数据工程与质量 · Data Engineering and Quality

具身数据感知误差溯源

Tracing Perception Errors in Embodied Data

更新于

30 秒结论

先读这个判断

版本说明:本轮以 15 篇可获取完整正文的论文为论证主干,逐篇核对问题、方法、结果与限制;未能取得可读全文的论文不再承担正文结论。

精读论文
15
正式证据事件
15
知识单元
EA-DATA, EA-SENSOR, EA-EVAL, ERR-PATTERN, ERR-EMBODIED
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

版本说明:本轮以 15 篇可获取完整正文的论文为论证主干,逐篇核对问题、方法、结果与限制;未能取得可读全文的论文不再承担正文结论。

  • 时间范围:2025 年 7 月 14 日至 2026 年 7 月 14 日。
  • 覆盖规模:936 条去重候选、130 篇可用全文、32 篇直接相关论文;同时覆盖观测、接触、同步标定、动作接口、评测和限制证据。
  • 停止依据:候选量、全文量、关键维度和连续两轮低新增率均达到范围综述要求;核心文章只引用代表性论文,不等于检索池只有十余篇。
  • 重要限制:这是公开论文快照,不是仅限同行评审论文的完整普查。

中心判断

近一年论文显示,具身系统里的“感知误差”不应只归因于视觉模型看错了。更有解释力的框架是:从数据采集、传感器可观测性、同步标定、监督字段、动作语义、本体控制到闭环评测的误差账本。许多部署失败在表面上像感知失败,实质上是上游数据缺失、模态错配、动作不可执行或接触隐变量不可观测在闭环中被放大。Lift3D-VLA、TACO、τ0-WM、DQAF 与 GigaWorld-1 分别从三维状态、接触纠错、异构监督、采集质量和闭环评估支撑这条链路(Lift3D-VLA: Lifting VLA Models to 3D Geometry and Dynamics-Aware Manipulation, TACO: TActile World Model as a Self-COrrector forScalable VLA Post-Training, τ_0-WM, Closing the Loop in Teleoperation: Episode-Level Data Quality Assessment and Feedback for High-Quality Demonstration Collection, GigaWorld-1: A Roadmap to Build World Models for Robot Policy Evaluation)。

核心机制

  1. 可观测性缺口是感知误差的第一来源。 2D VLA 会丢失 3D 几何、遮挡、接触和动态一致性,接触任务中的失败还常常视觉不可见;所以溯源时必须问“关键状态有没有被传感器看到”,而不是只问“模型有没有学会”。Lift3D-VLA、TACO、TacForeSight 与 HapTile 分别给出三维、触觉、力觉和多模态采集证据(Lift3D-VLA: Lifting VLA Models to 3D Geometry and Dynamics-Aware Manipulation, TACO: TActile World Model as a Self-COrrector forScalable VLA Post-Training, TacForeSight: Force-Guided Tactile World Model for Contact-Rich Manipulation, HapTile: A Haptic-Informed Vision-Tactile-Language-Action Dataset for Contact-Rich Imitation Learning)。
  2. episode 级遥测是把“感知失败”拆回数据原因的最小单元。 DQAF 表明,成功/失败标签不够,任务进度、运动平滑性、停顿、关节极限能解释一条示教为什么对学习有害;PSD metric 和 QoQ 则分别从运动频谱和下游贡献角度提供质量信号(Closing the Loop in Teleoperation: Episode-Level Data Quality Assessment and Feedback for High-Quality Demonstration Collection, An Efficient Metric for Data Quality Measurement in Imitation Learning, Quality over Quantity: Demonstration Curation via Influence Functions for Data-Centric Robot Learning)。
  3. 异构数据必须带监督可靠性标签。 人类视频、UMI、真实机器人和失败 推演 对 perception/action 的监督强度不同。tau0-WM 用 modality-specific supervision masks 处理缺失模态,PSI 用仿真过滤人类视频里的 pose estimation error、不可达轨迹和抓取不兼容问题,3PoinTr 用 visibility mask 保留被遮挡点的时序监督。

结论是:溯源账本要记录“这条数据能监督什么,不能监督什么”(τ_0-WM, Imitating What Works: Simulation-Filtered Modular Policy Learning from Human Videos, 3PoinTr: 3D Point Tracks for Learning Manipulation from Unconstrained Human Videos)。

  1. 动作语义错配会伪装成感知误差。 SPACE 指出不同机器人为实现同一状态变化可能需要不同控制命令;HapTile 强调动作—状态一致性、时间戳缺口和 episode 级切分;这说明误差归因不能只看图像,还要看坐标系、控制频率、控制器约定、动作归一化和目标本体适配是否一致(SPACE: Enabling Learning from Cross-Robot Data Toward Generalist Policies, HapTile: A Haptic-Informed Vision-Tactile-Language-Action Dataset for Contact-Rich Imitation Learning)。
  2. 溯源粒度要下探到 primitive、chunk 和关键事件。 SIEVE 表明数据选择应保留可复用 primitive 与 transition,不是盲目保留全量;WARP-RM 表明次优长程示教里可能有高价值 recovery chunks。也就是说,失败不是简单归因到“这条轨迹坏了”,而要追踪到任务阶段、接触事件、恢复片段和动作 chunk(SIEVE: Structure-Aware Data Selection for Imitation Learning with VLA Models, WARP-RM: A Warp-Augmented Relative Progress Reward Model for Data Curation)。
  3. 闭环评估决定归因是否成立。 GigaWorld-1 认为世界模型作为策略评估器的可靠性取决于长程 action-faithful 推演 consistency,而非短期视觉真实感;TACO 也把真实失败邻近状态转成可执行触觉纠正数据。由此可得:只有当数据诊断信号能解释或改善闭环成功率、恢复率、过力/碰撞/接管等指标时,它才算有效溯源(GigaWorld-1: A Roadmap to Build World Models for Robot Policy Evaluation, TACO: TActile World Model as a Self-COrrector forScalable VLA Post-Training)。

误差账本模板

层级常见误差来源溯源证据典型检核
观测/传感器2D 丢几何、遮挡、接触不可见、触觉漂移关键点可见性、3D/触觉/力事件、contact gate关键对象可见率、接触检测延迟、触觉/力异常
同步/标定/schema时间戳缺口、action-state 不一致、episode split 泄漏多模态 timestamp、控制循环、raw/rectified tactile同步误差、丢帧率、跨模态对齐残差
episode 质量停顿、振荡、过度纠正、关节极限DQAF 信号、PSD、质量分、人工反馈平滑度、停顿次数、接近关节极限的频率
监督可靠性缺失模态被当真值、人类视频不可执行supervision mask、sim filtering、visibility mask监督字段白名单、不可达率、pose estimation error
动作/本体坐标系、控制频率、adapter、动力学错配state delta、action adapter、controller metadataaction-state consistency、retargeting error
结构/阶段冗余、覆盖不均、坏轨迹中有好片段primitive/transition、progress/chunk labelsprimitive 覆盖、recovery chunk 成功率
闭环结果静态指标与真实部署不一致推演 consistency、真实执行对照成功率、恢复率、过力、碰撞、接管、sim-real ranking

对后续研究的启发

这个话题可以命名为:具身数据感知误差溯源:面向多模态采集与闭环部署的误差账本方法。这是基于 Lift3D-VLA、TACO 和 GigaWorld-1 的综合建议,而不是任一论文的原始命名(Lift3D-VLA: Lifting VLA Models to 3D Geometry and Dynamics-Aware Manipulation, TACO: TActile World Model as a Self-COrrector forScalable VLA Post-Training, GigaWorld-1: A Roadmap to Build World Models for Robot Policy Evaluation)。

核心研究问题建议拆成三问:

  1. 感知失败来自“看不见”、 “对不齐”、 “标错了”、 “动作不可执行”,还是“评估不闭环”?
  2. 每类错误在 episode、chunk、primitive、contact event 哪个粒度上最可诊断?
  3. 哪些数据质量信号能预测或改善真实闭环指标,而不是只提升离线指标?

条件、限制与未解决问题

  • 本轮证据尚未覆盖统一的 perception-error provenance 基准;这是本轮检索的覆盖缺口,不能表述成整个领域已证明不存在此类标准。
  • 多数方法能解释“哪类数据更好”,但本轮较少找到事故级追责链条:从某次失败反推具体数据源、模态、标定、动作字段和训练样本影响。这是本轮证据覆盖的边界,不是对整个领域的否定。
  • 触觉/力/3D 数据正在变重要,但本轮对跨传感器实例、磨损、漂移和维护成本的长期证据不足。这也是本轮检索与证据集的边界。
  • 世界模型可用于离线诊断和候选动作筛选,但 GigaWorld-1 和 TACO 都提醒:看起来真实的 推演 不等于接触和动作忠实,最终仍要回到真实闭环或高可信 sim-real ranking(GigaWorld-1: A Roadmap to Build World Models for Robot Policy Evaluation, TACO: TActile World Model as a Self-COrrector forScalable VLA Post-Training)。

中心判断

感知误差溯源的核心不是为失败贴上“视觉、触觉或模型”的标签,而是找到因果链上的第一偏离点。一次抓取失败可能始于遮挡后的位姿漂移,也可能始于时间戳错位、动作坐标系误解、接触状态不可见,甚至是世界模型把视觉真实感误当成动作可靠性。若只看最终结果,这些机制会被压成同一种失败。

可操作的误差账本至少要同时保存原始观测、校准版本、状态估计、动作命令、控制器解释、接触与力信号,以及每一步的置信度。复盘时先问输入状态是否完整,再问状态是否被正确对齐,最后才检查策略如何使用状态。这样的顺序能减少“换更强视觉模型”却修不到根因的无效迭代。

溯源还必须允许多因一果。视觉遮挡可能使位姿略偏,动作语义错配再把偏差放大,接触不可见最终导致滑移。第一偏离点用于确定最早修复位置,后续放大环节则用于评估防线是否有效;两者不能互相替代。

为了让账本真正可用,每个失败都应保存可重放的时间窗口,而不是只截取最后一帧。窗口至少覆盖动作提出、控制器解释、首次状态偏离、恢复尝试和最终结果。这样才能判断误差是突然发生,还是由小偏差逐步积累;也能检验某个修复究竟消除了根因,还是只掩盖了末端症状。

溯源指标必须与干预相连。若系统认为问题来自标定,就应在重新标定后看到同类失败显著下降;若归因于动作语义,就应通过坐标统一或适配器修复验证。不能被干预复现或证伪的标签,只是事后叙述,还不是工程诊断。

最终可以形成三层防线:采集时拒绝不可解释的数据,训练时隔离可靠性不同的监督,部署时用不确定性和过程信号及时接管。误差账本的价值不是增加记录字段,而是让三层防线共享同一套因果语言。

References

Accepted evidence

去重论文引用

  1. $τ_0$-WM: A Unified Video-Action World Model for Robotic ManipulationPengfei Zhou, Shengcong Chen, Di Chen, Jiaxu Wang, Rongjun Jin, Bingwen Zhu, Yike Pan, Songen Gu et al. · 2026
  2. 3PoinTr: 3D Point Tracks for Learning Manipulation from Unconstrained Human VideosAdam Hung, Bardienus Pieter Duisterhof, Jeffrey Ichnowski · 2026
  3. An Efficient Metric for Data Quality Measurement in Imitation LearningNoushad Sojib, Momotaz Begum · 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. GigaWorld-1: A Roadmap to Build World Models for Robot Policy EvaluationGigaWorld Team, Angyuan Ma, Boyuan Wang, Bohan Li, Chaojun Ni, Guo Li, Guan Huang, Guosheng Zhao et al. · 2026
  6. 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
  7. 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
  8. Lift3D-VLA: Lifting VLA Models to 3D Geometry and Dynamics-Aware ManipulationJiaming Liu, Qingpo Wuwu, Nuowei Han, Hao Chen, Zhuoyang Liu, Fan Fei, Yueru Jia, Chenyang Gu et al. · 2026
  9. PATCH: Action-Chunk-Conditioned Latent Patch Innovation Monitoring for Robot ManipulationYanan Zhou, Ranpeng Qiu, Yincong Chen, Jiajie Cui, Weiming Zhi · 2026
  10. Quality over Quantity: Demonstration Curation via Influence Functions for Data-Centric Robot LearningHaeone Lee, Taywon Min, Junsu Kim, Sinjae Kang, Fangchen Liu, Lerrel Pinto, Kimin Lee · 2026
  11. 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
  12. SPACE: Enabling Learning from Cross-Robot Data Toward Generalist PoliciesHaeone Lee · 2026
  13. 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
  14. TACO: TActile World Model as a Self-COrrector forScalable VLA Post-TrainingShengbang Liu, Yueru Jia, Yuyang Yan, Jiaming Liu, Xinran Zhang, Qiuxuan Feng, Yandong Guo, Shiji Zhou et al. · 2026
  15. WARP-RM: A Warp-Augmented Relative Progress Reward Model for Data CurationJustin Yu, Andrew Goldberg, Kavish Kondap, Karim El-Refai, Ethan Ransing, Qianzhong Chen, Mac Schwager, Fred Shentu et al. · 2026
完整证据附录(15 条正式证据事件)
  • Time range: 2025-07-14..2026-07-14
  • Events: 15
  • 每个事件一节,标题即锚点;trace-map 中的 event 链接跳转到这里。

ERR-TRACE-READ-0010

  • Claim: 示教数据质量不应只用相似度、互信息或人工启发式代理指标定义;QoQ 将高质量轨迹定义为对目标验证示范 loss 降低和策略性能提升有直接贡献的数据。
  • Stance: support | Confidence: direct
  • Paper: 2603.09056 Quality over Quantity: Demonstration Curation via Influence Functions for Data-Centric Robot Learning
  • Locator: VI CONCLUSIONS
  • Evidence: 论文指出人类遥操作会带来错误、操作约束、技能差异、噪声和次优行为;QoQ 用 influence functions 衡量训练 state-action 对验证示范的贡献,并在轨迹层聚合以降低噪声、保持覆盖,在仿真、真实机器人和 DROID in-the-wild 数据上改善策略成功率。
  • Quote: “In this work, we propose QoQ, a method that curates robotic datasets based on direct performance contribution to the learned policy. We define a quality scoring mechanism for state-action pairs derived from influence functions.”
  • Authors: haeone-lee; taywon-min; junsu-kim; et al.

ERR-TRACE-READ-0011

  • Claim: 面向部署环境的 end-user 示教,低质量常表现为过度纠正、振荡和突兀调整;轨迹功率谱密度 PSD 可作为无需 rollout、专家标签或重新训练的快速质量排序指标。
  • Stance: support | Confidence: direct
  • Paper: 2605.01544 An Efficient Metric for Data Quality Measurement in Imitation Learning
  • Locator: Abstract (full-text section)
  • Evidence: 论文把 poor-quality end-user demonstrations 具体化为 excessive corrective motions、oscillations 和 abrupt adjustments,并提出基于 demonstration trajectories PSD 的自动排序指标;实验比较未筛选、oracle、现有排序和 jerk/path-length 等 baseline,研究 PSD 筛选对下游 IL 成功率和平滑性的影响。
  • Quote: “Abstract Imitation learning (IL) has seen remarkable progress, yet field deployment of IL-powered robots remains hindered by the challenge of out-of-distribution (OOD) scenarios. Fine-tuning pre-trained policies with end-user demonstrations collected in deployment environments is a promising strategy to address this challenge. However, end-user demonstrations are frequently of poor quality, characterized by excessive corrective motions, oscillations, and abrupt adjustments that degrade both lear”
  • Authors: noushad-sojib; momotaz-begum

ERR-TRACE-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.

ERR-TRACE-READ-0013

  • Claim: 数据质量必须经闭环评估定义;对机器人世界模型来说,短期视觉真实感不如长程、动作忠实的 rollout 一致性更能决定其作为策略评估器的可靠性。
  • Stance: support | Confidence: direct
  • Paper: 2607.02642 GigaWorld-1: A Roadmap to Build World Models for Robot Policy Evaluation
  • Locator: Abstract (full-text section)
  • Evidence: 论文指出真实机器人策略评估受硬件和人工监督限制,是基础模型迭代瓶颈;WMBench 用真实 teleoperation 数据和匹配 policy rollouts 构造评估,并分析 7 个视频世界模型、4 种动作表示和 324,000 余次模拟 rollout。其结论强调 evaluator 质量由长程 action-faithful rollout consistency、可迁移物理先验、动作编码、记忆和评估导向 post-training 共同决定。
  • Quote: “Using WMBench, we analyze 7 video world models, 4 action representation schemes, and over 324,000 simulated policy rollouts paired with real robot executions, further enriching our analysis with large-scale community submissions from the CVPR 2026 GigaBrain Challenge, curated synthetic trajectories, and a training videos spanning more than 12,000 hours. Our experiments deliver three core insights: evaluator quality is dominated by long-horizon, action-faithful rollout consistency rather than sho”
  • Authors: gigaworld-team; angyuan-ma; boyuan-wang; et al.

ERR-TRACE-READ-0014

  • Claim: Lift3D-VLA 指出,纯 2D VLA 难以保真地表达可达性、遮挡、接触和随时间演化的几何约束,而现有 2D‑3D 转换又会损失几何保真度。
  • Stance: support | Confidence: direct
  • Paper: 2607.06564 Lift3D-VLA: Lifting VLA Models to 3D Geometry and Dynamics-Aware Manipulation
  • Locator: I Introduction
  • Evidence: 引言将操作需求归结为显式 3D 结构与时间一致性,并说明纯 2D 管线及有损的跨模态变换会削弱这些约束。
  • Quote: “Despite this progress, robotic manipulation fundamentally requires spatial reasoning in the physical world [ 57 , 82 , 12 , 13 , 59 ] : the robot must infer 3D structure, reason about geometric relationships (e.g., reachability, occlusion, and contact), and plan actions that remain temporally consistent as the geometry evolves. Purely 2D VLA pipelines often struggle to reliably capture these geometric constraints, particularly in cluttered or dynamic environments. A natural direction is to expli”
  • Authors: jiaming-liu; qingpo-wuwu; nuowei-han; et al.

ERR-TRACE-READ-0005

  • 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.

ERR-TRACE-READ-0007

  • 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

ERR-TRACE-READ-0008

  • Claim: τ0-WM 把真实机器人遥操作、UMI 式交互、人类第一视角视频与 rollout/失败轨迹合并训练,并用按模态的监督掩码处理各数据源的标注缺失。
  • Stance: conditional | Confidence: direct
  • Paper: 2606.01027 $τ_0$-WM: A Unified Video-Action World Model for Robotic Manipulation
  • Locator: Abstract (full-text section)
  • Evidence: 摘要直接列出四类交互数据和 modality-specific supervision masks。
  • Quote: “Abstract Robotic manipulation requires models that generate executable actions while anticipating and evaluating their future consequences before physical execution. We present -World Model ( -WM), a unified video-action world model that integrates policy learning, video prediction, and action evaluation within a single future-predictive framework. Built on a shared video diffusion backbone, -WM provides two complementary interfaces. First, a video action model jointly predicts future visual lat”
  • Authors: pengfei-zhou; shengcong-chen; di-chen; et al.

ERR-TRACE-READ-0002

  • 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.

ERR-TRACE-READ-0003

  • 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.

ERR-TRACE-READ-0006

  • Claim: 混合质量的长程示教不应粗粒度整条丢弃;次优 episode 里可能含有高价值恢复片段,质量控制需要下探到 frame/chunk 级别的进展信号。
  • Stance: conditional | Confidence: direct
  • Paper: 2606.28320 WARP-RM: A Warp-Augmented Relative Progress Reward Model for Data Curation
  • Locator: Abstract (full-text section)
  • Evidence: 论文指出长程遥操作包含 pauses、fumbles 和 recoveries,整条 episode 过滤会丢失 otherwise suboptimal executions 中嵌入的 high-advantage segments,也无法剪掉保留示教中的局部 hesitation;WARP-RM 学习 dense relative progress 并用 WARP-BC upweight high-advantage action chunks。
  • Quote: “Abstract Scaling imitation learning requires large datasets, yet human teleoperation inevitably produces mixed-quality demonstrations containing hesitations and recoveries. Prior frame-level progress reward models supervise on absolute temporal progress proxies that suffer from label noise, or require costly human annotations to define subtask boundaries. We present WARP (Warp-Augmented Relative Progress), a novel fully self-supervised algorithm for learning dense, signed relative progress magni”
  • Authors: justin-yu; andrew-goldberg; kavish-kondap; et al.

ERR-TRACE-READ-0015

  • Claim: 全局视觉异常、帧级变化或策略不确定性不足以判断感知误差是否会影响当前动作;部署监控需要把异常限定到 action chunk 将要使用的执行走廊。
  • Stance: limit | Confidence: direct
  • Paper: 2606.16690 PATCH: Action-Chunk-Conditioned Latent Patch Innovation Monitoring for Robot Manipulation
  • Locator: Abstract (full-text section)
  • Evidence: 作者指出开放工作空间中移动物体、瞬时遮挡和目标运动附近扰动会让部署脆弱;现有 runtime monitors 往往依赖全局 observation anomalies、policy uncertainty 或 frame-level visual changes,难以区分任务相关执行风险和无害视觉变化。PATCH 通过 active action chunk 的 projected execution corridor 累计持续残差作为介入信号。
  • Quote: “Abstract Learning-based manipulation policies have made substantial progress in real-world robot manipulation, particularly for short-horizon action generation. However, deployment in open workspaces remains fragile under unexpected local scene dynamics, such as moving objects, transient occlusions, or disturbances near the intended motion. Existing runtime monitors often rely on global observation anomalies, policy uncertainty, or frame-level visual changes, and struggle to distinguish task-rel”
  • Authors: yanan-zhou; ranpeng-qiu; yincong-chen; et al.

ERR-TRACE-READ-0009

  • 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

ERR-TRACE-READ-0001

  • Claim: TACO 用触觉感知世界模型联合预测未来视频和力序列,再将真实失败附近状态转成带纠正动作的想象片段,用于接触丰富任务的 VLA 后训练。
  • Stance: limit | Confidence: direct
  • Paper: 2607.02840 TACO: TActile World Model as a Self-COrrector forScalable VLA Post-Training
  • Locator: 5 Conclusion and Limitations
  • Evidence: 结论的 Recognize–Imagine–Label 回路明确连接了真实失败、视频—力联合想象与纠正动作标注。
  • Quote: “We presented TACO, a tactile-aware world-model-driven framework for scalable VLA post-training in contact-rich manipulation. Following a Recognize–Imagine–Label loop, TACO converts real-world failures into imagined corrections without repeated human intervention: a tactile-aware world model jointly denoises future video and force sequences, while a unified progress-action model recognizes failure-adjacent states and labels imagined segments with corrective actions. To incorporate this supervisio”
  • Authors: shengbang-liu; yueru-jia; yuyang-yan; et al.

ERR-TRACE-READ-0004

  • 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.

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。