数据工程与质量 · Data Engineering and Quality

近一年具身智能预训练模型对数据源与采集参数的要求

Data Requirements for Embodied Pretraining

更新于

30 秒结论

先读这个判断

本文回顾 2025 年 7 月 26 日至 2026 年 7 月 26 日的具身智能预训练、VLA、人类第一视角视频和跨本体数据研究。检索注册表包含 1,398 篇候选,40 篇已恢复完整非 OCR 全文,其中 20 篇通过深读、主张支持审计并进入本综述。

精读论文
20
正式证据事件
33
知识单元
EA-DATA, EA-HARDWARE, EA-SENSOR, EA-MODEL, EA-XEMBODIMENT
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

摘要

本文回顾 2025 年 7 月 26 日至 2026 年 7 月 26 日的具身智能预训练、VLA、人类第一视角视频和跨本体数据研究。检索注册表包含 1,398 篇候选,40 篇已恢复完整非 OCR 全文,其中 20 篇通过深读、主张支持审计并进入本综述。

核心结论是:通用预训练不要求所有采集设备、机位和采集方法尽可能一致;它要求任务、场景、物体、视点、行为和本体覆盖尽可能多样,同时对坐标系、时间、动作语义、机位角色、标定和标签可靠性实行严格数据契约。 最可靠的工程结构不是“全部统一设备”,也不是“什么数据都收”,而是“广泛异构池 + 显式条件化/对齐 + 小规模目标域锚点数据”。

研究边界

本文所称“预训练模型”,主要指从多任务、多数据集或多本体轨迹中学习通用视觉—语言—动作能力的 VLA 及相邻操作预训练方法。相机参数严格区分为三层:原始采集规格、预处理/传输规格和模型实际输入规格。论文只公开后两层时,不倒推摄像头原生能力。对论文未报告的采集码率和 FOV,本文明确记为未报告,不用经验数字补空。

核心结论

近一年的证据支持一个“分层一致性”框架:

  1. 覆盖层要多样。 应扩大任务、物体、场景、光照、背景、视点、行为策略、成功/恢复过程、操作者和数据源覆盖。
  2. 语义接口层要一致或可条件化。 必须能解释每条数据的坐标系、本体、控制器、动作定义、物理时间、机位角色、内外参、同步质量和标签信心。
  3. 目标域层要有高保真锚点。 保留一小批与目标机器人观测几何、动作接口和任务分布对齐的数据,用于中间训练、微调和闭环验收。

成果最直接的 ACE-Ego-0 实验将 1.48K 小时人类 egocentric 伪动作数据与 4.53K+ 小时机器人/仿真数据联合预训练。它并未将数据“先混了再说”,而是统一到相机坐标系,用形态 token 表示不同身体,按物理时间而非帧数组成动作块,并对人类伪标签做可靠性加权。去掉形态条件、时间对齐和标签可靠性后,RoboCasa 平均成功率都下降;其中将噪声人类伪动作与传感器记录的机器人动作等权对待,降幅最大。

1. 多样性究竟是收益,还是噪声

1.1 应当扩大的多样性

对通用模型,多样数据的价值在于补齐目标数据中没有的任务、语义和动作区域。ACE-Ego-0 的 Sweep Cubes 实验中,34 条机器人示范只覆盖 0.062 平方米的末端平面工作空间;419 条任务匹配人类视频覆盖 0.296 平方米,为前者的 4.8 倍,联合微调将 10 次试验中的成功率从 10% 提高到 40%。

大规模 EgoScale 实验同样说明规模与环境覆盖有价值:人类动作预训练从 1K 小时增至 20K 小时时,五个真实机器人任务的平均完成度从 0.30 增至 0.71。但最佳结果仍需少量精确对齐的人—机器人中间训练。因此应扩大的是“有新覆盖价值的多样性”,而不是“来源名称的数量”。

跨本体数据也不是越多越好,还要看均衡。OXE-AugE 指出,OXE 原始真实数据中排名前四的机器人类型占比超过 85%,容易让策略过拟合少数机器人—场景组合。所以,多本体采集应同时管理各本体、夹爪和任务的权重,而不是把大数据源自然变成主导分布。

1.2 必须统一、归一或显式标记的差异

动作语义不能靠文件名暗示。SPACE 显示,同一个动作数字可以因本体几何、底层控制器、硬件和执行延迟不同而导致不同结果。数据质量研究则将进度、平滑性、无效停顿、限位接近和操作者反馈作为 episode 层信号,提醒我们不能只用“成功/失败”两个标签衡量轨迹价值。

人类视频的伪动作也不能与传感器记录的动作等权。人类视频过滤工作表明,广泛视频只有在目标机器人可达、物体与抓取兼容、轨迹可执行时才是有效监督。全局选数据方法的反例还显示,只按一个总体分数排序会清空某些任务的覆盖;更合理的做法是在任务、本体或原子行为结构内分配预算后再评分。

2. 相机参数:已发表实例与可泛化要求

维度近一年已报告实例能否当通用门槛对采集的真正要求
分辨率HumanEgo 的 Aria RGB 原始采集为 2MP,策略输入为 240×320;SPARC 真机三视图输入为 224×224;部分 VLA 系统用 256×256 输入不能。模型输入不等于采集上限降采样和裁剪后仍能分辨目标、夹爪/手、接触区域和任务状态;原片保留足够的重裁剪与重标注裕量
帧率/采样率HumanEgo 人类 RGB 30 fps;EgoScale 大池为 30 fps;UMI-3D RGB 20 Hz、LiDAR 10 Hz;EmbodiedComp 真机双视图与动作 10 Hz不能。10、20、30 都是系统实例覆盖任务中最快的可见状态变化;用硬件时间戳对齐视觉、状态、动作与其他模态;训练时按物理时间重采样,不把不同帧数当作相同时长
码率EmbodiedComp 在特定桌面基准中从 0.10 到 0.06 bpp 的性能约降 5%,约 0.04 bpp 转折,约 0.02 bpp 快速失效;SPARC 显示按机位/区域分配比特更有效不能。bpp 还受分辨率、帧率、codec、模型和任务影响保存原始/高质量归档源;对传输版本画“码率—闭环成功率”曲线;优先保留任务关键机位和区域,不只看 PSNR/SSIM
FOVUMI-3D 在腕部使用 185° 鱼眼镜头;OXE-AugE 将不同相机的观测裁剪到接近 Bridge 数据的视场不能。185° 是一种腕部实现,不是通用值同时覆盖操作上下文和手/夹爪—目标交互区;记录内参、镜头模型和去/不去畸变策略;对目标域锚点数据保持观测几何可比
安置位置头部/第一视角、顶置/基座第三视角、腕部/掌内相机均有成功系统;SPARC 用侧、顶、腕三视图没有唯一最佳机位把机位当作带语义的观测角色;目标域数据要包含实际部署机位;多视图要保存同步、外参和 view ID

2.1 分辨率:要看裁剪后的有效像素,不是相机广告参数

在 HumanEgo 中,Project Aria Gen1 的 RGB 流以 30 fps 记录 2MP 图像,但训练的 RGB 输入为 240×320。该对比说明,原始分辨率的主要价值是为裁剪、跟踪、深度/位姿估计和以后重标注留出信息裕量;策略网络并不必然直接消费原始像素。SPARC 的真机侧视、顶视和腕视均缩放至 224×224,这也是模型/编码链的输入,不是论文对摄像头原生分辨率的统一要求。

因此,分辨率验收应变成任务测试:对最小物体、最细接触区、最大运动模糊和最不利裁剪分别评估跟踪/位姿误差与闭环成功率。“原片高于模型输入”是合理的工程结构;“所有数据源原生像素完全一致”不是证据支持的前提。

2.2 帧率:时间戳和物理时长比“全部 30 fps”更重要

人类 egocentric 视频里 30 fps 很常见:HumanEgo 与 EgoScale 都使用了 30 fps 人类 RGB 数据。但具身采集系统的合理频率可以不同。UMI-3D 的腕部 RGB 以 20 Hz 硬件触发,LiDAR 为 10 Hz,通过共同时钟和统一时空标定对齐;EmbodiedComp 的真实 UR5 采集以 10 Hz 记录腕部、第三人称视频、关节角和末端增量动作。

异构数据的问题不是“帧率必须相同”,而是“相同长度的 action chunk 是否代表相同物理时长”。ACE-Ego-0 去掉 time-aligned action chunking 后性能下降,原因正是不同帧率下固定步数覆盖了不同实际时长。工程上应保存原始时间戳和实际采样频率,再在训练时做物理时间对齐。对快速接触、冲击、工具操作或快速运动,还要用预实验检查 30 fps 是否会遗漏关键过渡,不能从慢速桌面任务直接外推。

2.3 码率:没有行业通用最低值,只有任务条件下的失效曲线

本次纳入的多数预训练论文没有报告原始采集视频的 codec、GOP、量化参数或 Mbps,因此现有证据不足以提出一个“预训练必须高于某 Mbps”的标准。少量专门的压缩工作提供了方法论而不是通用数字。

EmbodiedComp 在三个 VLA、桌面原子任务和其选定 codec 上观察到,机器人视觉性能从 0.10 到 0.06 bpp 约只下降 5%,约 0.04 bpp 出现转折,约 0.02 bpp 附近变得不可用。这是有价值的警示,却不是行业阈值:bpp 要与像素数、帧率、codec、视图数和任务一起解释。SPARC 进一步显示,不同机位和画面区域对动作的价值不均匀,把码率平均分给所有视图和区域是低效的。所以,采集与传输应分开:归档版保留高质量原始信息,在线版通过目标 VLA 的闭环成功率曲线选择 codec 和码率。

2.4 FOV 与机位:应对齐“观测几何”,不必对齐“摄像头型号”

FOV 的目标是在解析力、畸变和上下文覆盖之间取得任务相关平衡。UMI-3D 在夹爪腕部使用 185° 鱼眼镜头,以保留近距离交互和任务上下文;它保留原始鱼眼图像,并对相机与 LiDAR 做统一标定。OXE-AugE 则使用不同相机设备,但将真机观测裁剪到接近 Bridge 数据的视场。两者共同指向:硬件可以不同,进入模型的观测几何要么被对齐,要么被充分标记并在训练中条件化。

机位各有分工:顶置/基座第三视角稳定、适合观察全局,但对局部接触不敏感;头部 egocentric 视角能记录人的主动观察和长时序上下文,但存在自运动;腕部/掌内视角接近操作区,却容易被手、夹爪或物体遮挡,且运动模糊更强。主动感知预训练还表明,头戴相机运动可以是策略的“视点动作”,不应一律当成要被消除的噪声。

3. 建议的数据源架构

3.1 广泛池:覆盖优先

广泛池可同时纳入人类 egocentric 视频、机器人遥操、UMI、仿真、策略 rollout 和多机构数据。它的目标是获得广泛物体、场景、行为和视点,不必强求摄像头型号和原生分辨率相同。但每条数据都必须带上来源、本体、任务、机位、视觉参数、频率、时间戳、动作接口、控制器和监督可靠性。

3.2 数据契约:异构性可解释

建议的最小数据契约包括:

  • 相机型号、镜头模型、原始分辨率、名义/实际帧率、曝光、快门、codec 及码率或质量档;
  • 机位角色(head/base/top/side/wrist/palm)、内参、畸变参数、相对机器人基座/末端的外参与标定版本;
  • 硬件时间戳或可校验的共同时钟,丢帧、抖动、触发延迟和跨模态对齐误差;
  • 动作坐标系、绝对/增量语义、单位、控制频率、action chunk 覆盖的物理时长、底层控制器;
  • 本体形态、夹爪/灵巧手、关节限位与可达性;
  • 每个模态和动作通道的真值/伪标签/缺失标记,以及数据集级和时刻级的置信度。

3.3 目标域锚点:一致性优先

锚点集应尽可能匹配最终机器人的部署机位、观测裁剪、标定、动作接口、控制频率和任务难度。它不需要比广泛池大,却要高保真、可重复、有真实闭环结果。EgoScale、HumanEgo、UniDex 和 ACE-Ego-0 的证据都表明,外部多样数据可以带来先验和覆盖,但对准确控制的最后校准仍需可执行的目标域信号。UniDex 的结果尤其说明,经过重定向的人类数据可补充机器人示范,但在接触丰富任务上不能简单替代目标机器人数据。

4. 实际采集建议

以下是基于证据的工程建议,不是论文共同规定的行业标准:

  1. 原片与训练输入分层保存。 在存储允许时,原片保留比 224–320 级模型输入更高的空间信息,统一生成模型所需的派生版本;不应只存最后缩放图像。
  2. 将 30 fps 视为常见起点,而非通用终点。 用最快动作、最短接触和最大运动模糊做预试验;快速或接触任务若在 30 fps 丢失关键过渡,应提高帧率或增加更高频率的本体/力/触觉信号。
  3. 对多视图使用硬件时间戳和标定版本。 对齐误差至少要小于一个最快模态的采样周期;需做 3D 手—物或相机—LiDAR 重建时,毫秒级共同时钟更理想。
  4. 不用一个固定 FOV 覆盖所有机位。 全局机位优先完整任务状态,腕部机位优先夹爪/手与交互区;每个机位都用“关键对象在最不利姿态下不离开画面、仍有足够像素”作为验收。
  5. 码率用闭环任务定标。 分别测试原始、高、中、低码率的成功率、失败类型和延迟,选择失效转折前有安全余量的档位;不用 PSNR、SSIM 或人眼清晰度代替。
  6. 分层采样,不做大池简单随机混合。 先为任务、本体、机位、数据源和失败/恢复类型保留覆盖,再在层内按可执行性和标签可靠性加权。SIEVE 的结构化选择结果也支持先保留原语和组合结构,再谈数据效率。

条件、限制与研究空白

第一,本次样本的主流证据仍集中于桌面操作、夹爪或有限灵巧手。导航、足式运动、高速操作和室外传输未必共享相同分辨率/帧率/码率转折点。第二,大多数预训练论文对曝光、快门、滚动/全局快门、视频 codec、实际码率和 FOV 报告不完整,这是当前文献本身的复现性缺口。第三,多样来源的增益往往与新的动作表示、损失、过滤器和训练策略一起出现,因而很难将收益全部归因于“设备种类增加”。

第四,纯 RGB 对遮挡、弱纹理和接触任务存在可见上限。UMI-3D 用 LiDAR 解决视觉 SLAM 在遮挡、动态场景和跟踪失败中的可靠性问题;触觉视频基准显示大规模视觉—触觉表示可以从多任务数据中获得收益,但表示收益不等于已经解决了所有闭环控制的力觉需求。

结论与下一步

对用于具身预训练的数据来说,“一致”与“多样”并不矛盾,它们应用在不同层次。数据内容要多样,数据契约要统一,目标域锚点要对齐。相机型号可以不同,但机位角色、内外参、时间戳和观测几何不能不明;帧率可以不同,但物理时间语义不能混乱;码率可以因机位和任务自适应,但必须经过目标 VLA 的闭环验收。

如果要启动一个新的采集项目,最优先的投资顺序应是:先定义数据契约与目标域验收集,再用预实验确定各机位的分辨率、帧率、FOV 和码率,最后才扩大设备与来源的数量。这样获得的不是一个表面整齐的视频库,而是一个可以让模型正确理解差异的预训练资产。

References

Accepted evidence

去重论文引用

  1. 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
  2. ActiveMimic: Egocentric Video Pretraining with Active PerceptionXingyao Lin, Guojin Zhong, Tianyi Lu, Ziyi Ye, Yichen Zhu, Zuxuan Wu, Yu-Gang Jiang · 2026
  3. ATHENA: Accelerated Multi-Task Heterogeneous Influence Functions for Robot Data CurationTao Xu, Jiaxin Wang, Runhao Zhang, Jiayi Guan, Xianchao Zeng, Weixi Song, Xinyu Zhou, Zhetao Chen et al. · 2026
  4. Closing the Loop in Teleoperation: Episode-Level Data Quality Assessment and Feedback for High-Quality Demonstration CollectionGokul Narayanan, Yash Shahapurkar, Melih Erdogan, Brian Zhu, Eugen Solowjow · 2026
  5. Data Retrieval with Importance Weights for Few-Shot Imitation LearningAmber Xie, Rahul Chand, Dorsa Sadigh, Joey Hejna · 2025
  6. Developing Vision-Language-Action Model from Egocentric VideosTomoya Yoshida, Shuhei Kurita, Taichi Nishimura, Shinsuke Mori · 2025
  7. Dexterous Manipulation Policies from RGB Human Videos via 3D Hand-Object Trajectory ReconstructionHongyi Chen, Tony Dong, Tiancheng Wu, Liquan Wang, Yash Jangir, Yaru Niu, Yufei Ye, Homanga Bharadhwaj et al. · 2026
  8. EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human DataRuijie Zheng, Dantong Niu, Yuqi Xie, Jing Wang, Mengda Xu, Yunfan Jiang, Fernando Castañeda, Fengyuan Hu et al. · 2026
  9. Embodied Image Compression: Towards Codec for Robotic Visual SystemsZhenghao Chen, Zijie Yue, Haozhe Li, Yifan Zhao, Yunhui Guo, Zhi Wang · 2025
  10. 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
  11. HumanEgo: Zero-Shot Robot Learning from Minutes of Human Egocentric VideosZhi Wang, Botao He, Kelin Yu, Seungjae Lee, Ruohan Gao, Furong Huang, Yiannis Aloimonos · 2026
  12. 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
  13. OmniUMI: Towards Physically Grounded Robot Learning via Human-Aligned Multimodal InteractionShaqi Luo, Yuanyuan Li, Youhao Hu, Chenhao Yu, Chaoran Xu, Jiachen Zhang, Guocai Yao, Tiejun Huang et al. · 2026
  14. OXE-AugE: A Large-Scale Robot Augmentation of OXE for Scaling Cross-Embodiment Policy LearningGuanhua Ji, Harsha Polavaram, Lawrence Yunliang Chen, Sandeep Bajamahal, Zehan Ma, Simeon Adebola, Chenfeng Xu, Ken Goldberg · 2025
  15. SIEVE: Structure-Aware Data Selection for Imitation Learning with VLA ModelsChangti Wu, Bin Yu, Zhaolong Shen, Shijie Lian, Xiaopeng Lin, Cong Huang, Zhirui Zhang, Lei Zhang et al. · 2026
  16. SPACE: Enabling Learning from Cross-Robot Data Toward Generalist PoliciesHaeone Lee · 2026
  17. SPARC: Spatially Adaptive Rate Control for Vision-Language-Action ModelsSangYun Chung, Mincheol Shin, Jihyun Kim, Sangpil Kim · 2026
  18. SUGAR: A Scalable Human-Video-Driven Generalizable Humanoid Loco-Manipulation Learning FrameworkTianshu Wu, Xiangqi Kong, Yue Chen, Qize Yu, Hang Ye, Jia Li, Yizhou Wang, Hao Dong · 2026
  19. UMI-3D: Extending Universal Manipulation Interface from Vision-Limited to 3D Spatial PerceptionZiming Wang · 2026
  20. UniDex: A Robot Foundation Suite for Universal Dexterous Hand Control from Egocentric Human VideosGu Zhang, Qicheng Xu, Haozhe Zhang, Jianhan Ma, Long He, Yiming Bao, Zeyu Ping, Zhecheng Yuan et al. · 2026
完整证据附录(33 条正式证据事件)
  • Time range: 2025-07-26..2026-07-26
  • Events: 33
  • 每个事件一节,标题即锚点;trace-map 中的 event 链接跳转到这里。

EA-DQ-YEAR-READ-0008

  • Claim: 少样本部署场景下,外部大规模数据的“质量”主要取决于对目标任务分布的相关性;只按最近邻相似度检索会受噪声和先验数据分布偏差影响。
  • Stance: support | Confidence: direct
  • Paper: 2509.01657 Data Retrieval with Importance Weights for Few-Shot Imitation Learning
  • Locator: Abstract (full-text section)
  • Evidence: IWR 将 retrieval-based imitation learning 的常用最近邻规则解释为目标数据分布 KDE 的极限,指出其高方差、易受噪声影响且不考虑 prior data distribution;方法用目标/先验分布概率比进行 importance-weighted retrieval,并在仿真和 Bridge 真实评估中改善现有检索方法。
  • Quote: “Abstract While large-scale robot datasets have propelled recent progress in imitation learning, learning from smaller task specific datasets remains critical for deployment in new environments and unseen tasks. One such approach to few-shot imitation learning is retrieval-based imitation learning, which extracts relevant samples from large, widely available prior datasets to augment a limited demonstration dataset. To determine the relevant data from prior datasets, retrieval-based approaches mo”
  • Authors: amber-xie; rahul-chand; dorsa-sadigh; et al.

EA-DQ-YEAR-READ-0009

  • Claim: 跨本体机器人数据的质量问题不只是轨迹数量,而是本体/夹爪分布是否均衡;高度不平衡的数据集会让策略过拟合少数 robot-scene 组合。
  • Stance: support | Confidence: direct
  • Paper: 2512.13100 OXE-AugE: A Large-Scale Robot Augmentation of OXE for Scaling Cross-Embodiment Policy Learning
  • Locator: Abstract (full-text section)
  • Evidence: 论文指出 OXE 聚合 60 多个机器人数据集,但 top four robot types 占超过 85% 真实数据,带来过拟合风险;OXE-AugE 用 9 种不同机器人本体扩增 16 个 OXE 子集,形成 4.4M trajectories,并研究扩增对 cross-embodiment learning 的影响。
  • Quote: “Abstract Large and diverse datasets are needed for training generalist robot policies that have potential to control a variety of robot embodiments—robot arm and gripper combinations—across diverse tasks and environments. As re-collecting demonstrations and retraining for each new hardware platform are prohibitively costly, we show that existing robot data can be augmented for transfer and generalization. The Open X-Embodiment (OXE) dataset, which aggregates demonstrations from over 60 robot dat”
  • Authors: guanhua-ji; harsha-polavaram; lawrence-yunliang-chen; et al.

EA-EGO-2026-0007

  • Claim: 在 EgoScale 的测量区间内,egocentric human action pretraining 确有规模收益:1K 到 20K 小时使真实机器人平均任务完成度从 0.30 升到 0.71。
  • Stance: support | Confidence: direct
  • Paper: 2602.16710 EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data
  • Locator: 3.3 Policy Performance Scales with Pretraining Data Size
  • Evidence: 五个数据规模的同架构实验报告单调提升,并限制结论不外推到测量区间之外。
  • Quote: “Average task completion rises monotonically from 0.30 at 1k hours to 0.71 at 20k hours”
  • Authors: ruijie-zheng; dantong-niu; yuqi-xie; et al.

EA-DQ-YEAR-READ-0010

  • Claim: 在多任务 VLA 微调中,数据质量治理要同时考虑样本效用和任务覆盖;单一全局排序会让某些任务被几乎淘汰,导致任务级覆盖坍缩。
  • Stance: support | Confidence: direct
  • Paper: 2606.16208 ATHENA: Accelerated Multi-Task Heterogeneous Influence Functions for Robot Data Curation
  • Locator: C.4 Retention Balance, Single-Task Curation, and Real-Robot Failure Modes
  • Evidence: ATHENA 指出 VLA 性能不只取决于规模,也取决于 demonstration quality,大规模冗余数据甚至可能伤害性能;在六任务真实机器人设置中,naive global influence ranking 让 Stack Bowls 只保留 13 条示教,而 MII 结合 task-local 和 cross-task influence utilities 后保留分布更均衡。
  • Quote: “To further ablate the role of Multitask Influence Interaction (MII), we visualize the retained task distributions after data curation in Fig. 8 . We consider the six-task real-robot setting with 120 demonstrations per task and an overall retention ratio of 66.7%. Without MII, naively ranking demonstrations with a single global influence score results in a highly skewed retained set: Pick Fruits, Shelf Retrieval, and Wipe Board retain 115, 113, and 104 demonstrations, respectively, whereas Stack”
  • Authors: tao-xu; jiaxin-wang; runhao-zhang; et al.

EA-PRETRAIN-DATA-2026-0003

  • Claim: 多相机 VLA 不应把码率在机位和画面区域间均分;应优先保留对当前动作有用的视图和区域。
  • Stance: support | Confidence: direct
  • Paper: 2606.16253 SPARC: Spatially Adaptive Rate Control for Vision-Language-Action Models
  • Locator: 1 Introduction and 3 Method
  • Evidence: 论文指出不同机位和图像区域对控制的价值不均匀,SPARC 通过时序 mask 自适应分配比特。
  • Quote: “Uniform bitrate allocation across cameras and image regions is therefore fundamentally inefficient.”
  • Authors: sangyun-chung; mincheol-shin; jihyun-kim; et al.

EA-PRETRAIN-DATA-2026-0002

  • Claim: 任务匹配的人类 egocentric 视频能补齐少量机器人示范的动作覆盖空洞,但收益是在对齐与质量加权管线中实现的。
  • Stance: support | Confidence: direct
  • Paper: 2606.17200 ACE-Ego-0: Unifying Egocentric Human and Robotic Data for VLA Pretraining
  • Locator: 5.3 Human Data for Augmented Fine-Tuning, Figure 6
  • Evidence: 419 条人类视频的工作空间覆盖是 34 条机器人示范的 4.8 倍,联合微调将 10 试验成功率从 10% 提高到 40%。
  • Quote: “The 419 episodes of task-matched human video spread across 0.296 m 2 , 4.8 broader coverage”
  • Authors: hao-li; ganlong-zhao; yufei-liu; et al.

EA-PRETRAIN-DATA-2026-0005

  • Claim: VLA 对压缩往往呈‘轻压缩稳定、越过任务特定转折后骤降’,因此码率验收应看闭环成功曲线,不应只看人眼画质。
  • Stance: conditional | Confidence: direct
  • Paper: 2512.11612 Embodied Image Compression: Towards Codec for Robotic Visual Systems
  • Locator: 5.3 Experiment Result and Discussions
  • Evidence: 该基准中 RVS 从 0.10 到 0.06 bpp 约下降 5%,约 0.04 bpp 出现转折,0.02 bpp 附近快速失效。
  • Quote: “RVS curves stay flat from 0.10 to 0.06 bpp (5% drop), then kink sharply around 0.04 bpp”
  • Authors: zhenghao-chen; zijie-yue; haozhe-li; et al.

EA-PRETRAIN-DATA-2026-0006

  • Claim: 当动作学习依赖多视图时,数据包应同步保存机位标识、视频、机器人状态和动作;10 Hz 是该 UR5 系统实例,不是预训练的通用帧率。
  • Stance: conditional | Confidence: direct
  • Paper: 2512.11612 Embodied Image Compression: Towards Codec for Robotic Visual Systems
  • Locator: Appendix C Subjective Data Collection
  • Evidence: 真实管线同步记录腕部与第三人称 RealSense、关节角和末端增量动作,频率为 10 Hz。
  • Quote: “Joint angles, two camera streams (wrist view and third-person view, captured by two Intel realsense cameras), and actions”
  • Authors: zhenghao-chen; zijie-yue; haozhe-li; et al.

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-EGO-2026-0008

  • Claim: 大规模 human pretraining 仍需少量精确 aligned human-robot mid-training 才能最好地落到可执行控制;规模和本体对齐是互补条件。
  • Stance: conditional | Confidence: direct
  • Paper: 2602.16710 EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data
  • Locator: 3.2 Large-Scale Human Pretraining Is Key to Strong Dexterous Manipulation Policy Performance
  • Evidence: 四类 checkpoint 的消融中,pretrain+midtrain 最好;human pretraining 提供结构,mid-training 负责控制锚定。
  • Quote: “combining human pretraining with a small amount of aligned mid-training yields the best overall performance”
  • Authors: ruijie-zheng; dantong-niu; yuqi-xie; et al.

EA-EGO-2026-0011

  • Claim: Retargeted ego-human 数据只能部分替代目标机器人示范:在论文的 Make Coffee co-training 实验中,没有 robot data 时成功始终接近 0。
  • Stance: conditional | Confidence: direct
  • Paper: 2603.22264 UniDex: A Robot Foundation Suite for Universal Dexterous Hand Control from Egocentric Human Videos
  • Locator: 5.4 UniDex-Cap for Human-Robot Data Co-train
  • Evidence: 作者明确总结 human data helps but robot data is indispensable,并给出约 2:1 的局部替代斜率。
  • Quote: “Retargeted human data helps, but robot data is indispensable”
  • Authors: gu-zhang; qicheng-xu; haozhe-zhang; et al.

EA-UMI-READ-0003

  • Claim: UMI-style data is a scalable foundation, but becomes substantially more useful for contact-rich tasks when upgraded from mainly visuomotor supervision to multimodal physical interaction data.
  • Stance: conditional | Confidence: direct
  • Paper: 2604.10647 OmniUMI: Towards Physically Grounded Robot Learning via Human-Aligned Multimodal Interaction
  • Locator: Abstract (full-text section)
  • Evidence: The HTML full text repeatedly identifies limited physical interaction signals as a bottleneck of existing UMI-like systems and proposes synchronized RGB, depth, trajectory, tactile sensing, internal grasping force, and external wrench data to improve contact-rich policy learning.
  • Quote: “Abstract UMI-style interfaces enable scalable robot learning, but existing systems remain largely visuomotor, relying primarily on RGB observations and trajectory while providing only limited access to physical interaction signals. This becomes a fundamental limitation in contact-rich manipulation, where success depends on contact dynamics such as tactile interaction, internal grasping force, and external interaction wrench that are difficult to infer from vision alone. We present OmniUMI, a uni”
  • Authors: shaqi-luo; yuanyuan-li; youhao-hu; et al.

EA-EGO-2026-0013

  • Claim: Physics refiner 和 interaction reward 是把 Ego-centric 视频数据变成可执行技能的必要中间层;只跟踪运动会在接触任务中失败。
  • Stance: conditional | Confidence: direct
  • Paper: 2605.20373 SUGAR: A Scalable Human-Video-Driven Generalizable Humanoid Loco-Manipulation Learning Framework
  • Locator: 4.4 Component Analysis
  • Evidence: 组件消融中去除 Refiner 显著退化,去除 interaction reward 时机器人只模仿弯腰而无法抬起物体。
  • Quote: “Removing the Refiner leads to substantial performance degradation”
  • Authors: tianshu-wu; xiangqi-kong; yue-chen; et al.

EA-EGO-2026-0014

  • Claim: 缩小 human/robot 图像外观差距并不足以让 ego 数据可训练;Water Flowers 消融中 visual-only 最高约 32.5%,显式 hand-object 6DoF ICT 才带来大幅闭环提升。
  • Stance: conditional | Confidence: direct
  • Paper: 2605.24934 HumanEgo: Zero-Shot Robot Learning from Minutes of Human Egocentric Videos
  • Locator: 4.4 What Drives Performance of HumanEgo?
  • Evidence: raw RGB、inpainting、robot RGB 和 ICT 的阶梯消融把视觉外观与空间关系作用分离。
  • Quote: “Monocular RGB encodes appearance , not the 3D spatial relationships that manipulation demands.”
  • Authors: zhi-wang; botao-he; kelin-yu; et al.

EA-EGO-2026-0017

  • Claim: 自动 RGB-only ego 标签存在明显 fidelity ceiling:严格阈值下左右 wrist pose recovery 仅约 66% 和 62%,规模化以噪声为代价。
  • Stance: conditional | Confidence: direct
  • Paper: 2606.06194 ActiveMimic: Egocentric Video Pretraining with Active Perception
  • Locator: 4.3 Egocentric Video Yields Effective Pretraining Labels
  • Evidence: HOT3D ground truth 上的 10% sample 验证给出 head/wrist 三类严格阈值 recovery rate。
  • Quote: “Under the strict tier ( , rot6d L2 ), head recovery reaches 78.82%, with left and right wrist recovery at 65.93% and 61.72%, respectively;”
  • Authors: xingyao-lin; guojin-zhong; tianyi-lu; et al.

EA-EGO-2026-0018

  • Claim: 把 camera motion 当作 viewpoint action 可提供真实的 active-perception prior,但能力必须在有 head-camera/robot fine-tuning 的系统中承接。
  • Stance: conditional | Confidence: direct
  • Paper: 2606.06194 ActiveMimic: Egocentric Video Pretraining with Active Perception
  • Locator: 4.4 The Head Camera Enables Pretrained Active Perception
  • Evidence: Restocking 中 egocentric-pretrained model 的 placement 为 24/27,SFT-only 为 6/27;移除 head camera 降到 1/27。
  • Quote: “ActiveMimic scores 24 out of 27 on placement, whereas ActiveMimic sft-only achieves only 6 out of 27”
  • Authors: xingyao-lin; guojin-zhong; tianyi-lu; et al.

EA-PRETRAIN-DATA-2026-0004

  • Claim: 带宽要求必须在目标 VLA 和任务上用闭环成功率标定,感知画质或单一固定 bpp 不能替代。
  • Stance: conditional | Confidence: direct
  • Paper: 2606.16253 SPARC: Spatially Adaptive Rate Control for Vision-Language-Action Models
  • Locator: C.1 Analysis of Key Components
  • Evidence: VLABench 中 0.0333/0.0685 bpp 的 SPARC 成功率接近但低于未压缩,而极低 bpp 下所有变体失效。
  • Quote: “SPARC (Ours) 0.0333 37.5 0.0685 38.3 Uncompressed - 40.5 - 40.5”
  • Authors: sangyun-chung; mincheol-shin; jihyun-kim; 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-EGO-2026-0003

  • Claim: 从 egocentric 视频恢复的 object trajectory 不是完整机器人动作:该路线无法获得 gripper state,只能把动作表示为 9D 物体位姿增量。
  • Stance: limit | Confidence: direct
  • Paper: 2509.21986 Developing Vision-Language-Action Model from Egocentric Videos
  • Locator: III-C Policy Training
  • Evidence: 策略训练段明确说明 gripper state 缺失,并以 object pose displacement 作为替代动作。
  • Quote: “Because gripper states cannot be obtained from Section III-B , each action is represented by a 9-dimensional vector”
  • Authors: tomoya-yoshida; shuhei-kurita; taichi-nishimura; et al.

EA-EGO-2026-0004

  • Claim: Ego-centric 轨迹构建存在规模—质量冲突:保留更多疑似背景/错误轨迹会显著降低真实机器人表现。
  • Stance: limit | Confidence: direct
  • Paper: 2509.21986 Developing Vision-Language-Action Model from Egocentric Videos
  • Locator: IV-C Ablation Study
  • Evidence: BGTS=1.0 保留 86,427 episodes 但真实机器人分数低于 BGTS=0.7 的 45,157 episodes。
  • Quote: “Setting an appropriate curation threshold is crucial to balancing the scale and quality of our dataset”
  • Authors: tomoya-yoshida; shuhei-kurita; taichi-nishimura; et al.

EA-EGO-2026-0005

  • Claim: 单目 RGB 人类视频恢复出的 hand-object 轨迹常不具物理可执行性;对象几何、手尺度/姿态误差会形成穿模、无效接触和抓取失败。
  • Stance: limit | Confidence: direct
  • Paper: 2602.09013 Dexterous Manipulation Policies from RGB Human Videos via 3D Hand-Object Trajectory Reconstruction
  • Locator: III-B Dexterous Grasp and Manipulation Learning
  • Evidence: 方法段明确说明重建运动正确时,机器人—对象交互仍可能因几何误差而无效。
  • Quote: “the resulting robot–object interactions are not always physically feasible due to reconstruction errors”
  • Authors: hongyi-chen; tony-dong; tiancheng-wu; et al.

EA-EGO-2026-0006

  • Claim: 当前 VideoManip 依赖静态或近静态相机,并在真实闭环中用固定 hand-object 相对位姿绕过手部遮挡,限制了动态第一视角数据的可用范围。
  • Stance: limit | Confidence: direct
  • Paper: 2602.09013 Dexterous Manipulation Policies from RGB Human Videos via 3D Hand-Object Trajectory Reconstruction
  • Locator: V Conclusion, Limitations, and Future Work
  • Evidence: 作者在限制段明确列出 dynamic camera 未覆盖;实验段说明对象点云被 LEAP Hand 遮挡时采用固定相对位姿近似。
  • Quote: “The current framework assumes static or approximately static camera setups”
  • Authors: hongyi-chen; tony-dong; tiancheng-wu; et al.

EA-EGO-2026-0009

  • Claim: Ego-centric 数据的动作接口会决定训练成败:wrist-only 缺失手指/接触时序,小 fingertip 误差又会映射成不合理关节和接触丢失。
  • Stance: limit | Confidence: direct
  • Paper: 2602.16710 EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data
  • Locator: 3.6 Hand Action Space Design for Human Pretraining
  • Evidence: 动作空间消融中 wrist-only 普遍较差,fingertip mapping 在 Cards/Bottle 等接触敏感任务不稳定。
  • Quote: “Small errors in fingertip pose often lead to implausible joint configurations after mapping”
  • Authors: ruijie-zheng; dantong-niu; yuqi-xie; et al.

EA-EGO-2026-0010

  • Claim: 将 egocentric hand trajectories 转为机器人可执行数据仍需 human-in-the-loop retargeting:基础坐标/形态偏差和 contact-rich 片段要人工校准。
  • Stance: limit | Confidence: direct
  • Paper: 2603.22264 UniDex: A Robot Foundation Suite for Universal Dexterous Hand Control from Egocentric Human Videos
  • Locator: 3.2.1 Kinematic Retargeting
  • Evidence: 论文的两阶段 retargeting 先自动 IK,再用 GUI 调整 6DoF offset;接触片段需人工复核。
  • Quote: “The whole pipeline is a two-stage, human-in-the-loop retargeting procedure”
  • Authors: gu-zhang; qicheng-xu; haozhe-zhang; 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-EGO-2026-0012

  • Claim: 从人类视频恢复的 motion prior 会因遮挡、接触伪影和 retargeting 误差而物理不合理,不能直接当作 humanoid policy 的示范。
  • Stance: limit | Confidence: direct
  • Paper: 2605.20373 SUGAR: A Scalable Human-Video-Driven Generalizable Humanoid Loco-Manipulation Learning Framework
  • Locator: 1 Introduction
  • Evidence: 引言直接列出三类误差并说明它们使数据 unsuitable for direct policy learning。
  • Quote: “Severe occlusion, contact artifacts, and retargeting errors render this data physically implausible for direct imitation”
  • Authors: tianshu-wu; xiangqi-kong; yue-chen; et al.

EA-EGO-2026-0015

  • Claim: HumanEgo 的高成功率依赖强 hand/object tracking 前端;单目绝对深度、动态遮挡、模块级联误差和亚厘米接触精度仍是未解决困难。
  • Stance: limit | Confidence: direct
  • Paper: 2605.24934 HumanEgo: Zero-Shot Robot Learning from Minutes of Human Egocentric Videos
  • Locator: 5 Conclusion
  • Evidence: 作者在 limitation 段逐项列出 stereo hand tracking、occlusion-robust tracking、cascading failures 和 1 cm plateau。
  • Quote: “monocular substitutes drop real-world success sharply”
  • Authors: zhi-wang; botao-he; kelin-yu; et al.

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-DQ-YEAR-READ-0015

  • Claim: SIEVE 按可复用原语组合和转换接口分配选择预算,再优先保留各组合模式中稳定、中心的轨迹;论文报告其用 50% 示教和 50% 训练步数可优于全量训练。
  • Stance: limit | Confidence: direct
  • Paper: 2607.06442 SIEVE: Structure-Aware Data Selection for Imitation Learning with VLA Models
  • Locator: Introduction
  • Evidence: 引言的贡献列表同时说明了结构暴露、学习友好轨迹选择和半量数据超过全量训练的结果。
  • Quote: “Our contributions are as follows: • We propose a primitive-compositional view of trajectory utility, realized by Primitive Discovery and Structural Exposure Allocation, which allocate selection budgets according to reuse-aware primitive and transition exposure under diminishing returns. • We introduce Learning-Friendly Trajectory Selection, which selects medoid trajectories within each composition-pattern bucket to favor central, stable, and predictable realizations for behavior cloning. • We pr”
  • Authors: changti-wu; bin-yu; zhaolong-shen; et al.

EA-ALIGN-READ-0012

  • Claim: DQAF 不用成功/失败二值标签代替数据质量,而是联合子任务进度、动作平滑度、停顿和运动学极限生成 episode 级评估与给操作者的纠正反馈。
  • Stance: support | Confidence: direct
  • Paper: 2605.26349 Closing the Loop in Teleoperation: Episode-Level Data Quality Assessment and Feedback for High-Quality Demonstration Collection
  • Locator: Abstract (full-text section)
  • Evidence: 摘要明确列出了质量信号、结构化评估和可执行的自然语言反馈。
  • Quote: “Abstract Industrial automation is at a pivotal moment, as Physical AI is driving a transition from rigid, hand-engineered automation systems toward more flexible and adaptive systems. This shift has created a growing demand for large-scale, real-world robot demonstration data, making teleoperation an increasingly important mechanism for data collection. However, high-quality teleoperated demonstrations remain difficult to obtain in practice, as novice operators often produce episodes that are ta”
  • Authors: gokul-narayanan; yash-shahapurkar; melih-erdogan; 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.

EA-TACTILE-2026-0002

  • Claim: HT-Bench 的进步仍停留在表征层:当前四项任务没有直接测量真实机器人闭环操作,因此不能据此宣称策略或部署收益。
  • Stance: limit | Confidence: direct
  • Paper: 2606.19161 HT-Bench: Benchmarking and Learning Dexterous Full-Hand Tactile Representations with Egocentric Vision
  • Locator: 6 Limitations and Future Work
  • Evidence: 作者在限制章节明确列出硬件/本体覆盖和闭环下游评测缺失。
  • Quote: “While these tasks assess structural, cross-modal, and temporal understanding, they do not directly measure downstream robotic performance.”
  • 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。

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。