SparseDriveV2: Scoring is All You Need for End-to-End Autonomous Driving

SparseDriveV2: Scoring is All You Need for End-to-End Autonomous Driving

⚠️ AI 生成 · 建议对照原文 本页为自动整理的学习笔记；关键数据与引用如需引用，请回查 PDF / 官方版本。

学习档位 中文笔记

类型 文献 · 更新 2026-07-19

标签 autonomous-driving · planning · end-to-end-learning · trajectory-scoring · navsim

所属 端到端驾驶

Migration status: imported from ad_projs@a823662; source anchors and claims remain needs-source-verification.

• Paper: https://arxiv.org/abs/2603.29163
• Code: https://github.com/swc-17/SparseDriveV2 at 696ef77924eb9e0a4b4047d013a50e9854bfa026

中文学习笔记（自动生成，需核验）

Topic: ad-end-to-end-driving · Tier: needs-review · Year: 2026 · Venue: —
Evidence level: partial · 本地全文: 是 · 建议阅读: ~55 分钟
Paper: https://arxiv.org/abs/2603.29163
Code: —
Generator: grok

为什么值得读

系统验证「静态超密轨迹词表+可扩展打分」是否足以逼近甚至超越动态生成方案，给出因式分解路径/速度词表与粗到细打分的可扩展范式，在轻量骨干下于 NAVSIM/Bench2Drive 报告强结果，对端到端多模态规划与 SparseDrive 系列演进有直接参考价值。

一句话结论

足够密的因式分解静态词表配合粗粒度路径/速度打分+细粒度轨迹打分即可实现高效高质端到端规划，无需动态轨迹生成。

问题定义

端到端多模态规划通常对候选轨迹打分选优；静态词表受粗离散与计算/显存限制覆盖不足，动态生成虽更细但更复杂。核心问题是：动态生成是否必要，还是静态词表在足够密且可高效打分时已能达到可比性能。

前置知识

端到端自动驾驶与规划；轨迹词表/锚点、模仿学习与多目标蒸馏（如 Hydra-MDP）；路径–速度因式分解与组合；交叉注意力/可变形聚合；NAVSIM、Bench2Drive 等基准与 PDMS/EPDMS 等指标；SparseDrive 类稀疏场景特征设计。

核心贡献

• 从缩放视角系统研究静态词表密度（以 Hydra-MDP 为代表），发现性能随锚点数持续提升、在计算约束前未饱和。
• 提出可扩展词表表示：将轨迹分解为几何路径与速度剖面，通过组合实现组合覆盖且保持紧凑结构。
• 提出可扩展打分：先对路径与速度分别粗打分剪枝，再对少量组合轨迹细粒度打分，使超密词表可算。
• 综合二者提出 SparseDriveV2，在 NAVSIM 与 Bench2Drive 上达到文中报告的 SOTA 级结果，并强调为打分类方法刷新记录。

方法流程

传感器观测经场景编码器得特征 F，自车状态编码得 E；离线从示范中提取并聚类得到路径词表 P 与速度词表 V，组合得轨迹词表 T=C(p,v)；推理时对路径/速度分别嵌入并与场景/状态交互做粗打分，取 Top-K 后组合轨迹并再条件化、细打分，选最高分轨迹作为规划输出。

关键模块和设计取舍

（1）轨迹因式分解：路径 p 按固定空间间隔 Δs 采样几何形状，速度剖面 v 为固定时间间隔 Δt 上的标量速度；D(τ) 分解、C(p,v) 用累计距离插值重构。（2）词表构建：路径/速度分别 K-Means 聚类，|T|=Np×Nv，文中示例相对先验可达约 32× 更密（1024×256 vs 8192）。（3）场景编码：遵循 SparseDrive 哲学直接用多视角图像特征交互，不做显式 BEV。（4）粗打分：路径可用交叉注意力或沿路径可变形聚合，速度用交叉注意力；细打分仅在少量组合轨迹上进行。取舍：用因式组合换覆盖、用两阶段打分控复杂度，坚持纯打分、无动态生成子模块。

数据集、实验设置与指标

缩放实验基于 Hydra-MDP，报告 NAVSIM v2 上 EPDMS 与 NVIDIA L20 48GB 训练显存；主结果称 NAVSIM 上 PDMS/EPDMS，Bench2Drive 上 Driving Score 与 Success Rate；骨干为轻量 ResNet-34。摘录未给出完整训练超参、数据划分与全部对比设置细节。

主要结果（需原文证据）

Hydra-MDP 静态锚点缩放：1024→16384 锚点时 EPDMS 从 85.02 升至 87.35，各档相对前一档约 +0.45~+0.78，32768 OOM。SparseDriveV2（ResNet-34）：NAVSIM 上 92.0 PDMS、90.1 EPDMS；Bench2Drive 上 89.15 Driving Score、70.00 Success Rate；词表密度约 32× 于先验（1024×256 vs 8192）。更细对比与消融待来源核验。

局限、失败场景与适用边界

摘录未系统列失败场景；隐含边界包括：词表仍依赖示范聚类、对分布外几何/速度覆盖有限；超密组合依赖粗打分剪枝质量；路径/速度独立粗分可能漏掉强时空耦合模式；计算上限仍受显存约束（缩放表中 OOM）；是否全面优于所有动态/混合方法需完整实验核对。适用：以打分选优为主的端到端多模态规划、可接受离线词表的场景。

与前序 / 同期 / 后续方法的关系

打分类：VADv2 概率规划；Hydra-MDP 多目标蒸馏；DriveSuprim 粗到细精炼——受词表粗离散限制。动态生成：回归（如 ipad）与扩散/流匹配（DiffusionDrive/V2、GoalFlow）等，表达力强但模块更复杂。混合：GTRS 结合扩散提案与静态词表及稳健打分。SparseDriveV2 坚持纯静态打分，用因式词表+可扩展打分回应「动态是否必要」。

官方代码与复现建议

官方代码与模型：https://github.com/swc-17/SparseDriveV2。复现建议：先复现 Hydra-MDP 词表密度缩放表验证环境；再实现路径/速度分解与组合及两阶段打分；用 ResNet-34 对齐 NAVSIM/Bench2Drive 协议；注意路径长度 Smax、Δs/Δt、Np/Nv 与 Top-K。摘录未给完整配置，细节以仓库与全文为准。

推荐阅读顺序

摘要与引言问题陈述 → Table 1 缩放实验 → §3.1 问题形式化 → §3.2–3.3 因式词表与构建 → §3.4 粗/细打分与图 1 总览 → Related Work 定位 → 实验与结果页（摘录未全）→ 结论与代码。

自测问题

1. Q: SparseDriveV2 认为静态词表性能瓶颈的主要原因是什么？ A: 并非静态本质缺陷，而是在计算约束下动作空间覆盖不足；缩放实验显示密度增加时性能持续提升且未先饱和。
2. Q: 轨迹如何被因式分解与重构？ A: 分解为几何路径 p（固定空间间隔）与速度剖面 v（固定时间间隔上的速度）；重构时累计 st=Σ vk Δt，沿路径在距离 st 处插值得到 τ=C(p,v)。
3. Q: 可扩展打分如何控制超密词表的计算量？ A: 先对路径、速度独立粗打分并 Top-K，再仅对少量组合轨迹做细粒度时空打分，使精确推理不随全词表线性爆炸。
4. Q: 文中报告的词表密度相对先验大约是多少？ A: 约 32× 更密，示例为 1024×256 对比 8192。
5. Q: 摘要中 SparseDriveV2 在 NAVSIM 与 Bench2Drive 的主要数字是什么？ A: NAVSIM：92.0 PDMS、90.1 EPDMS；Bench2Drive：89.15 Driving Score、70.00 Success Rate（ResNet-34 骨干）。

原文证据

• page 1 / Abstract: SparseDriveV2 scales the trajectory vocabulary to be 32× denser than prior methods, while still enabling efficient scoring over such super-dense candidate set. With a lightweight ResNet-34 as backbone, SparseDriveV2 achieves 92.0 PDMS and 90.1 EPDMS on NAVSIM, with 89.15 Driving Score and 70.00 Success Rate on Bench2Drive.
• page 2 / Table 1: #Anchors EPDMS Memory (MB): 1024 85.02 9531; … 16384 87.35 (+0.57) 38877; 32768 – OOM
• page 2–3 / Introduction: as the number of trajectory anchors increases, planning performance consistently improves, without exhibiting saturation before computational constraints are reached.
• page 5 / Fig.1 caption: SparseDriveV2 factorizes (a) a spatiotemporal trajectory into (b) a geometric path and a velocity profile, and reconstructs the trajectory by (c) composing the two components.
• page 7–8 / §3.3–3.4: the size of the trajectory vocabulary scales as |T| = Np × Nv … coarse factorized scoring independently over paths and velocity profiles to prune low-quality candidates, followed by fine-grained scoring on a small set of composed trajectories.

Evidence-backed Reading Notes

Topic	Evidence-backed note	Source	Short original cue
Problem	Large candidate trajectory sets improve coverage but direct scoring can become computationally expensive.	[PDF p.1, Abstract]	Introduction
Representation	The note treats the method as factorized trajectory vocabulary + scoring because the paper’s method pages introduce the relevant representation/module vocabulary.	[PDF p.3, Method]	Method
Core mechanism	Factorize paths and velocity profiles, apply coarse scoring for pruning, then fine score composed trajectories.	[PDF p.3, Method]	Scoring is All You Need
Input / Output	Input: scene features and trajectory vocabulary. Output: scored ego trajectory candidates.	[PDF p.3, Method]	NAVSIM
Training / Evaluation	The paper reports NAVSIM and Bench2Drive metrics; metric protocol must be bound to benchmark version.	[PDF p.11, Method]	PDMS
Relationship	GTRS emphasizes generalized scoring; DiffusionDrive generates with diffusion; SparseDrive is the sparse-scene predecessor.	[PDF p.3, Method]	Related Work
Failure/Risk	Vocabulary/checkpoint mismatch, candidate coverage, benchmark version drift, and factorization assumptions are key risks.	[PDF p.3, Method]	Scoring is All You Need
Reproduction boundary	Requires matching trajectory vocabulary/checkpoint artifacts.	[PDF p.3, Method]	NAVSIM

Code Cross-References

Local path	Why it matters
696ef77924eb9e0a4b4047d013a50e9854bfa026:README.md	Code/repo anchor for implementation cross-check.
696ef77924eb9e0a4b4047d013a50e9854bfa026:navsim/planning/script/config/pdm_scoring/__init__.py	Code/repo anchor for implementation cross-check.
696ef77924eb9e0a4b4047d013a50e9854bfa026:navsim/planning/script/config/pdm_scoring/default_run_pdm_score.yaml	Code/repo anchor for implementation cross-check.
696ef77924eb9e0a4b4047d013a50e9854bfa026:navsim/planning/script/config/pdm_scoring/default_run_pdm_score_fast.yaml	Code/repo anchor for implementation cross-check.
696ef77924eb9e0a4b4047d013a50e9854bfa026:navsim/planning/script/config/pdm_scoring/default_run_pdm_score_fast_v1.yaml	Code/repo anchor for implementation cross-check.
696ef77924eb9e0a4b4047d013a50e9854bfa026:navsim/planning/script/config/pdm_scoring/default_scoring_parameters_v1.yaml	Code/repo anchor for implementation cross-check.

Reproduction Notes

• Treat this note as paper/code reading material first; do not interpret mini-data smoke tests as paper reproduction. [PDF p.3, Method]
• Before running experiments, verify the local code entry points above against the paper method terminology and dataset protocol. [PDF p.3, Method]
• If a claim is not linked to a PDF page or code path in this note, treat it as an implementation hypothesis rather than established paper fact.

Local Asset Identity

Discovery evidence

• topic: ad-end-to-end-driving
• sources: openalex
• retrieved_at: 2026-07-20
• query: end-to-end autonomous driving planning perception
• doi: 10.48550/arxiv.2603.29163
• score_total: 40
• suggested_tier: watch

Relevance

(no prose relevance explanation — numeric score only or HTTP source)

Snippets

(no snippet evidence in candidate pool)

延伸解读与背景补充

生成：2026-07-21 · 来源条数 2 · 模型 heuristic · 需人工核验数字

问题再表述

围绕「SparseDriveV2: Scoring is All You Need for End-to-End Autonomous Driving」的核心问题与动机（待结合全文核验）。

方法要点补强

• 见原文方法章节；以下为基于摘要/摘录的要点提示。
• SparseDriveV2: Scoring is All You Need for End-to-End Autonomous Driving …

与相近工作的关系

与相近工作的关系待核验；请对照 related work。

常见误读

• 勿仅凭摘要推断未给出的数值指标。

自测问题

1. 这篇工作的输入/输出表示是什么？（SparseDriveV2: Scoring is All You Need for End-to-End Autonomous Driving）
2. 训练目标与评测协议各是什么？
3. 主要失败模式或局限是什么？

外部解读索引

• [2603.29163] SparseDriveV2: Scoring is All You Need for End-to-End Autonomous Driving — 2026-07-21 — 官方摘要/二次页面（自动抓取）
• [2603.29163] SparseDriveV2: Scoring is All You Need for End-to-End Autonomous Driving — 2026-07-21 — 官方摘要/二次页面（自动抓取）

关键图表

方法结构（重绘）

flowchart LR
  A["输入 / 观测"] --> B["表示 / 编码"]
  B --> C["推理 / 解码"]
  C --> D["输出 / 动作或检测"]
  %% method sketch for: SparseDriveV2: Scoring is All You Need for End-to-End Autonomous Driving

方法结构示意（重绘；细节以原论文为准，待 PDF 核验）。

论文摘录图/表

来源：原论文约 p.4（arch）；学习用途摘录。

来源：原论文约 p.2（table）；学习用途摘录。

英文自动分析（可折叠）

展开英文 Paper Card / AI deep analysis

Paper Card

Field	Content
One-line takeaway	SparseDriveV2 shows that scaling and factorizing trajectory vocabulary/scoring can make scoring-based E2E planning highly competitive.
Problem	Large candidate trajectory sets improve coverage but direct scoring can become computationally expensive.
Representation	factorized trajectory vocabulary + scoring
Input / Output	Input: scene features and trajectory vocabulary. Output: scored ego trajectory candidates.
Core Mechanism	Factorize paths and velocity profiles, apply coarse scoring for pruning, then fine score composed trajectories.
Training / Evaluation	The paper reports NAVSIM and Bench2Drive metrics; metric protocol must be bound to benchmark version.
Reproduction Status	Requires matching trajectory vocabulary/checkpoint artifacts.
Compare With	GTRS emphasizes generalized scoring; DiffusionDrive generates with diffusion; SparseDrive is the sparse-scene predecessor.
Failure/Risk	Vocabulary/checkpoint mismatch, candidate coverage, benchmark version drift, and factorization assumptions are key risks.

原文摘录与素材（可折叠）

展开 Extract / Selections / Local assets

Source Anchors

Anchor	What to verify	Source	Short original cue
Title and abstract	Use to verify paper identity and top-level contribution.	[PDF p.1, Abstract]	SparseDriveV2: Scoring is All You Need for End-to-End Autonomous Driving
Motivation	Use to verify the problem statement and why the work is needed.	[PDF p.1, Abstract]	Introduction
Core method	Use to verify the main modeling mechanism and module names.	[PDF p.3, Method]	Method
Key module terms	Use to verify exact component names before editing the note.	[PDF p.3, Method]	Scoring is All You Need
Dataset and protocol	Use to verify data dependencies: NAVSIM / Bench2Drive.	[PDF p.3, Method]	NAVSIM
Metrics and results	Use to verify metric names and reported benchmark context.	[PDF p.11, Method]	PDMS
Experiments or ablation	Use to verify which claims are experimentally supported.	[PDF p.12, Method]	Experiments
Position in related work	Use to verify the claimed relationship to neighboring methods.	[PDF p.3, Method]	Related Work
Conclusion or limits	Use to verify final claims and remaining constraints.	[PDF p.14, Experiments]	Conclusion

反向链接

• 端到端驾驶
• 自动驾驶数据集与评测矩阵
• 自动驾驶数据集与评测矩阵
• 自动驾驶方法谱系
• 自动驾驶方法谱系

---

学习导航

标记为已读

• 所属 Topic: ad-end-to-end-driving
• 学习路径: 学习路线 · 论文库
• 上一篇: SparseDrive 文献笔记
• 下一篇: UniAD: Planning-oriented Autonomous Driving
• 返回 Topic