---
title: "AlphaDrive: Unleashing the Power of VLMs in Autonomous Driving via Reinforcement Learning and Reasoning"
canonical_url: "https://www.modelscope.cn/papers/124998"
md_url: "https://www.modelscope.cn/papers/124998.md"
arxiv_id: 2503.07608
published: 2025-03-10
last_updated: 2025-03-10
authors:
  - "Bo Jiang"
  - "Shaoyu Chen"
  - "Qian Zhang"
  - "Wenyu Liu"
  - "Xinggang Wang"
model_name: AlphaDrive
model_developer: "Horizon Robotics、Huazhong University of Science and Technology"
domain:
  - "计算机视觉"
  - "机器人学"
  - "自动驾驶"
  - "视觉语言模型"
  - "强化学习"
type:
  - "计算机视觉"
  - "机器人学"
  - "自动驾驶"
  - "视觉语言模型"
  - "强化学习"
  - "Computer Vision and Pattern Recognition"
  - Robotics
arxiv_url: "https://arxiv.org/abs/2503.07608"
pdf_url: "https://arxiv.org/pdf/2503.07608"
code_link: "https://github.com/hustvl/AlphaDrive"
---

# AlphaDrive: Unleashing the Power of VLMs in Autonomous Driving via Reinforcement Learning and Reasoning

> OpenAI o1 and DeepSeek R1 achieve or even surpass human expert-level performance in complex domains like mathematics and science, with reinforcement learning (RL) and reasoning playing a crucial role. In autonomous driving, recent end-to-end models have…

「AlphaDrive: Unleashing the Power of VLMs in Autonomous Driving via Reinforcement Learning and Reasoning」是 ModelScope 魔搭社区收录的论文，arXiv 2503.07608，作者为 Bo Jiang, Shaoyu Chen, Qian Zhang et al.，发表于 2025-03-10，属于 计算机视觉、机器人学、自动驾驶 领域。

- **ArXiv**: 2503.07608
- **Published**: 2025-03-10
- **Authors**: Bo Jiang, Shaoyu Chen, Qian Zhang, Wenyu Liu, Xinggang Wang
- **Model**: AlphaDrive
- **Developer**: Horizon Robotics、Huazhong University of Science and Technology
- **Domain**: 计算机视觉, 机器人学, 自动驾驶, 视觉语言模型, 强化学习
- **ArXiv URL**: https://arxiv.org/abs/2503.07608
- **PDF**: https://arxiv.org/pdf/2503.07608
- **Code**: https://github.com/hustvl/AlphaDrive

Source: https://www.modelscope.cn/papers/124998

---

> AlphaDrive：融合强化学习与推理的视觉-语言模型助力自动驾驶规划

## 摘要

本文提出了一种名为AlphaDrive的视觉-语言模型（VLM），专门用于自动驾驶中的高层次规划。现有的端到端模型虽然在规划性能上有所提升，但在处理复杂和长尾问题时仍存在局限性，主要由于缺乏常识和推理能力。AlphaDrive通过引入基于组相对策略优化（GRPO）的强化学习（RL）框架，并结合两阶段训练策略，即监督微调（SFT）和RL，显著提升了规划性能和训练效率。

具体而言，AlphaDrive设计了四种针对规划任务的GRPO奖励机制：规划准确性奖励、动作加权奖励、规划多样性奖励和规划格式奖励。这些奖励机制使得模型能够更好地理解不同驾驶行为的重要性，鼓励多样化的解决方案，并确保输出格式的规范性。此外，两阶段训练策略中，第一阶段利用大规模预训练模型生成高质量的规划推理数据集，第二阶段通过RL进一步优化模型，从而有效缓解了早期训练中的不稳定性和幻觉现象。

实验结果表明，AlphaDrive在大规模驾驶数据集上的表现优于仅使用SFT训练的模型，特别是在规划准确性和训练数据利用率方面有显著提升。尤其值得注意的是，经过RL训练后，AlphaDrive展现出了多模态规划的新兴能力，这对于提高驾驶安全性和效率具有重要意义。

## Abstract

OpenAI o1 and DeepSeek R1 achieve or even surpass human expert-level performance in complex domains like mathematics and science, with reinforcement learning (RL) and reasoning playing a crucial role. In autonomous driving, recent end-to-end models have greatly improved planning performance but still struggle with long-tailed problems due to limited common sense and reasoning abilities. Some studies integrate vision-language models (VLMs) into autonomous driving, but they typically rely on pre-trained models with simple supervised fine-tuning (SFT) on driving data, without further exploration of training strategies or optimizations specifically tailored for planning. In this paper, we propose AlphaDrive, a RL and reasoning framework for VLMs in autonomous driving. AlphaDrive introduces four GRPO-based RL rewards tailored for planning and employs a two-stage planning reasoning training strategy that combines SFT with RL. As a result, AlphaDrive significantly improves both planning performance and training efficiency compared to using only SFT or without reasoning. Moreover, we are also excited to discover that, following RL training, AlphaDrive exhibits some emergent multimodal planning capabilities, which is critical for improving driving safety and efficiency. To the best of our knowledge, AlphaDrive is the first to integrate GRPO-based RL with planning reasoning into autonomous driving. Code will be released to facilitate future research.