端到端自动驾驶规划方法研究综述：场景表示、规划生成与评估验证

A Review of End-to-End Autonomous Driving Planning Methods: Scene Representation, Planning Generation, and Evaluation and Validation

摘要: 端到端自动驾驶规划通过减少人工规则和模块接口限制，使模型能够围绕最终驾驶任务进行联合优化。然而，场景表示、辅助任务和模型结构日益复杂，并不意味着其实际驾驶能力得到充分验证，场景信息能否有效支撑规划、不同技术机制如何生成驾驶行为，以及现有评估能否可靠反映模型能力，仍需系统分析。为此，本文围绕“场景表示–规划生成–评估验证”这一主线，对端到端自动驾驶规划研究进行综述。首先，界定端到端规划的任务边界与分类依据，归纳鸟瞰视角表示、占据表示、向量化场景表示以及语义增强与潜在世界状态表示，分析其从空间关系、三维占用、道路拓扑、高层语义和未来演化等方面对规划的支撑作用。其次，依据规划结果生成与改进的主导机制，将现有方法划分为模仿学习、多任务联合学习、生成模型和视觉–语言–动作模型四类，比较其规划依据、生成机制、输出形式及能力边界。随后，比较开环评估、闭环仿真评估和语言–动作一致性评估的评价对象、主要指标与适用场景，分析其在衡量离线规划质量、连续驾驶能力和语义–行为一致性方面的适用性与局限。最后，归纳现有研究的共性挑战，并展望面向闭环可信与可验证安全的发展方向。

Abstract: End-to-end autonomous driving planning reduces reliance on manually designed rules and module interfaces, enabling models to be jointly optimized toward the final driving task. However, increasingly complex scene representations, auxiliary tasks, and model architectures do not necessarily imply that actual driving capabilities have been adequately validated. Whether scene information can effectively support planning, how different technical mechanisms generate driving behaviors, and whether existing evaluation methods can reliably reflect model capability still require systematic investigation. To address these issues, this paper reviews end-to-end autonomous driving planning along the main line of “scene representation - planning generation - evaluation and validation”. First, the task boundaries and classification criteria of end-to-end planning are defined. Major scene representations, including bird’s-eye-view representation, occupancy representation, vectorized scene representation, semantic-enhanced representation, and latent world-state representation, are summarized, with emphasis on how they support planning through spatial relationships, three-dimensional occupancy, road topology, high-level semantics, and future evolution. Second, according to the dominant mechanisms used to generate and refine planning results, existing methods are categorized into four groups: imitation learning, multi-task joint learning, generative models, and vision-language-action models. Their planning basis, generation mechanisms, output forms, and capability boundaries are then compared. Subsequently, open-loop evaluation, closed-loop simulation evaluation, and language-action consistency evaluation are compared in terms of evaluation objects, major metrics, and applicable scenarios, and their applicability and limitations in measuring offline planning quality, continuous driving capability, and semantic-behavior consistency are analyzed. Finally, the common challenges faced by existing studies are summarized, and future research directions toward closed-loop trustworthiness and verifiable safety are discussed.