Embodied AI Evidence Hub具身智能证据知识库

数据工程与质量 · Data Engineering and Quality

近一年空间数据生产难点及具身机器人与智能驾驶数据难点异同

Spatial Data Production for Robotics and Autonomous Driving

更新于

30 秒结论

先读这个判断

本文考察 2025 年 7 月 27 日至 2026 年 7 月 27 日之间的论文，问题是：空间数据的生产难点究竟在哪里，具身机器人与智能驾驶有哪些共性和结构性差异？本轮为 scoping review，候选池 225 篇，恢复 44 篇完整非 OCR 全文，最终选取 17 篇通过主张支持审计的论文作为正式证据。这一范围能支撑生产机制的比较，但不等于全部产业成本调查；新近 arXiv 论文中的一些结果也仍需要更大规模复现。

精读论文
17
正式证据事件
19
知识单元
EA-DATA, EA-SENSOR, EA-4D
审计状态
当前结算 run 已通过仓库校验

v0 · 完整正文

研究综述

研究边界

本文考察 2025 年 7 月 27 日至 2026 年 7 月 27 日之间的论文，问题是：空间数据的生产难点究竟在哪里，具身机器人与智能驾驶有哪些共性和结构性差异？本轮为 scoping review，候选池 225 篇，恢复 44 篇完整非 OCR 全文，最终选取 17 篇通过主张支持审计的论文作为正式证据。这一范围能支撑生产机制的比较，但不等于全部产业成本调查；新近 arXiv 论文中的一些结果也仍需要更大规模复现。

中心判断

空间数据的主要瓶颈不是“捕获更多像素或点”，而是生产可用于决策的时空真值：需要把多模态观测放进同一坐标系和时间轴，补足遮挡与接触中的不可观状态，绑定动作、意图和拓扑语义，覆盖自然分布中稀缺的失败，并用真实或可反应的闭环检验其后果。具身机器人最难的是“执行真值”：某个身体在接触中究竟做了什么、是否可执行；智驾最难的是“覆盖真值”：道路世界是否被持续、拓扑正确地记录，稀有风险是否足够密集地被复现。

一、空间真值先是一条生产链，之后才是一个文件

空间数据至少包含四层真值：度量几何（位姿、深度、尺度），拓扑与语义（车道属性、物体身份、可通行/可操作性），时间动态（运动、遮挡、状态转移），以及动作后果（执行一个动作后，世界如何反应）。任何一层失真，都可以让整条数据变得不可学习或不可验证。

智驾的 HRDX 是一个典型例子：城市峡谷中 GNSS 会受多径与遮挡干扰，因此需要 LiDAR–惯导里程计与 INS 融合；相机、LiDAR 还要受统一主时钟约束，扫描需要运动补偿。但坐标对齐之后，仍需专家在全局注册的稠密点云上标注道路几何、拓扑和法规属性。KITScenes 更进一步：它使用航拍影像提供无遮挡俯视图，再用车载传感器定位立体交通设施，将两层地图融合、人工复核，通过拓扑检查、规划仿真和闭环试驾验收。这说明“地图标注”实际上是多源测绘、版本更新和软件验收的组合。

机器人一侧的同类问题更近。ActiveMimic 表明，第一视角视频中的手腕轨迹与相机自运动天然耦合，坐标系不统一就会把头部/相机位移错当成手部动作。HapTile 则把时间对齐推到触觉、动作和状态层：检查时间戳缺口、损坏轨迹和动作—状态一致性。因此，两个行业的第一个共性是：时空对齐不是数据清洗的小步骤，而是数据是否有意义的先决条件。

二、真正贵的不是观测，而是不可观状态的补全

驾驶场景里，遮挡导致的未知通常是远处车辆、弱势交通参与者或被挡住的车道结构；机器人里，未知往往是“最后一厘米”的接触力、滑移、物体局部变形和夹爪状态。前者能部分借助多车/路侧视角、长程地图与航拍视角补全；后者往往必须把触觉、力/力矩和本体感觉放入控制环。

这也是纯视觉机器人数据的边界。UMI-3D 指出，单目视觉 SLAM 在遮挡、动态变化、弱特征区域会跟踪失败，需要 LiDAR 中心的 3D 位姿估计扩大可采任务分布。3PoinTr 则不简单丢弃含遮挡的轨迹，而只屏蔽不可见的点—时刻损失，尽量保留任务关键物体点。HT-Bench 已将触觉—第一视角配对数据扩展到百万级帧和任务级分布外评测，但这仍首先是“接触表征”的底座，不自动等于长时闭环操作能力。

由此可以推断，空间数据的边际价值不与传感器数量单调增长，而取决于新模态是否补足了当前任务的不可观变量。若这一推断正确，那么去掉某个模态后，关联的闭环失败应当在接触、遮挡或恶劣环境压力测试中系统增加；如果没有，该模态更可能是冗余成本。

三、两类系统都缺长尾，但缺的长尾不同

智驾可以被动积累海量里程，问题是真正危险的事件在自然分布中极稀缺。SinD 2.0 把冲突描述得很清楚：日常合规交互占据大部分时长，而系统边界测试需要视觉遮挡、狭窄可行域、路权争夺等稀有语义触发条件。更难的是，自动风险标签会继承轨迹估计、地图和信号灯绑定误差，可重现并不等于已成为人工验证的真值。

机器人的长尾并非主要来自一个“稀有场景库”，而是对象、工具、材料、接触相位、操作阶段和本体的组合爆炸。一条成功示教通常只说明“在这一次接触中做对了”，却不告诉模型如何在滑移、对齐偏差、空抓或外界扰动中恢复。TacForeSight 因此专门采集恢复交互：若目标能力包含扰动后重新建立接触，数据集就必须包含这些过程，而不能只增加成功次数。

这一差异决定了数据飞轮的不同形态：智驾应以“车队日志→风险挖掘→可反应仿真→闭环压测”为主；机器人应以“成功示教→人为扰动→失败/恢复→目标本体复验”为主。

四、合成数据可以放大覆盖，但无法凭空制造真值

两个领域都在用世界模型、仿真和重建补数据。但这些方法的价值是“扩展已有真值的支撑域”，而不是取消真值生产。

OccSim 能从单帧和未来自车动作生成公里级 3D occupancy，并用合成流预训练占用预测模型。可论文同时承认，语义 occupancy 本身仍依赖人工语义标注，可用训练帧数不足 10 万。CARLA-GS 将视觉重建、LLM 意图推理和 CARLA 物理执行拆成模块，用于合成危险交互；但在其小规模实验中，人工检查的 LLM 输出有效率只有 29.4%。这些结果不是否定合成数据，而是将它定位为一个需要校验的生产工序。

具身侧也一样。仿真过滤人类视频的研究 显示，人类视频只有经过机器人可达性、任务兼容性和抓取适合性过滤，才更接近可执行示教。这意味着“合成/仿真”的最可靠用法不是宣称与现实无差别，而是做筛选器、压力测试器和数据补全器。

五、具身机器人与智驾：同一组质量问题，不同的物理约束

比较维度具身机器人智能驾驶共同机制
核心空间单元物体—手/夹爪—工具的局部关系自车—道路拓扑—多交通体的大范围关系都需要对象身份、坐标与时间连续性
最难真值接触、力、滑移、可执行动作地图拓扑、占用、交互风险与 ODD 覆盖纯视觉都存在遮挡与弱可观性
动作的可迁移性弱：控制器、本体、夹爪改变动作含义相对较强：车辆运动学更统一，但车型与控制栈仍有差异动作标签必须绑定执行上下文
数据获取主动交互为主，单条成本高，可主动施加扰动被动车队日志为主，量大，但稀有风险密度低都需要有目的地重采失败和边界样本
语义寿命对象/任务/本体变化快，控制参数影响大地图与交通法规相对稳定，但必须持续更新数据都需要版本、设备与语义谱系
最终验收目标机器人在真实接触闭环中的成功、恢复与安全可反应多交通体闭环与真车/高可信仿真验证开环准确率不能单独定义数据质量

具身侧最根本的特殊性是“动作属于身体”。SPACE 指出，同一条记录命令在不同控制器、硬件个体与部署动力学下会产生不同运动。ACE-Ego-0 也显示，联合人类视频和机器人数据时，必须显式处理坐标、本体、物理时间与标签可信度。智驾侧更重要的特殊性是“真值属于 ODD”：一座城市、一种天气、一类路口上的有效空间覆盖，无法自动代表另一地区的驾驶文化和风险分布。

六、对数据平台的启示

第一，数据 schema 应围绕“可证伪的事件”，而不是只围绕传感器文件。每个 episode/log 至少应保留：坐标系与标定版本、主时钟与模态延迟、可信监督字段、动作与执行控制器、接触/风险事件、失败与恢复段，以及真实闭环结果。StandardE2E 对智驾的警示是，格式、API、坐标规约和模态覆盖的碎片化本身就会阻断跨数据集实验；但统一接口不能把缺失的模态或不同的标签语义变得等价。

第二，质量门禁应采用乘法观，而非加分表。可以把有效空间数据近似表达为：

有效性 ≈ 时空对齐 × 关键状态可观性 × 语义/动作正确性 × 长尾覆盖 × 闭环可验证性

任一项接近零，更高分辨率或更多时长都很难补救。因此验收指标应从“采集了多少”改为“可用监督字段比例、对齐残差、遮挡/接触恢复率、风险事件密度、闭环增益与失败拒识率”。

第三，两个行业都应保留少量高质量“锚定真值”来校正大规模弱监督。智驾的锚定真值是高精度位姿、人工复核地图、真实风险事件和闭环试验；机器人的锚定真值是目标机器人的真实动作、接触、失败恢复和终局成功。人类视频、仿真数据、自动标注和世界模型可以绕着它们扩展，却不应取代它们。

条件、分歧与待验证点

本轮证据尚不支持三个过度强结论。第一，不能说所有机器人任务都必须上触觉和 3D；任务若在开放空间中纯视觉已可观，新模态可能只增加成本。第二，不能由一个生成模型的高真实度推出它能用于安全裁决；动作响应、交互反应和闭环相关性必须单独验证。第三，不能把本轮公开论文的数据集成本直接换算成企业规模的单位经济性；论文通常不报告完整的标注工时、设备维护、事故责任和数据合规成本。

下一步最值得做的实验，是在同一生产预算下比较三种路线：增加被动观测、增加主动失败/恢复采集，以及增加合成数据与人工复核。真正有意义的因变量不是开环准确率，而是目标 ODD/目标机器人上的闭环成功、恢复、失败拒识和安全边界。

结论

具身机器人和智能驾驶并非两个毫不相干的数据行业。它们都在将不完全、异步、有遮挡的多模态观测，转化为能对动作后果负责的时空真值。二者的差异在于真值的最终锚点：机器人锚定在特定身体的接触执行，智驾锚定在特定 ODD 中的地图、多体交互与安全覆盖。所以，最值得投资的不是某一种“万能空间数据”，而是一条能够持续对齐、暴露不可观状态、富集失败并回到闭环验收的数据生产系统。

References

  1. ActiveMimic: Egocentric Video Pretraining with Active Perception
  2. HapTile: A Haptic-Informed Vision-Tactile-Language-Action Dataset for Contact-Rich Imitation Learning
  3. UMI-3D: Extending Universal Manipulation Interface from Vision-Limited to 3D Spatial Perception
  4. 3PoinTr: 3D Point Tracks for Learning Manipulation from Unconstrained Human Videos
  5. HT-Bench: Benchmarking and Learning Dexterous Full-Hand Tactile Representations with Egocentric Vision
  6. TacForeSight: Force-Guided Tactile World Model for Contact-Rich Manipulation
  7. SPACE: Enabling Learning from Cross-Robot Data Toward Generalist Policies
  8. ACE-Ego-0: Unifying Egocentric Human and Robotic Data for VLA Pretraining
  9. KITScenes Multimodal
  10. HRDX: A Large-Scale Vector HD-Map Dataset
  11. SinD 2.0
  12. OccSim: Multi-kilometer Simulation with Long-horizon Occupancy World Models
  13. CARLA-GS
  14. StandardE2E
  15. Imitating What Works

Accepted evidence

去重论文引用

  1. 3PoinTr: 3D Point Tracks for Learning Manipulation from Unconstrained Human VideosAdam Hung, Bardienus Pieter Duisterhof, Jeffrey Ichnowski · 2026
  2. 4D-CAAL: 4D Radar-Camera Calibration and Auto-Labeling for Autonomous DrivingShanliang Yao, Zhuoxiao Li, Runwei Guan, Kebin Cao, Meng Xia, Fuping Hu, Sen Xu, Yong Yue et al. · 2026
  3. ACE-Ego-0: Unifying Egocentric Human and Robotic Data for VLA PretrainingHao Li, Ganlong Zhao, Yufei Liu, Haotian Hou, Guoquan Ye, Tongyan Fang, Chunxiao Liu, Siyuan Huang et al. · 2026
  4. ActiveMimic: Egocentric Video Pretraining with Active PerceptionXingyao Lin, Guojin Zhong, Tianyi Lu, Ziyi Ye, Yichen Zhu, Zuxuan Wu, Yu-Gang Jiang · 2026
  5. CARLA-GS: Decoupling Representation, Reasoning, and Physics Simulation for Autonomous Driving Corner-Case SynthesisKaicong Huang, Meng Ma, Ruimin Ke · 2026
  6. Embody4D: A Generalist Data Engine for Embodied 4D World ModelingPeiyan Tu, Hanxin Zhu, Jingwen Sun, Shaojie Ren, Cong Wang, Yuyan Xu, Jiayi Luo, Xiaoqian Cheng et al. · 2026
  7. HapTile: A Haptic-Informed Vision-Tactile-Language-Action Dataset for Contact-Rich Imitation LearningAmirhosein Alian, Yongqiang Zhao, Shiyi Gu, Xuyang Zhang, Zhuo Chen, Christopher E. Mower, Haitham Bou-Ammar, Shan Luo · 2026
  8. HRDX: A Large-Scale Vector HD-Map DatasetSahith Reddy Chada, Isht Dwivedi, Nirav Savaliya · 2026
  9. HT-Bench: Benchmarking and Learning Dexterous Full-Hand Tactile Representations with Egocentric VisionYuzhe Huang, Jiaping Wu, Jiaming Jiang, Hezhe Lin, Aikebaier Aierken, Yunlong Wang, Kun Cheng, Ziyuan Jiao et al. · 2026
  10. Imitating What Works: Simulation-Filtered Modular Policy Learning from Human VideosAlbert J. Zhai, Kuo-Hao Zeng, Jiasen Lu, Ali Farhadi, Shenlong Wang, Wei-Chiu Ma · 2026
  11. OccSim: Multi-kilometer Simulation with Long-horizon Occupancy World ModelsTianran Liu, Shengwen Zhao, Mozhgan Pourkeshavarz, Weican Li, Nicholas Rhinehart · 2026
  12. SinD 2.0: A Multi-City UAV Dataset with Semantic Risk Annotations for SOTIF-Oriented Safety Validation at Signalized IntersectionsYunwei Li, Shengjie Fu, Chunrong Chen, Chengxiang Zhao, Yuchen Fan, Mingyu Zhu, Yanchao Xu, Yuxin Zhang et al. · 2026
  13. SPACE: Enabling Learning from Cross-Robot Data Toward Generalist PoliciesHaeone Lee · 2026
  14. StandardE2E: A Unified Framework for End-to-End Autonomous Driving DatasetsStepan Konev · 2026
  15. TacForeSight: Force-Guided Tactile World Model for Contact-Rich ManipulationYujie Zang, Yuhang Zheng, Xian Nie, Yupeng Zheng, Shuai Tian, Songen Gu, Chen Gao, Zining Wang et al. · 2026
  16. The Road Ahead in Autonomous Driving: The KITScenes Multimodal DatasetRichard Schwarzkopf, Fabian Immel, Alexander Blumberg, Jonas Merkert, Nils Rack, Kaiwen Wang, Fabian Konstantinidis, Julian Truetsch et al. · 2026
  17. UMI-3D: Extending Universal Manipulation Interface from Vision-Limited to 3D Spatial PerceptionZiming Wang · 2026
完整证据附录（19 条正式证据事件）
  • Time range: 2025-07-27..2026-07-27
  • Events: 19
  • 每个事件一节,标题即锚点;trace-map 中的 event 链接跳转到这里。

EA-4DDATA-READ-0015

  • Claim: Embody4D targets the sparse-view limitation of robot video data with monocular-to-novel-view video transformation and a 3D-aware compositional synthesis pipeline for training data.
  • Stance: support | Confidence: direct
  • Paper: 2605.01799 Embody4D: A Generalist Data Engine for Embodied 4D World Modeling
  • Locator: Abstract (full-text section)
  • Evidence: The abstract ties fixed or sparse viewpoints to partial observations and introduces both novel-view video generation and a compositional synthesis pipeline to address data scarcity.
  • Quote: “Abstract Embodied agents require robust and comprehensive 3D spatiotemporal representations to support spatial reasoning, manipulation understanding, and downstream decision making. However, existing robot data are typically captured from fixed or sparse viewpoints, providing only partial and view-dependent observations, which limits multi-view perception and generalization across viewpoints. Given the difficulty of collecting additional viewpoints in real-world settings, we propose Embody4D, a”
  • Authors: peiyan-tu; hanxin-zhu; jingwen-sun; et al.

EA-SPATIAL-2026-0003

  • Claim: Production-grade HD-map ground truth is a multi-source and multi-pass QA product: occlusion-free aerial data and onboard 3D sensing must be fused, map changes checked, geometry and topology manually cross-reviewed, and structural consistency validated.
  • Stance: support | Confidence: direct
  • Paper: 2606.02956 The Road Ahead in Autonomous Driving: The KITScenes Multimodal Dataset
  • Locator: Appendix E Annotation Protocol and Quality Control
  • Evidence: The annotation appendix describes complementary map passes, fusion, manual review and multiple automated integrity checks.
  • Quote: “Map creation is split into two complementary annotation passes. Road-level content, like lane geometry, road markings, lane topology, crosswalks, as well as BEV traffic light and sign positions, are annotated from aerial imagery, which provides a geo-referenced, occlusion-free top-down view.”
  • Authors: richard-schwarzkopf; fabian-immel; alexander-blumberg; et al.

EA-SPATIAL-2026-0005

  • Claim: Fleet-grade spatial truth requires an explicit pose-and-time production stack: urban GNSS degradation is mitigated with LiDAR-inertial and INS fusion, cameras and LiDAR are disciplined to a master clock, and scans are motion-compensated to image timestamps.
  • Stance: support | Confidence: direct
  • Paper: 2606.17080 HRDX: A Large-Scale Vector HD-Map Dataset
  • Locator: 3.3 Localization and Sensor Synchronization
  • Evidence: The dataset construction section describes the degradation mechanism and concrete fusion, clock and motion-compensation measures.
  • Quote: “In dense urban areas, GNSS degrades due to multipath and occlusions. We mitigate this by integrating LiDAR–Inertial Odometry (LIO) and fusing it with INS, yielding robust, drift-free localization.”
  • Authors: sahith-reddy-chada; isht-dwivedi; nirav-savaliya

EA-SPATIAL-2026-0006

  • Claim: Vector HD-map ground truth is built on a globally registered dense 3D map and still requires expert geometry/topology annotation plus visual verification; high-quality coordinates alone do not supply regulatory semantics.
  • Stance: support | Confidence: direct
  • Paper: 2606.17080 HRDX: A Large-Scale Vector HD-Map Dataset
  • Locator: 3.5 Data Annotation and Ground-Truth Labels
  • Evidence: The annotation section states how the metric substrate and expert semantic labeling are combined.
  • Quote: “To construct reliable ground truth for static map elements, we accumulate LiDAR frames at an interval of 5m across each scene into a globally registered, high-density 3D map. This merged point cloud serves as the metric substrate for annotation. Expert annotators use the fused pointcloud to annotate road geometry in vector form (polylines and polygons) and encode topology (e.g., intersection centerlines), followed by visual verification to ensure spatial precision.”
  • Authors: sahith-reddy-chada; isht-dwivedi; nirav-savaliya

EA-DQ-YEAR-READ-0003

  • Claim: 人类视频能扩大机器人学习数据来源，但其质量取决于机器人可执行性和任务兼容性；仿真过滤可把不可达、估计错误或 grasp 不兼容的轨迹排除或标注出来。
  • Stance: conditional | Confidence: direct
  • Paper: 2602.13197 Imitating What Works: Simulation-Filtered Modular Policy Learning from Human Videos
  • Locator: 3.3 Trajectory and Grasp Filtering via Simulation
  • Evidence: PSI 将人类演示转换为 6DoF object pose trajectories 后在仿真中执行，用于过滤不适合机器人学习的数据；不适合原因包括 pose estimation errors 和机器人 physically unachievable trajectories，并生成 grasp suitability labels 以学习 task-oriented grasping。
  • Quote: “Now that we have converted the human demonstrations into 6 DoF object pose trajectories, the next step is to execute them on a robot in simulation. This serves two purposes. One is to filter out those that may not be suitable for robot learning. There are two main reasons a trajectory may be unsuitable. First, pose estimation errors can lead to inaccurate trajectories. Second, the extracted trajectory may not be physically achievable by the robot. In either case, it would be harmful to train the”
  • Authors: albert-j-zhai; kuo-hao-zeng; jiasen-lu; et al.

EA-4DDATA-READ-0009

  • Claim: 3PoinTr 保留含暂时遮挡点的轨迹，只屏蔽不可见的单个点—时间损失，因而能继续利用操作中任务关键的物体点监督。
  • Stance: conditional | Confidence: direct
  • Paper: 2603.08485 3PoinTr: 3D Point Tracks for Learning Manipulation from Unconstrained Human Videos
  • Locator: 4.3 Results: 3D Point Track Prediction
  • Evidence: 结果段对比了删除整条不可见轨迹的基线与仅屏蔽不可见 point-timestep 损失的 3PoinTr。
  • Quote: “The primary advantage of 3PoinTr is that it trains on data General Flow ignores. Real-world points are often temporarily occluded; General Flow removes any trajectory with invisible point-timestep pairs during preprocessing, whereas 3PoinTr retains all trajectories and masks losses for individual invisible point-timestep pairs. This provides additional supervision over task-critical object points that are temporarily occluded during manipulation. For example, in the Throw Away Paper task, every”
  • Authors: adam-hung; bardienus-pieter-duisterhof; jeffrey-ichnowski

EA-4DDATA-READ-0007

  • Claim: HapTile 的数据质量控制在机器人控制环中同步全部模态，检查空轨迹、损坏轨迹和时间戳缺口，并验证动作—状态一致性。
  • Stance: conditional | Confidence: direct
  • Paper: 2606.04825 HapTile: A Haptic-Informed Vision-Tactile-Language-Action Dataset for Contact-Rich Imitation Learning
  • Locator: 3.2 Synchronization and Data Quality Control
  • Evidence: 数据质量段明确记录了控制环同步、时间戳缺口检查、损坏轨迹剔除和 action-state consistency 检查。
  • Quote: “All data modalities are synchronized through the robot control loop. For policy learning, actions are converted to a unified 7D end-effector delta representation (1) where are translational deltas, are rotational deltas, and is the gripper command. This decouples learning from the exact robot configuration, enabling cross-embodiment by focusing the policy on local contact adjustment from tactile feedback. Several quality checks are applied to every collected trajectory. Empty or corrupted trajec”
  • Authors: amirhosein-alian; yongqiang-zhao; shiyi-gu; et al.

EA-4DDATA-READ-0008

  • Claim: 如果目标包含扰动恢复，数据集必须显式收集nominal demonstrations和recovery interaction data；否则模型很难学习接触丢失后的再建立。
  • Stance: conditional | Confidence: direct
  • Paper: 2606.11184 TacForeSight: Force-Guided Tactile World Model for Contact-Rich Manipulation
  • Locator: IV-B 2 Perturbation-Aware Evaluation
  • Evidence: TacForeSight在perturbation-aware设置中额外收集恢复示教，让外部扰动发生在执行中并要求示教者重新建立稳定接触；完整模型在三个扰动任务上报告90%、85%、85%平均完成/成功分数。
  • Quote: “Policies in this setting are trained using both nominal demonstrations and recovery interaction data.”
  • Authors: yujie-zang; yuhang-zheng; xian-nie; et al.

EA-PRETRAIN-DATA-2026-0001

  • Claim: 异构来源应扩大，但在联合预训练前必须将空间坐标、本体形态、物理时间和标签可靠性显式对齐或条件化；否则会降低动作学习性能。
  • Stance: conditional | Confidence: direct
  • Paper: 2606.17200 ACE-Ego-0: Unifying Egocentric Human and Robotic Data for VLA Pretraining
  • Locator: 5.2 Ablation Studies, Figure 5(b)
  • Evidence: 三项组件消融均降低 RoboCasa 成功率，其中去掉人类伪动作可靠性加权的降幅最大。
  • Quote: “Removing morphology tokens makes the success rate drop from 72.8% to 70.9%”
  • Authors: hao-li; ganlong-zhao; yufei-liu; et al.

EA-SPATIAL-2026-0001

  • Claim: 4D radar auto-labeling remains bounded by the quality of the cross-modal teacher and correspondence: severe occlusion or lighting can degrade visual segmentation, while image-plane overlap makes radar point-to-instance assignment ambiguous.
  • Stance: limit | Confidence: direct
  • Paper: 2601.21454 4D-CAAL: 4D Radar-Camera Calibration and Auto-Labeling for Autonomous Driving
  • Locator: IV-C Discussion
  • Evidence: The discussion explicitly names visual-segmentation degradation and overlap ambiguity as residual annotation limits after the proposed calibration and refinement pipeline.
  • Quote: “For auto-labeling, annotation quality is bounded by visual instance segmentation performance, which may degrade under severe occlusion or extreme lighting. Additionally, significant object overlap in the image plane leads to ambiguous point-to-instance correspondences.”
  • Authors: shanliang-yao; zhuoxiao-li; runwei-guan; et al.

EA-SPATIAL-2026-0002

  • Claim: Synthetic occupancy generation does not escape the ground-truth bottleneck: OccSim reports that semantic occupancy still requires manual semantic annotation, leaving fewer than 100,000 training frames and constraining exploration of scaling limits.
  • Stance: limit | Confidence: direct
  • Paper: 2603.28887 OccSim: Multi-kilometer Simulation with Long-horizon Occupancy World Models
  • Locator: 10 Potential limitation and Future work
  • Evidence: The limitations section directly attributes the small training volume to manual semantic occupancy annotation.
  • Quote: “Compared to image data, semantic occupancy data requires manual semantic annotation, which limits the total amount of available training data to fewer than 100,000 frames. This volume is significantly smaller than the amount of available RGB video data, and it also restricts our ability to explore the upper limits of the model’s performance.”
  • Authors: tianran-liu; shengwen-zhao; mozhgan-pourkeshavarz; et al.

EA-UMI-READ-0004

  • Claim: Original vision-only UMI data can fail to be useful in scenes with occlusion, dynamic changes, feature-poor regions, or tracking failure; adding LiDAR-centric 3D sensing improves data quality and expands the feasible task distribution.
  • Stance: limit | Confidence: direct
  • Paper: 2604.14089 UMI-3D: Extending Universal Manipulation Interface from Vision-Limited to 3D Spatial Perception
  • Locator: Abstract (full-text section)
  • Evidence: The HTML full text states that monocular visual SLAM makes UMI vulnerable to occlusions, dynamic scenes, and tracking failures, and reports that LiDAR-centric SLAM improves pose-estimation robustness and demonstration data quality under challenging real-world conditions.
  • Quote: “Abstract We present UMI-3D, a multimodal extension of the Universal Manipulation Interface (UMI) for robust and scalable data collection in embodied manipulation. While UMI enables portable, wrist-mounted data acquisition, its reliance on monocular visual SLAM makes it vulnerable to occlusions, dynamic scenes, and tracking failures, limiting its applicability in real-world environments. UMI-3D addresses these limitations by introducing a lightweight and low-cost LiDAR sensor tightly integrated i”
  • Authors: ziming-wang

EA-SPATIAL-2026-0004

  • Claim: Driving-data fragmentation is itself a production bottleneck: dataset-specific file formats, APIs, calibration conventions and modality coverage make preprocessing repeatedly reimplemented and cross-dataset experiments expensive and brittle.
  • Stance: limit | Confidence: direct
  • Paper: 2606.04271 StandardE2E: A Unified Framework for End-to-End Autonomous Driving Datasets
  • Locator: 1 Introduction
  • Evidence: The introduction explicitly describes recurring per-project preprocessing and brittle adapter work caused by heterogeneous dataset conventions.
  • Quote: “In practice, this fragmentation imposes a recurring tax on end-to-end driving research. Per project, researchers re-implement preprocessing pipelines from raw protobuf, parquet, pickled-frame, or city-wide GPKG sources into a representation their model can consume.”
  • Authors: stepan-konev

EA-EGO-2026-0016

  • Claim: Ego-centric wrist trajectory 与相机自运动天然耦合；若不统一参考系，动作监督会把 camera rotation/translation 错算为 hand motion。
  • Stance: limit | Confidence: direct
  • Paper: 2606.06194 ActiveMimic: Egocentric Video Pretraining with Active Perception
  • Locator: 3 Method
  • Evidence: 方法段明确说明 current-frame wrist pose 与 first-frame camera path 的坐标差异会混合两类位移。
  • Quote: “using these wrist poses directly as action supervision would therefore conflate wrist movement with camera motion”
  • Authors: xingyao-lin; guojin-zhong; tianyi-lu; et al.

EA-SPATIAL-2026-0007

  • Claim: Synthetic driving corner cases still require human or programmatic validation: even with explicit collision zones and CARLA execution, the LLM may fail under multiple constraints and only 29.4% of generated outputs were judged valid by manual inspection.
  • Stance: limit | Confidence: direct
  • Paper: 2607.07601 CARLA-GS: Decoupling Representation, Reasoning, and Physics Simulation for Autonomous Driving Corner-Case Synthesis
  • Locator: IV-B Quantitative Evaluation
  • Evidence: The evaluation directly reports LLM instability and manual validity for the generated outputs.
  • Quote: “However, LLM instability can still limit planning reliability. As indicated by the zone hit rate and variance, the LLM may fail to reach the target zone under multiple constraints, even with explicit prompts, as shown in Case 3 of Fig. 6 . Manual inspection shows that 29.4% of LLM-generated outputs are valid.”
  • Authors: kaicong-huang; meng-ma; ruimin-ke

EA-SPATIAL-2026-0008

  • Claim: Safety-critical driving data is hard to produce because the desired distribution conflicts with natural occurrence: geographic coverage is narrow, routine safe interactions dominate, and raw trajectories lack semantic labels for triggers such as occlusion and constrained maneuver space.
  • Stance: limit | Confidence: direct
  • Paper: 2607.16943 SinD 2.0: A Multi-City UAV Dataset with Semantic Risk Annotations for SOTIF-Oriented Safety Validation at Signalized Intersections
  • Locator: I Introduction
  • Evidence: The introduction explicitly identifies geographic homogeneity, rare critical events and missing trigger semantics as separate data limitations.
  • Quote: “Effective safety validation also requires a high degree of criticality, which conflicts with the sparsity of dangerous scenarios in naturalistic driving data. Rigorous evaluation must push the system to its performance boundaries, yet naturalistic datasets are dominated by safe, routine interactions.”
  • Authors: yunwei-li; shengjie-fu; chunrong-chen; et al.

EA-SPATIAL-2026-0009

  • Claim: Rule-generated risk labels remain provisional ground truth because they inherit errors from trajectories, maps and signal binding; auditability does not substitute for manual accuracy and threshold-sensitivity validation.
  • Stance: limit | Confidence: direct
  • Paper: 2607.16943 SinD 2.0: A Multi-City UAV Dataset with Semantic Risk Annotations for SOTIF-Oriented Safety Validation at Signalized Intersections
  • Locator: V-E Narrow Feasible-Area Scenarios
  • Evidence: The paper explicitly describes the inheritance of upstream errors and the unfinished manual audit.
  • Quote: “This design makes each label reproducible and inspectable, but it also means that the labels inherit map, signal-binding, and trajectory-estimation errors. We therefore use the semantic layer as a queryable scenario-mining index whose thresholds and evidence can be inspected and revised, rather than as a finalized benchmark of human-verified annotation accuracy.”
  • Authors: yunwei-li; shengjie-fu; chunrong-chen; et al.

EA-ALIGN-READ-0001

  • Claim: A recorded robot action is not a universal supervision signal: the same command can produce different motions across controllers, embodiments, hardware units, and deployment-time dynamics.
  • Stance: limit | Confidence: direct
  • Paper: 2606.24049 SPACE: Enabling Learning from Cross-Robot Data Toward Generalist Policies
  • Locator: 3.2 Inconsistency of Control Commands across Robots
  • Evidence: SPACE predicts Cartesian state deltas as a shared end-effector-space representation and uses an action adapter to convert them into robot-specific control commands, improving cross-robot and dynamics-shift robustness.
  • Quote: “Recent work has scaled robot learning by training policies on data from multiple embodiments [ 27 , 23 , 32 ] , often using the Cartesian delta action space [ 23 , 32 ] since it is less dependent on robot-specific kinematics and invariant to base-frame translation [ 18 , 14 ] . In practice, this is typically realized by predicting Cartesian delta control commands that are fed to the underlying robot controller [ 23 , 32 ] . Figure 2: Different robots (e.g., UR5 vs. Franka Research 3) require dif”
  • Authors: haeone-lee

EA-TACTILE-2026-0001

  • Claim: 近一年触觉表征研究开始从小规模单任务管线走向大规模全手触觉—第一视角配对数据和多任务、任务级 OOD 基准；HT-Bench 以约 1000 万 RGB 帧、780 万触觉帧和 226 项任务测量接触结构、跨模态对齐与时间动态。
  • Stance: support | Confidence: direct
  • Paper: 2606.19161 HT-Bench: Benchmarking and Learning Dexterous Full-Hand Tactile Representations with Egocentric Vision
  • Locator: Abstract; 3 HT-Bench: A Multi-Task Tactile Evaluation Benchmark
  • Evidence: 摘要和基准设计章节直接给出数据规模、四项评测任务与任务级 OOD 划分。
  • Quote: “comprising 10M RGB frames and 7.8M tactile frames collected across 226 tasks.”
  • Authors: yuzhe-huang; jiaping-wu; jiaming-jiang; et al.

References

Interactive Wiki

进入交互阅读

静态页采用默认版本；v0、v0、v0可在同一交互 Wiki 中阅读，不建立相互竞争的索引页。

用 AI 继续研究 v0 v0 v0

Method and limits

研究方法、证据边界与引用说明

本页由当前结算的证据层自动派生，不改写已结算 run。论文只有在完整非 OCR 全文、结构化论文笔记和 claim-support audit 通过后，才进入接受证据。

页面结论是对所列论文与研究窗口的综合，不等同于无限外推。引用本专题时，请同时保留本页 canonical 地址、访问日期，并在需要时回溯原论文。