the problem of insufficient control continuity and driving stability in conventional end-to-end autonomous driving models under complex urban scenarios, this paper proposes an end-to-end autonomous driving approach based on a VAE-MLPPO parallel network. The method integrates a Variational Autoencoder (VAE), Long Short-Term Memory (LSTM) networks, and Proximal Policy Optimization (PPO). Specifically, the VAE compresses input images to remove redundant information and improve training efficiency, while fused camera perception and vehicle state data are processed by the MLPPO network to extract spatiotemporal features and generate continuous control commands. Experimental validation in CARLA complex urban scenarios, including the CoRL2017 and No Crash benchmarks, demonstrates that the proposed method achieves 6.3% and 9.3% higher success rates than the BEV + PPO baseline in the most challenging tasks. In addition, roundabout trajectory tests show reductions of 0.11 m in average centerline deviation and 0.03 rad in angular deviation. Ablation studies further confirm the effectiveness of the VAE in feature compression and the LSTM in enhancing control continuity and driving stability. These results indicate that the VAE-MLPPO method can significantly improve task success rate, driving stability, and model training efficiency in complex urban driving scenarios.