arXiv cs.AI· Ding Fan, Xuewen Luo, Hwa Hui Tew, Susilawati, Vishnu Monn Baskaran, Fucai Ke·· 2 天前AI 评分22
M-Diffusion Planner：用多头扩散模型与强化学习实现个性化驾驶
Drive As You Like: Multi-Head Diffusion with Reinforcement Learning for Personalized Driving
AI 导读
M-Diffusion Planner 通过双解耦设计实现个性化自动驾驶：大语言模型推断驾驶需求并选择策略，由共享扩散规划器的对应头执行。研究采用代理策略比率估计与约束式头特定更新改进 Group Relative Policy Optimization，在无风格标注下用策略专属奖励微调各头。nuPlan 开环与闭环评测显示其具备策略对齐行为与有竞争力的规划性能。
正文 · 原文
Abstract:Personalized autonomous driving requires understanding the driving needs implicit in user requests. This is challenging when driving demonstrations lack style annotations and behavioral adaptation risks degrading planning quality. We propose M-Diffusion Planner with a dual-decoupling design: separating intent understanding from trajectory planning, and general planning capability learning from strategy alignment. A large language model infers driving needs and selects a strategy, which is executed by the corresponding head of a shared diffusion planner. Following imitation pretraining, strategy-specific rewards specialize individual heads without style annotations. We innovatively adapt Group Relative Policy Optimization using surrogate policy-ratio estimation and constrained head-specific updates to mitigate degradation of pretrained planning capabilities. Open-loop and closed-loop evaluations on nuPlan demonstrate distinct strategy-aligned behaviors and competitive planning performance. Further experiments assess intent-to-strategy mapping and collision risk, and analyze the effects of reward design and optimization on strategy alignment and planning quality.
| Subjects: | Robotics (cs.RO); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2508.16947 [cs.RO] |
| (or arXiv:2508.16947v3 [cs.RO] for this version) | |
| https://doi.org/10.48550/arXiv.2508.16947 arXiv-issued DOI via DataCite |
Submission history
From: Fan Ding [view email]
[v1]
Sat, 23 Aug 2025 08:33:11 UTC (888 KB)
[v2]
Thu, 23 Jul 2026 17:41:04 UTC (3,651 KB)
[v3]
Thu, 15 Oct 2026 08:01:33 UTC (2,390 KB)
来源：arXiv cs.AI · arxiv.org