Research KB 登录

IDX-001 Index

NAVSIM 规划线

id
updated
type index
title NAVSIM 规划线
indexes BMK-006, CPT-017, PPR-2608.19085, PPR-2411.15139, PPR-2512.07745, PPR-2605.15120, PPR-2603.24581, PPR-2412.01718, PPR-2406.06978, PPR-2503.12820, PPR-2503.05689, PPR-2506.06659, PPR-2506.06664, PPR-2402.13243, PPR-2405.19620, PPR-2603.29163, PPR-2505.15111, PPR-2503.11650, PPR-2504.19580, PPR-2507.17596, PPR-2212.10156, PPR-2205.15997, PPR-2303.12077, PPR-2402.11502, PPR-2503.07656, PPR-2408.03601, PPR-2312.03031, PPR-2505.09315, PPR-2507.00603, PPR-2405.04390, PPR-2406.08481, PPR-2510.12796, PPR-2601.22032, PPR-2607.29031, PPR-2605.09701, PPR-2605.31476, PPR-2512.10226, PPR-2405.17398, PPR-2309.17080, PPR-2311.16038, PPR-2311.17918, PPR-2412.18607, PPR-2506.24113, PPR-2506.08052, PPR-2507.04049, PPR-2510.08562, PPR-2409.00588, PPR-2408.00415, PPR-2308.01898, PPR-2404.07762, PPR-2310.12432, PPR-2502.13144, PPR-2506.09985, PPR-2603.14482, PPR-2506.13757, PPR-2504.01941, PPR-2605.19771, PPR-2601.05083, PPR-2510.04333

单轨迹回归起点

  • PPR-2205.15997（TransFuser） — 自注意力在多个分辨率融合透视视图与鸟瞰视图特征；其去 LiDAR 变体是本线常用基线。
  • PPR-2212.10156（UniAD） — 感知、预测、规划并入单一网络，以规划为目标重排任务优先级，统一 query 接口通信。
  • PPR-2303.12077（VAD） — 把场景建成完全向量化表示，智能体运动与地图元素作为实例级规划约束。
  • PPR-2312.03031（BEV-Planner） — 对开环评估的诊断：模型过度依赖 ego 状态，并提出轨迹是否沿道路行驶的指标。
  • PPR-2405.19620（SparseDrive） — 对称稀疏感知模块统一检测、跟踪、在线建图，配并行运动规划器。
  • PPR-2402.11502（GenAD） — 生成式表述：实例中心场景 tokenizer 加潜空间轨迹分布。
  • PPR-2309.17080（GAIA-1）、PPR-2311.16038（OccWorld）、PPR-2311.17918（Drive-WM）、PPR-2405.17398（Vista）、PPR-2412.18607（DrivingGPT）、PPR-2506.24113（Epona） —面向未来预测的世界模型，预测媒介分别为离散 token、三维占用、多视角视频与视频扩散。

候选构造

  • PPR-2402.13243（VADv2） — 概率式规划：动作空间离散为大规划词表并 token 化，规划 token 与场景 token 交互后输出分布。
  • PPR-2406.06978（Hydra-MDP） — 保留候选集，监督改为人类示范与规则教师的多教师蒸馏，多头解码器对齐不同评估指标。
  • PPR-2503.12820（Hydra-MDP++） — Hydra-MDP 的扩展，教师指标加入红绿灯合规、车道保持与扩展舒适。
  • PPR-2411.15139（DiffusionDrive） — 以先验锚点构造高斯混合起点并截断扩散调度，配级联扩散解码器。
  • PPR-2503.05689（GoalFlow） — 以场景信息选出的目标点约束流匹配，生成多模轨迹。
  • PPR-2510.08562（ResAD） — 不直接预测未来轨迹，改预测相对确定性惯性参考的残差。
  • PPR-2505.09315（TransDiffuser） — 不依赖预定义轨迹词表或场景先验处理模式坍缩。
  • PPR-2506.06664（GTRS） — 扩散生成细粒度候选，打分器在超稠密轨迹集上训练。
  • PPR-2505.15111（iPad） — 候选置于特征提取与辅助任务中心，用候选锚定注意力迭代精化。
  • PPR-2506.06659（DriveSuprim） — 由粗到细的渐进候选过滤，配旋转增广与自蒸馏。
  • PPR-2603.29163（SparseDriveV2） — 对 Hydra-MDP 做稠密静态词表的缩放研究，简化其余组件。
  • PPR-2504.19580（ARTEMIS） — 自回归逐点生成轨迹，专家混合按场景路由查询。
  • PPR-2408.03601（DRAMA） — 以 Mamba 序列建模融合相机、LiDAR 鸟瞰视图与 ego 状态。
  • PPR-2507.17596（PRIX） — 仅相机、无显式鸟瞰视图，生成式规划头直接从原始像素出轨迹。
  • PPR-2503.07656（DriveTransformer） — 任务并行、稀疏表示与流式处理的统一框架。

打分信号

  • PPR-2406.06978、PPR-2503.12820 — 规则教师与人类示范的多教师蒸馏子分数。
  • PPR-2601.05083（DrivoR） — 相机感知寄存器 token 压缩多相机特征，打分解码器模仿 oracle 并预测安全、舒适、效率子分数。
  • PPR-2605.15120（CLOVER） — 打分手拟合真实规划指标子分数，并把打分手选目标回灌生成器。
  • PPR-2507.04049（DIVER） — 把扩散过程当随机策略，用组相对策略优化优化轨迹级多样性与安全奖励。
  • PPR-2512.07745（DiffusionDriveV2） — 尺度自适应乘性噪声探索，锚内与锚间截断的两级组相对策略优化。
  • PPR-2502.13144（RAD） — 三维高斯泼溅闭环中大规模试错，模仿学习作正则项。
  • PPR-2506.08052（ReCogDrive） — 视觉语言模型的驾驶先验经强化学习注入扩散规划器。
  • PPR-2504.01941（WoTE） — 在鸟瞰视图世界模型中推演候选的未来状态并据此评估。
  • PPR-2605.09701（DriveFuture） — 未来感知潜作为扩散规划器的显式条件。
  • PPR-2608.19085（DA-WAM） — 每条候选轨迹生成各自的未来潜状态，并以其为打分的条件。

未来潜与决策的耦合

  • PPR-2406.08481（LAW） — 以当前特征与 ego 轨迹为条件预测未来场景特征，作自监督任务。
  • PPR-2405.04390（DriveWorld） — 动态记忆库与静态场景传播的时空预训练，任务提示解耦下游特征。
  • PPR-2510.12796（DriveVLA-W0） — 以预测未来图像补偿动作监督的稀疏，分自回归与扩散两种世界模型。
  • PPR-2601.22032（Drive-JEPA） — 视频联合嵌入预测架构预训练后，以多模轨迹蒸馏衔接规划。
  • PPR-2506.09985（V-JEPA 2）、PPR-2603.14482（V-JEPA 2 的稠密特征后续版本） —驾驶域外的视频联合嵌入预测骨干。
  • PPR-2507.00603（World4Drive） — 意图条件潜未来加世界模型选择器。
  • PPR-2607.29031（Auto-JEPA） — 冻结视觉骨干，预测连续未来意图嵌入并从固定轨迹记忆中检索与排序。
  • PPR-2605.31476（IDOL） — 逆动力学把相邻未来潜解码为运动更新。
  • PPR-2512.10226（LCDrive） — 把未来后果表达为动作对齐的潜链式思维 token。
  • PPR-2603.24581（Latent-WAM） — 空间感知压缩世界编码器与动态潜世界模型；测试时丢弃动力学分支。
  • PPR-2605.19771（BeyondDrive） — 流匹配式负轨迹生成器合成安全关键但贴近专家的轨迹。

评估口径

  • BMK-006（NAVSIM） — 非反应式仿真基准；第一代 PDMS 与第二代两阶段 EPDMS 的子分数集与聚合方式不同。评分概念见 CPT-017。
  • PPR-2412.01718（HUGSIM） — 三维高斯泼溅的照片级闭环仿真器，提出 HD-Score。
  • PPR-2408.00415（DriveArena）、PPR-2308.01898（UniSim）、PPR-2404.07762（NeuroNCAP）、PPR-2310.12432（CAT） —闭环仿真与安全关键场景生成。
  • PPR-2510.04333（RAP） — 放弃照片真实感，用三维光栅化生成反事实增广数据。

方法前置（驾驶域外）

  • PPR-2409.00588（DPPO） — 用策略梯度微调扩散式策略的算法框架。
  • PPR-2506.13757（AutoVLA） — 把推理与动作生成并入单个自回归生成模型。
  • PPR-2503.11650（Centaur） — 以 Cluster Entropy 不确定性做测试时训练。
  • PPR-2406.06978 的规则教师谱系与 PPR-2402.13243 的大词表采样谱系见“候选构造”节。

未决 / 待补

  • 本索引成员当前均为 Tier-0 源卡；本线尚无投影 Claim 与 VER 卡。轴上的结构判断需要 claim 级承载才能进入对象层（届时由 SYN 承担，见 D7 与 D29 的分工）。
  • 51 张引用展开卡的 version/source-hash 为显式占位，arXiv 源未拉取。
  • 成员的实质内容以其卡片正文为准；本卡只做导航。

关联（59）

  • PPR-2205.15997 TransFuser: Imitation With Transformer-Based Sensor Fusion for Autonomous Driving
  • PPR-2212.10156 Planning-oriented Autonomous Driving
  • PPR-2303.12077 VAD: Vectorized Scene Representation for Efficient Autonomous Driving
  • PPR-2308.01898 UniSim: A Neural Closed-Loop Sensor Simulator
  • PPR-2309.17080 GAIA-1: A Generative World Model for Autonomous Driving
  • PPR-2310.12432 CAT: Closed-loop Adversarial Training for Safe End-to-End Driving
  • PPR-2311.16038 OccWorld: Learning a 3D Occupancy World Model for Autonomous Driving
  • PPR-2311.17918 Driving Into the Future: Multiview Visual Forecasting and Planning with World Model for Autonomous Driving
  • PPR-2312.03031 Is Ego Status All You Need for Open-Loop End-to-End Autonomous Driving?
  • PPR-2402.11502 GenAD: Generative End-to-End Autonomous Driving
  • PPR-2402.13243 VADv2: End-to-End Vectorized Autonomous Driving via Probabilistic Planning
  • PPR-2404.07762 NeuroNCAP: Photorealistic Closed-loop Safety Testing for Autonomous Driving
  • PPR-2405.04390 DriveWorld: 4D Pre-Trained Scene Understanding via World Models for Autonomous Driving
  • PPR-2405.17398 Vista: A Generalizable Driving World Model with High Fidelity and Versatile Controllability
  • PPR-2405.19620 SparseDrive: End-to-End Autonomous Driving via Sparse Scene Representation
  • PPR-2406.06978 Hydra-MDP: End-to-end Multimodal Planning with Multi-target Hydra-Distillation
  • PPR-2406.08481 Enhancing End-to-End Autonomous Driving with Latent World Model
  • PPR-2408.00415 DriveArena: A Closed-Loop Generative Simulation Platform for Autonomous Driving
  • PPR-2408.03601 DRAMA: An Efficient End-to-end Motion Planner for Autonomous Driving with Mamba
  • PPR-2409.00588 Diffusion Policy Policy Optimization
  • PPR-2411.15139 DiffusionDrive: Truncated Diffusion Model for End-to-End Autonomous Driving
  • PPR-2412.01718 HUGSIM: A Real-Time, Photo-Realistic and Closed-Loop Simulator for Autonomous Driving
  • PPR-2412.18607 DrivingGPT: Unifying Driving World Modeling and Planning with Multi-Modal Autoregressive Transformers
  • PPR-2502.13144 RAD: Training an End-to-End Driving Policy via Large-Scale 3DGS-based Reinforcement Learning
  • PPR-2503.05689 GoalFlow: Goal-Driven Flow Matching for Multimodal Trajectories Generation in End-to-End Autonomous Driving
  • PPR-2503.07656 DriveTransformer: Unified Transformer for Scalable End-to-End Autonomous Driving
  • PPR-2503.11650 Centaur: Robust End-to-End Autonomous Driving with Test-Time Training
  • PPR-2503.12820 Hydra-MDP++: Advancing End-to-End Driving via Expert-Guided Hydra-Distillation
  • PPR-2504.01941 End-to-End Driving with Online Trajectory Evaluation via BEV World Model
  • PPR-2504.19580 ARTEMIS: Autoregressive End-to-End Trajectory Planning With Mixture of Experts for Autonomous Driving
  • PPR-2505.09315 TransDiffuser: Diverse Trajectory Generation with Decorrelated Multi-modal Representation for End-to-end Autonomous Driving
  • PPR-2505.15111 iPad: Iterative Proposal-Centric End-to-End Autonomous Driving
  • PPR-2506.06659 DriveSuprim: Towards Precise Trajectory Selection for End-to-End Planning
  • PPR-2506.06664 Generalized Trajectory Scoring for End-to-end Multimodal Planning
  • PPR-2506.08052 ReCogDrive: A Reinforced Cognitive Framework for End-to-End Autonomous Driving
  • PPR-2506.09985 V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning
  • PPR-2506.13757 AutoVLA: A Vision-Language-Action Model for End-to-End Autonomous Driving with Adaptive Reasoning and Reinforcement Fine-Tuning
  • PPR-2506.24113 Epona: Autoregressive Diffusion World Model for Autonomous Driving
  • PPR-2507.00603 World4Drive: End-to-End Autonomous Driving via Intention-Aware Physical Latent World Model
  • PPR-2507.04049 DIVER: Reinforced Diffusion Breaks Imitation Bottlenecks in End-to-End Autonomous Driving
  • PPR-2507.17596 PRIX: Learning to Plan From Raw Pixels for End-to-End Autonomous Driving
  • PPR-2510.04333 RAP: 3D Rasterization Augmented End-to-End Planning
  • PPR-2510.08562 ResAD: Normalized Residual Trajectory Modeling for End-to-End Autonomous Driving
  • PPR-2510.12796 DriveVLA-W0: World Models Amplify Data Scaling Law in Autonomous Driving
  • PPR-2512.07745 DiffusionDriveV2: Reinforcement Learning-Constrained Truncated Diffusion Modeling in End-to-End Autonomous Driving
  • PPR-2512.10226 Latent Chain-of-Thought World Modeling for End-to-End Driving
  • PPR-2601.05083 Driving on Registers
  • PPR-2601.22032 Drive-JEPA: Video JEPA Meets Multimodal Trajectory Distillation for End-to-End Driving
  • PPR-2603.14482 V-JEPA 2.1: Unlocking Dense Features in Video Self-Supervised Learning
  • PPR-2603.24581 Latent-WAM: Latent World Action Modeling for End-to-End Autonomous Driving
  • PPR-2603.29163 SparseDriveV2: Scoring is All You Need for End-to-End Autonomous Driving
  • PPR-2605.09701 DriveFuture: Future-Aware Latent World Models for Autonomous Driving
  • PPR-2605.15120 CLOVER: Closed-Loop Value Estimation and Ranking for End-to-End Autonomous Driving Planning
  • PPR-2605.19771 Beyond Imitation: Learning Safe End-to-End Autonomous Driving from Hard Negatives
  • PPR-2605.31476 IDOL: Inverse-Dynamics-Guided Future Prediction for End-to-End Autonomous Driving
  • PPR-2607.29031 Auto-JEPA: A Latent World Model of Continuous Intent for End-to-End Autonomous Driving
  • PPR-2608.19085 DA-WAM: Decision-Aligned Future Latents for Driving World Models
  • BMK-006 NAVSIM non-reactive planning benchmark (v1 PDMS / v2 EPDMS two-stage)
  • CPT-017 PDM Score (PDMS)