数据工程与质量 · Data Engineering and Quality

近一年触觉数据与视觉数据联合训练的方法和进展

Visual–Tactile Joint Training

更新于

30 秒结论

先读这个判断

本文关注 2025 年 7 月 22 日至 2026 年 7 月 22 日间,触觉数据与视觉数据被共同用于机器人表征、策略、VLA 或世界模型训练的方法。证据池来自 1,789 篇候选论文、247 篇可读全文,最终纳入 27 篇完成全文精读与主张核验的论文。这里的“联合训练”不只指把两种输入送入同一个网络,也包括未来触觉监督、视觉—触觉动作世界模型、训练期触觉蒸馏,以及视觉慢环与触觉快环的分层控制。

精读论文
27
正式证据事件
30
知识单元
EA-SENSOR, EA-ALIGN, EA-MODEL, EA-DATA, 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

研究边界

本文关注 2025 年 7 月 22 日至 2026 年 7 月 22 日间,触觉数据与视觉数据被共同用于机器人表征、策略、VLA 或世界模型训练的方法。证据池来自 1,789 篇候选论文、247 篇可读全文,最终纳入 27 篇完成全文精读与主张核验的论文。这里的“联合训练”不只指把两种输入送入同一个网络,也包括未来触觉监督、视觉—触觉动作世界模型、训练期触觉蒸馏,以及视觉慢环与触觉快环的分层控制。

这批论文大多仍是 arXiv 预印本,任务、触觉硬件和机器人平台分散。因而本文能判断方法方向和已出现的闭环证据,不能据此宣布跨传感器、跨本体的通用视觉—触觉基础模型已经成熟。

中心判断

近一年的核心进展不是“触觉成为第二种视觉”,而是视觉和触觉开始按物理功能、因果角色与控制频率分工:视觉提供接触前的全局语义和几何,触觉描述动作造成的局部接触后果;训练时,未来触觉被用来约束动作相关表示或预测接触演化;执行时,视觉负责较慢的任务与名义动作,触觉负责较快的滑移、压力和失配纠正。真正有效的系统因此越来越像一条“同步接触数据—对齐表示—动作条件预测—门控融合—高频纠偏”的接触执行栈,而不是一个把所有 token 等权混合的多模态编码器。

这一判断可被证伪:如果后续跨传感器、跨任务实验证明,简单早期拼接在同等数据、算力和控制延迟下稳定优于表示对齐、门控或快慢分层,而且不产生视觉动力学污染与缺失模态退化,那么“功能—频率分层”就不再是必要结构。当前证据尚未出现这种反例。

一、数据单元从“配对帧”变成“接触事件”

视觉—触觉训练首先遇到的不是网络问题,而是数据单位问题。视觉帧通常连续且信息密集,触觉却只在 approach、contact、slip、release、recovery 等短暂阶段显著变化。如果按固定帧数均匀抽样,模型看到的大量样本可能只是无接触背景或同一次按压的近重复观测。

规模化正在发生。HT-Bench 把全手触觉与第一视角视频配对扩展到约 1,000 万 RGB 帧、780 万触觉帧和 226 项任务;OmniVTA/OmniViTac 则报告 21,879 条视觉—触觉—动作轨迹、86 个任务、126 个物体,并使用机器人与手持采集器的同构夹爪降低本体差异。进步不只是“更多帧”,而是同步动作、触觉和视觉,保留连续接触结构,并开始覆盖多类接触模式。

但规模很容易被近重复放大。RCT 表明,同一次机器人按压的相邻触觉帧不是独立样本;按帧随机切分会把同一物理交互泄漏到训练和测试。更有启发性的是,在相同的 held-out-contact-sequence 测试池上,从一次按压的浅接触到深接触均匀取 5 帧,只用 7,055 个训练帧就取得 68.18% 的触觉—文本 Recall@1,高于使用 22,576 帧完整按压的 62.44%。这说明训练数据的有效尺度应以独立接触事件、材料、传感器实例和任务覆盖计量,而不是只数帧。

二、融合位置从“输入拼接”转向“动作表示对齐”

早期融合的直觉是:把视觉 embedding、触觉 embedding 和机器人状态拼在一起,让注意力自己学习关系。近一年的结果正在修正这一假设,因为不同层的表示承担不同功能。VLM 侧更接近语义与视觉感知,最终动作态已经压缩为即时电机命令;未来触觉描述的是 action chunk 之后的接触后果,最合适的监督位置可能在两者之间。

Representation-Aligned Tactile Grounding 用冻结线性探针比较多个接口,发现未来触觉在动作专家中间层最容易读出。把未来触觉 latent 监督放在这一层,五项真实接触任务的平均成功率达到 74%,高于 VLM 侧监督的 58% 和最终动作态监督的 62%。这项结果的重要性不在于又多了一个辅助损失,而在于它把问题从“是否使用触觉”改写成“触觉监督应塑造哪一种动作表示”。

融合还出现了方向性。Tactile-WAM 指出,无约束地把触觉反向注入视觉动力学可能损害视频与动作预测;AT-VLA 则把低频视觉—语言理解与高频触觉反馈分开。由此可见,触觉更适合作为接触门控、动作后果监督或局部残差,而不是要求它与视觉在每一层、每一帧都对称融合。

三、学习目标从“识别接触”升级为“预测动作后果”

表征对齐回答“触觉监督放在哪里”,世界模型路线回答“模型应学习什么”。Visuo-Tactile World Models 将外部视觉、指尖触觉、腕部位姿和关节状态按时间戳同步,学习接触丰富任务的联合动力学;Dream-Tac 进一步同时生成未来视觉、未来触觉和动作 chunk;ContactWorld 的消融则把关键条件归结为空间结构、时间连续性和跨模态兼容性。

这条路线比触觉分类更接近控制,因为它要求模型回答:“执行这段动作后,画面和接触场会怎样变化?”视觉未来负责全局对象运动,触觉未来负责局部压力、滑移和接触稳定性,两者共同约束动作忠实。但“能生成”仍不等于“会控制”。如果预测只用于重建损失,模型可能得到漂亮的未来视频,却没有进入动作选择、候选验证或纠偏。

Inference-time Policy Steering via Vision and Touch 代表另一种用法:在推理时用视觉—触觉世界模型给候选动作评分。它把世界模型从训练期表征器变成动作验证器,也提醒我们区分三种价值层级:重建或检索提升、策略排序提升、真实闭环成功率提升。只有最后两层才能直接支撑部署判断。

四、控制架构从单环策略变成“慢规划 + 快反射”

视觉和触觉的采样率、延迟和信息角色不同。把所有输入塞进一个 5–10 Hz 的 VLA 循环,会让触觉错过滑移与过力发生后的短窗口;反过来让高频触觉承担任务语义,也会浪费算力并放大噪声。

OmniVTA 给出了目前较完整的分层实例:双流世界模型预测视觉与触觉未来,门控策略根据预测接触状态调整模态权重,60 Hz 触觉控制器根据预测—观测差异产生修正动作。在其 object-diversity 评测中,加入触觉反射闭环后,擦拭成功率由 0.66 提到 0.80,切割由 0.50 提到 0.85。这里的增益不能只归因于“多一种传感器”,因为对照组保留了视觉—触觉策略,区别是触觉是否进入高频闭环。

TouchWorld 也采用类似分工:视觉—语言负责子任务规划,触觉世界模型预测接触目标,视觉—触觉策略生成名义动作,触觉残差负责在线修正。两项工作共同支持一个系统性结论:联合训练的最小有效单元不再是单一多模态 Transformer,而是跨频率协作的策略栈。

五、部署出现两条分叉:实时触觉与训练期触觉

并非所有平台都能长期承受触觉硬件的成本、磨损和标定维护。一条路线保留部署时触觉,用它处理遮挡、滑移和突发扰动;另一条路线把触觉当作 privileged training signal,部署时只保留视觉和本体状态。

HapticVLA 先用触觉构造过力、压力集中和滑移相关的安全奖励训练 teacher,再把动作分布蒸馏给无触觉 student。在三项易损物体任务、每项 20 次测试中,无触觉 student 的平均成功率为 86.7%。这证明“训练时有触觉、推理时无触觉”在封闭任务上可行,但不能外推为视觉已替代触觉:训练仍依赖目标硬件的触觉数据,而且新材料、突发滑移和外力扰动可能超出视觉可预测范围。

因此,两条路线不是互斥替代,而是风险分层。对接触过程稳定、材料分布窄、硬件成本敏感的任务,可优先尝试触觉蒸馏;对插入、柔性物、易碎物、动态扰动和安全关键任务,实时触觉仍更有依据。

方法谱系与适用条件

路线训练信号与融合位置部署时需要触觉最适合解决主要风险
配对表征学习视觉—触觉对比、遮蔽或重建可选材料、接触结构、跨模态检索帧级泄漏;表征分数不等于控制收益
触觉增强 VLA当前触觉作为 token、adapter 或门控输入是遮挡、接触状态、动作修正时序错位;无约束融合污染视觉先验
未来触觉 grounding未来触觉 latent 监督动作相关中间层通常是让动作表示编码接触后果接口与目标选择依赖架构和传感器
视觉—触觉世界模型动作条件下联合预测视觉与触觉未来视用法而定规划、动作验证、失败附近后训练预测逼真不等于动作忠实
慢—快分层闭环视觉慢策略 + 触觉高频 residual/reflex是滑移、压力、扰动恢复系统复杂、同步与稳定性要求高
触觉蒸馏触觉 teacher/奖励监督视觉—状态 student否传感器成本敏感、分布较封闭的任务分布外接触不可观测;训练仍需触觉

条件、分歧与尚未解决的问题

第一,表征进展与控制进展必须分账。HT-Bench 证明大规模视觉—触觉表征训练可行,但其四项评测没有真实机器人闭环,不能直接推导策略收益。相对而言,OmniVTA、TouchWorld 和表示对齐 grounding 给出了真实任务证据,但任务和硬件仍窄。

第二,评测不能只看最终成功。SoftVTBench 强调柔性物操作还应独立记录滑移、掉落、过度形变和安全交互。一个更高的成功率可能掩盖更大的峰值力、传感器损伤或恢复失败。

第三,跨传感器通用性仍是薄弱环节。触觉图像、marker displacement、局部力场与腕部六维力/力矩的空间含义和采样频率不同。共同 latent 可以缓解接口差异,但不能消除安装位置、软胶磨损、换件重标定和传感器实例偏差。

第四,世界模型必须验证动作忠实。视觉上合理的未来不能作为接触安全的充分证据;需要同时检查触觉预测、动作—状态因果顺序、候选动作排序与真实执行的一致性。

下一步最有价值的实验不是再做一个“视觉 vs 视觉+触觉”的总消融,而是逐层回答:按接触事件切分是否优于按帧切分?触觉应进入输入、动作中间层还是未来预测头?门控是否防止模态污染?缺失触觉时性能如何退化?高频 residual 是否在同等动作策略下真正提高恢复率?这些问题能把“多模态有效”拆成可复现、可部署的机制证据。

结论

近一年视觉—触觉联合训练已经越过“证明触觉有用”的阶段,开始回答触觉在何时、何处、以什么形式影响动作。最可信的方向是:用接触事件而非帧堆积数据;用未来触觉约束动作相关表示而非无差别拼接;用世界模型预测动作造成的接触后果;用视觉慢环保持语义与几何,用触觉快环修正滑移、压力和失配。与此同时,跨硬件泛化、长期维护、缺失模态退化和安全过程评测仍未解决。对系统设计者而言,触觉的价值不应由输入通道数量证明,而应由它是否让机器人更早发现错误、更快纠偏、并在真实闭环中减少危险接触来证明。

References

Accepted evidence

去重论文引用

  1. AT-VLA: Adaptive Tactile Injection for Enhanced Feedback Reaction in Vision-Language-Action ModelsXiaoqi Li, Muhe Cai, Jiadong Xu, Juan Zhu, Hongwei Fan, Yan Shen, Guangrui Ren, Hao Dong · 2026
  2. ContactWorld: What Matters in Vision-Tactile World Models for Contact-Rich ManipulationZhiyuan Zhang, Pokuang Zhou, Kaidi Zhang, Adeesh Desai, Temitope Amosa, Davood Soleymanzadeh, Jiuzhou Lei, Minghui Zheng et al. · 2026
  3. Deform360: A Massive Multi-view Visuotactile Dataset for Deformable World ModelsHongyu Li, Wanjia Fu, Xiaoyan Cong, Zekun Li, Binghao Huang, Hanxiao Jiang, Xintong He, Yiqing Liang et al. · 2026
  4. Dream-Tac: A Unified Tactile World Action Model for Contact-Rich Robot ManipulationYunfan Lou, Yifan Ye, Yankai Fu, Jun Cen, Xiaowei Chi, Yaoxu Lyu, Peidong Jia, Sirui Han et al. · 2026
  5. HapticVLA: Contact-Rich Manipulation via Vision-Language-Action Model without Inference-Time Tactile SensingKonstantin Gubernatorov, Mikhail Sannikov, Ilya Mikhalchuk, Egor Kuznetsov, Makar Artemov, Ogunwoye Faith Ouwatobi, Marcelino Fernando, Artem Asanov 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. 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
  8. In-the-Wild Compliant Manipulation with UMI-FTHojung Choi, Yifan Hou, Chuer Pan, Seongheon Hong, Austin Patel, Xiaomeng Xu, Mark R. Cutkosky, Shuran Song · 2026
  9. Inference-time Policy Steering via Vision and TouchYilin Wu, Zilin Si, Zeynep Temel, Oliver Kroemer, Andrea Bajcsy · 2026
  10. Learning Tactile-Aware Quadrupedal Loco-Manipulation PoliciesPokuang Zhou, Yuhao Zhou, Quan Khanh Luu, Seungho Han, Heng Zhang, Binghao Huang, Yunzhu Li, Arash Ajoudani et al. · 2026
  11. Multisensory Continual Learning: Adapting Pretrained Visuomotor Policies to ForceJaden Clark, Changhao Wang, Yihuai Gao, Seongheon Hong, Hojung Choi, Mark Cutkosky, Yifan Hou, Shuran Song · 2026
  12. OmniVTA: Visuo-Tactile World Modeling for Contact-Rich Robotic ManipulationYuhang Zheng, Songen Gu, Weize Li, Yupeng Zheng, Yujie Zang, Shuai Tian, Xiang Li, Ce Hao et al. · 2026
  13. PATCH: Action-Chunk-Conditioned Latent Patch Innovation Monitoring for Robot ManipulationYanan Zhou, Ranpeng Qiu, Yincong Chen, Jiajie Cui, Weiming Zhi · 2026
  14. RCT: A Robot-Collected Touch-Vision-Language Dataset for Tactile GeneralizationJingbo He, Michael Färber, Roberto Calandra · 2026
  15. Representation-Aligned Tactile Grounding for Contact-Rich Robotic ManipulationRuilin Chen, Jingkai Jia, Tong Yang, Xinyu Zhou, Qiao Sun, Jiangwei Zhong, Shizeng Zhang, Nuo Chen et al. · 2026
  16. RGB-S: Image-Aligned Tactile Saliency for Robust Dexterous ManipulationShengcheng Luo · 2026
  17. SoftVTBench: A Safety-Aware Visuo-Tactile Benchmark for Physically Constrained Robotic Manipulation of Deformable ObjectsBowen Jing, Mingxin Wang, Ruiyang Hao, Chenchen Ge, Hanwen Shen, Junjie He, Yang Cui, Yiming Hou et al. · 2026
  18. TAC-LOCO: Unified Whole-Body Control for Quadrupedal TACtile-Informed LOCO-ManipulationMuqun Hu, Yuhao Zhou, Kabir Ray Malik, Chi Lin, Won Suk Lee, Yu She, Yan Gu · 2026
  19. 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
  20. 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
  21. Tactile-WAM: Touch-Aware World Action Model with Tactile Asymmetric AttentionSiyu Wu, Linjing You, Junjie Zhu, Yaozu Liu, Changhao Zhang, Jian Liu, Weiqiang Wang, Qi Li et al. · 2026
  22. TAMEn: Tactile-Aware Manipulation Engine for Closed-Loop Data Collection in Contact-Rich TasksLongyan Wu, Jieji Ren, Chenghang Jiang, Junxi Zhou, Shijia Peng, Ran Huang, Guoying Gu, Li Chen et al. · 2026
  23. TouchWorld: A Predictive and Reactive Tactile Foundation Model for Dexterous ManipulationJianyi Zhou, Feiyang Hong, Yunhao Li, Yicheng Zhao, Yongjue Cen, Zirui Liu, Jiakang Huang, Zirui Chen et al. · 2026
  24. Uncertainty Quantification for Flow-Based Vision-Language-Action ModelsRalf Römer · 2026
  25. Validate the Dream Before You Trust Its Verdict: Admissibility for World-Model SimulatorsChristian Oefinger · 2026
  26. Visuo-Tactile World ModelsCarolina Higuera, Sergio Arnaud, Byron Boots, Mustafa Mukadam, Francois Robert Hogan, Franziska Meier · 2026
  27. You Only Touch Once: 6-DoF Object Pose Estimation from Single Tactile ContactPengfei Ye, Yuxiang Ma, Haonan Chen, Guangming Wang, Yixiong Jing, Brian Sheil, Yilun Du, Edward Adelson · 2026
完整证据附录(30 条正式证据事件)
  • Time range: 2025-07-22..2026-07-22
  • Events: 30
  • 每个事件一节,标题即锚点;trace-map 中的 event 链接跳转到这里。

EA-TWM-READ-0008

  • Claim: TAMEn 用动捕精度模式与 VR 便携模式平衡数据质量和环境多样性,并把人在环的触觉可视化恢复数据纳入金字塔式数据配方。
  • Stance: support | Confidence: direct
  • Paper: 2604.07335 TAMEn: Tactile-Aware Manipulation Engine for Closed-Loop Data Collection in Contact-Rich Tasks
  • Locator: Abstract (full-text section)
  • Evidence: 摘要明确列出精度/便携双模式采集、触觉恢复遥操作和人在环恢复数据。
  • Quote: “To balance data quality and environmental diversity, we implement a dual-mode acquisition pipeline: a precision mode leveraging motion capture for high-fidelity demonstrations, and a portable mode utilizing VR-based tracking for in-the-wild acquisition and tactile-visualized recovery teleoperation. Building on this hardware, we unify large-scale tactile pretraining, task-specific bimanual demonstrations, and human-in-the-loop recovery data into a pyramid-structured data regime, enabling closed-l”
  • Authors: longyan-wu; jieji-ren; chenghang-jiang; et al.

EA-TWM-READ-0005

  • Claim: 触觉数据的有效形态不止原始 tactile image,还包括 marker displacement 等显式接触几何与滑移特征。
  • Stance: support | Confidence: direct
  • Paper: 2606.04825 HapTile: A Haptic-Informed Vision-Tactile-Language-Action Dataset for Contact-Rich Imitation Learning
  • Locator: 4.2 Vision-Based Tactile Sensing and Marker Tracking
  • Evidence: HapTile 的每个夹爪手指安装视觉触觉传感器,接触会带来图像变化和 marker displacement;论文把 marker-motion 信号保存进数据集并用于 haptic feedback,实验也比较 vision-only、vision+tactile image 与 vision+tactile+marker 表征。
  • Quote: “We smooth the signal with separate rise and release factors, producing a stable tactile motion estimate. The resulting marker-motion signal is stored in HapTile and is also used to drive haptic feedback.”
  • Authors: amirhosein-alian; yongqiang-zhao; shiyi-gu; et al.

EA-TWM-READ-0002

  • Claim: 触觉世界模型可以被扩展为同时生成未来视觉、未来触觉和动作 chunk 的世界动作模型。
  • Stance: support | Confidence: direct
  • Paper: 2606.08737 Dream-Tac: A Unified Tactile World Action Model for Contact-Rich Robot Manipulation
  • Locator: Abstract (full-text section)
  • Evidence: Dream-Tac 把 world action model 扩展到触觉,联合建模当前视觉、触觉、语言指令下的未来视觉观测、未来触觉观测和动作 chunk,并加入 contact-gated visuotactile fusion 与 contact-aware attention bias。
  • Quote: “In this paper, we propose Dream-Tac, a unified Tactile-World Action Model that jointly models actions, future visual observations, and tactile dynamics. Specifically, Dream-Tac introduces (i) contact-gated visuotactile fusion to selectively integrate tactile signals and (ii) a contact-aware attention bias to better regulate cross-modal interactions during manipulation.”
  • Authors: yunfan-lou; yifan-ye; yankai-fu; et al.

EA-TWM-READ-0006

  • Claim: 腕部六维力/力矩可作为未来触觉 latent 的先行条件,用于预测短时域接触变化。
  • Stance: support | Confidence: direct
  • Paper: 2606.11184 TacForeSight: Force-Guided Tactile World Model for Contact-Rich Manipulation
  • Locator: IV-D 1 World Model Conditioning
  • Evidence: TacForeSight 的 TacForceWM 从双指触觉观测出发,以高频腕部 force/torque 为条件预测短时域触觉 latent dynamics;ablation 中 wrist wrench 条件的未来触觉预测优于无条件版本,MSE 从 0.027 降到 0.017,cosine 从 0.954 提升到 0.992。
  • Quote: “We first ablate the conditioning modality used in the tactile world model. We compare four conditioning designs: no external condition, RGB image condition, robot state condition, and wrist wrench condition. Prediction quality is evaluated using Mean Squared Error (MSE), cosine similarity (Cos), and symmetric KL divergence (KL sym ), which respectively measure element-wise prediction error, latent directional consistency, and latent distribution consistency. As shown in Table II , wrist wrench c”
  • Authors: yujie-zang; yuhang-zheng; xian-nie; et al.

EA-TWM-READ-0003

  • Claim: 在接触丰富操作中,触觉世界模型的关键不是简单增加模态,而是让表征同时具备空间结构、时间连续性和跨模态兼容性。
  • Stance: support | Confidence: direct
  • Paper: 2606.13877 ContactWorld: What Matters in Vision-Tactile World Models for Contact-Rich Manipulation
  • Locator: Abstract (full-text section)
  • Evidence: ContactWorld 在 12 个接触丰富任务上比较视觉与触觉表征;点云把平均规划成功率从腕部视角 20.7% 和前视 22.0% 提升到 32.1%,点云加触觉力场进一步到 36.1%。作者强调触觉效果取决于跨模态表征兼容,而非模态数量本身。
  • Quote: “Abstract Contact-rich manipulation requires world models to reason over complex contact dynamics from multimodal sensory observations. However, it remains unclear which representation properties fundamentally support stable long-horizon planning in contact-rich settings. In this paper, we present ContactWorld, a benchmark and systematic empirical study of vision-tactile world models spanning 12 contact-rich manipulation tasks, including insertion, disassembly, screwing, and exploratory interacti”
  • Authors: zhiyuan-zhang; pokuang-zhou; kaidi-zhang; et al.

EA-TWM-READ-0004

  • Claim: 触觉世界模型也可以在推理期作为候选动作验证器,而不只是训练期的动态模型。
  • Stance: support | Confidence: direct
  • Paper: 2606.14981 Inference-time Policy Steering via Vision and Touch
  • Locator: 5 Experiments
  • Evidence: ViTaL 学习 visuo-tactile latent world model,结合视觉和文本条件触觉 verifier,对候选动作进行长时域视觉模式选择和短时域触觉 refinement;真实机器人任务包括 wiping、insertion 和 pipette transfer。
  • Quote: “We evaluate ViTaL on three real-world contact-rich manipulation tasks. Our experiments are designed to answer three questions: whether multimodal visuo-tactile steering improves policy over unimodal guidance (Sec. 5.1 ); whether a learned latent world model and semantically aligned verifier provides reliable rewards for predicted outcomes (Sec. 5.2 ); and whether the proposed bi-level optimization balances performance and efficiency compared to naive visual-tactile fusion (Sec. 5.3 ). Robot Setu”
  • Authors: yilin-wu; zilin-si; zeynep-temel; et al.

EA-TWM-READ-0015

  • Claim: Flow-based VLA 在真实部署时缺少置信度机制会形成安全与恢复缺口;velocity-field disagreement 可把动作预测的不可靠性转成失败检测和主动微调信号。
  • Stance: support | Confidence: direct
  • Paper: 2606.18043 Uncertainty Quantification for Flow-Based Vision-Language-Action Models
  • Locator: Abstract (full-text section)
  • Evidence: 作者将真实非平稳环境中的分布外场景描述为 VLA 可能“无预警失败”的关键限制,并提出用小 ensemble 的 velocity-field disagreement 量化 epistemic uncertainty;LIBERO 实验显示该不确定性与下游表现、失败检测和主动采样相关。
  • Quote: “This presents a critical limitation for real-world deployment in non-stationary environments, where models inevitably encounter scenarios outside their pretraining distribution and may fail without warning. To address this, we derive an efficient method for quantifying epistemic uncertainty in flow-matching models by leveraging velocity-field disagreement (VFD) across a small ensemble.”
  • Authors: ralf-rmer

EA-UMI-READ-0015

  • Claim: 软物体和形变场景中的数据质量主要受可观测性约束:必须同时记录多视角全局运动、触觉局部形变和可用于评测的 3D 轨迹。
  • Stance: support | Confidence: direct
  • Paper: 2607.05390 Deform360: A Massive Multi-view Visuotactile Dataset for Deformable World Models
  • Locator: Abstract (full-text section)
  • Evidence: 论文认为形变物体有高维状态和复杂材料属性,接触诱发的局部形变常被末端执行器或物体遮挡;已有数据集常缺对象多样性、依赖合成数据,或缺高保真标注与接触形变。Deform360 采集 198 个日常物体、1,980 个交互序列、215 小时以上数据、41 个环视相机和双臂触觉 UMI gripper,并用 markerless 3D tracking 提取稠密几何与运动。
  • Quote: “To address this, we present Deform360, a large-scale visuotactile dataset featuring 198 daily-life objects, 1,980 interaction sequences, and over 215 hours of observations from 41 surround-view cameras and bimanual tactile grippers to capture both global motion and contact-induced local deformations. Leveraging a novel markerless visuotactile 3D tracking pipeline to extract dense geometry and motion, we systematically evaluate current state-of-the-art world models, comparing 2D video models agai”
  • Authors: hongyu-li; wanjia-fu; xiaoyan-cong; et al.

EA-UMI-READ-0002

  • Claim: UMI-style data is useful for contact-rich manipulation when the collection interface records force/torque, depth, pose, and grasp-force signals; the paper also implies that vision/trajectory-only data is insufficient for force-sensitive tasks.
  • Stance: conditional | Confidence: direct
  • Paper: 2601.09988 In-the-Wild Compliant Manipulation with UMI-FT
  • Locator: Abstract (full-text section)
  • Evidence: The HTML full text reports that UMI-FT mounts compact six-axis force/torque sensors on each finger, uses multimodal demonstrations to train adaptive compliance policies, and shows diverse in-the-wild data outperforming limited scene-diversity data in a skewer task.
  • Quote: “Abstract Many manipulation tasks require careful force modulation. With insufficient force the task may fail, while excessive force could cause damage. The high cost, bulky size and fragility of commercial force/torque (F/T) sensors have limited large-scale, force-aware policy learning. We introduce UMI-FT, a handheld data-collection platform that mounts compact, six-axis force/torque sensors on each finger, enabling finger-level wrench measurements alongside RGB, depth, and pose. Using the mult”
  • Authors: hojung-choi; yifan-hou; chuer-pan; et al.

EA-TWM-READ-0001

  • Claim: VT-WM 的训练序列同步记录腕部位姿、关节位置、外部视觉和两个指尖触觉视频,并使用时间戳对齐后降采样训练。
  • Stance: conditional | Confidence: direct
  • Paper: 2602.06001 Visuo-Tactile World Models
  • Locator: B.0.1 Training dataset
  • Evidence: 训练数据段明确列出了同步的本体状态、外部视频与双指触觉视频数据流。
  • Quote: “Each sequence contains multimodal data streams: proprioceptive information (wrist pose, joint positions), exocentric video from the camera, and video from each Digit 360 fingertip sensor. All data streams were synchronized using timestamps and downsampled to 6 FPS for training the world model. Our training dataset for V-WM and VT-WM consists of 124 demonstrations totaling 112k datapoints, with each demonstration averaging 40 seconds. For validation, we use 26 demonstrations spanning all tasks, c”
  • Authors: carolina-higuera; sergio-arnaud; byron-boots; et al.

EA-TWM-READ-0007

  • Claim: 触觉信息用于 VLA 或世界模型时,低频视觉语言理解和高频触觉反馈应在架构上分离。
  • Stance: conditional | Confidence: direct
  • Paper: 2605.07308 AT-VLA: Adaptive Tactile Injection for Enhanced Feedback Reaction in Vision-Language-Action Models
  • Locator: 5 Conclusion
  • Evidence: AT-VLA 把系统分为慢速视觉语言流和快速触觉流,慢速流负责任务理解和视觉定位,快速流以高频处理触觉反馈;作者采用 3:1 的快慢流频率比,并在真实接触丰富任务中验证 adaptive tactile injection、tactile gate、adaptive cross-attention 和 reaction dual-stream 的作用。
  • Quote: “In summary, AT-VLA introduces an adaptive framework that seamlessly integrates tactile sensing into vision-language-action models. Through the Adaptive Tactile Injection mechanism, AT-VLA dynamically balances pretrained visual-language knowledge with newly learned tactile representations, preserving model integrity while enhancing action precision. The Tactile Reaction Dual-Stream mechanism further enables rapid, high-frequency tactile responses by decoupling slow perceptual reasoning from fast”
  • Authors: xiaoqi-li; muhe-cai; jiadong-xu; et al.

EA-TWM-READ-0014

  • Claim: 在视觉遮挡或视觉退化下,显式把触觉接触投影到图像域可以改善空间对齐;但这种收益依赖对运动学和标定误差导致的空间不确定性建模。
  • Stance: conditional | Confidence: direct
  • Paper: 2606.08765 RGB-S: Image-Aligned Tactile Saliency for Robust Dexterous Manipulation
  • Locator: Abstract (full-text section)
  • Evidence: 作者称视觉观测不可靠或被遮挡时,稀疏异构触觉与稠密视觉表示的对齐是核心挑战;方法使用正运动学和相机标定投影触觉传感器位置,并用力调制高斯 saliency maps 建模运动学和标定误差带来的空间不确定性。
  • Quote: “We then render force-modulated Gaussian saliency maps to model spatial uncertainty arising from kinematic and calibration errors.”
  • Authors: shengcheng-luo

EA-TWM-READ-0010

  • Claim: 触觉不是简单加 token 就能提升;接触任务中的 RGB 未来可能视觉上合理但物理上不完整,而无约束触觉注入还可能污染视频和动作预测。
  • Stance: conditional | Confidence: direct
  • Paper: 2606.26663 Tactile-WAM: Touch-Aware World Action Model with Tactile Asymmetric Attention
  • Locator: Abstract (full-text section)
  • Evidence: 作者指出 insertion、assembly、search、reorientation 依赖 slip、jamming、contact normals 和小对齐误差,这些状态在 RGB 中弱可见或不可见;同时他们定义 tactile pollution:无约束触觉 token 注入会迫使视觉 dynamics model 吸收稀疏局部事件式接触信号,从而退化视频和动作预测。
  • Quote: “In contact-rich manipulation, however, visually plausible futures can be physically incomplete: insertion, assembly, search, and reorientation often depend on slip, jamming, contact normals, or small alignment errors that are weakly visible or hidden in RGB. A natural solution is to predict future tactile states, however, we identify tactile pollution , a failure mode where unconstrained tactile-token injection degrades video and action prediction by forcing a visual dynamics model to absorb spa”
  • Authors: siyu-wu; linjing-you; junjie-zhu; et al.

EA-TWM-READ-0011

  • Claim: 力觉/触觉/音频能揭示图像不可直接观测的交互状态;但这些模态硬件和任务依赖强,所以需要在少量多传感数据上适配既有视觉策略,而不是预训练时穷尽所有传感器。
  • Stance: conditional | Confidence: direct
  • Paper: 2606.30988 Multisensory Continual Learning: Adapting Pretrained Visuomotor Policies to Force
  • Locator: Abstract (full-text section)
  • Evidence: 作者称接触丰富任务常依赖 vision 之外的 sensory data,force、tactile 或 audio feedback 能揭示 images 中不可直接观察的 interaction states;但这些模态 hardware- and task-specific,且大规模多传感数据稀缺。他们提出 MuSe,将 limited multisensory data 融入 pretrained vision-only policies,并以 force-torque sensing 做真实任务案例。
  • Quote: “Abstract Robot manipulation often depends on sensory data beyond vision, especially in contact-rich tasks where force, tactile, or audio feedback reveals interaction states not directly visible from images. Yet such modalities are hardware- and task-specific, and large multisensory datasets remain scarce, making it impractical to pretrain policies with every sensor they may encounter. We study multi-sensory continual learning : adapting a pretrained robot policy to new tasks with newly introduce”
  • Authors: jaden-clark; changhao-wang; yihuai-gao; et al.

EA-TWM-READ-0009

  • 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.

EA-TWM-READ-0012

  • Claim: 只看任务完成会低估传感器感知误差的真实风险;柔性物操作需要把安全交互、滑移/掉落和过度形变作为独立评测目标。
  • Stance: limit | Confidence: direct
  • Paper: 2607.04234 SoftVTBench: A Safety-Aware Visuo-Tactile Benchmark for Physically Constrained Robotic Manipulation of Deformable Objects
  • Locator: 1 Introduction
  • Evidence: 作者指出现有 manipulation benchmarks 多以 success 为中心,很少评估执行过程是否物理安全;SoftVTBench 分开报告 Goal Success 和 Safety Success,后者要求无掉落并限制峰值形变。实验显示 success-only evaluation 会显著高估策略表现,而触觉感知可改善 Safety Success 并降低物体形变。
  • Quote: “Experiments show that success-only evaluation substantially overstates policy performance, as a large fraction of goal-completing rollouts violate physical safety, and that adding tactile sensing improves Safety Success while keeping Goal Success comparable and reduces object deformation during execution.”
  • Authors: bowen-jing; mingxin-wang; ruiyang-hao; et al.

EA-TWM-READ-0013

  • Claim: 世界模型/生成式仿真器不能仅凭视觉逼真度作为具身策略评测裁判;必须验证其是否按动作正确响应,并建立 admissibility/validation 梯度后才能把闭环判决当证据。
  • Stance: gap | Confidence: direct
  • Paper: 2607.07196 Validate the Dream Before You Trust Its Verdict: Admissibility for World-Model Simulators
  • Locator: Abstract (full-text section)
  • Evidence: 作者指出机器人中 World Models 越来越被用于模拟动作后果并给出 success/safety verdict,但视频生成指标如 FVD 奖励视觉真实感,却忽略世界是否对 policy actions 正确响应;他们主张作为 test oracle 的 WM 需要先通过 accreditation,并提出 L0-L4 admissibility ladder。
  • Quote: “Abstract Across robotics, World Models (WMs) are increasingly used to evaluate action policies by simulating the consequences of actions in an imagined world, and returning a success or safety verdict. Yet a verdict is only as trustworthy as the WM that produced it, and the WM itself needs to be certified. In video-generation WMs , fidelity metrics such as Fréchet Video Distance (FVD) reward visual realism, but ignore whether the world responds correctly to the policy’s actions, including those”
  • Authors: christian-oefinger

EA-LOCOMANIP-2026-0012

  • Claim: Adding tactile-command tracking at the low level raised insertion success from 0.70 to 0.85, full reorientation-plus-insertion from 0.60 to 0.80, and valve tightening from 0.80 to 0.85.
  • Stance: support | Confidence: direct
  • Paper: 2604.27224 Learning Tactile-Aware Quadrupedal Loco-Manipulation Policies
  • Locator: IV-B Experimental Results and Analyze
  • Evidence: The paper compares variants with the same tactile-aware high level but different low-level tactile tracking.
  • Quote: “Finally, comparing Baseline 3 (P3) with our full method demonstrates the benefit of incorporating tactile commands into the low-level policy: our method further increases the success rate from 0.70 to 0.85 on Task 1 insertion (+0.15), from 0.60 to 0.80 on Task 1 whole (+0.20), and from 0.80 to 0.85 on Task 2 (+0.05), and achieves 1.00 on Task 3 (compared to 0.20 for P1).”
  • Authors: pokuang-zhou; yuhao-zhou; quan-khanh-luu; et al.

EA-LOCOMANIP-2026-0021

  • Claim: In 10 matched hardware trials, tactile-informed TAC-LOCO achieved 90% dynamic loco-manipulation success versus 50% for Deep WBC with a fixed gripper.
  • Stance: support | Confidence: direct
  • Paper: 2607.10132 TAC-LOCO: Unified Whole-Body Control for Quadrupedal TACtile-Informed LOCO-Manipulation
  • Locator: 6.5 Baseline comparison
  • Evidence: The hardware baseline comparison isolates learned grasp regulation under the same command set.
  • Quote: “We conduct 10 hardware trials using the same set of loco-manipulation commands as in Sec. 6 . Table 4 compares the success rates of our policy and the baseline, showing that our tactile-informed policy achieves a substantially higher success rate. Figure 11 shows that the baseline suffers from gradual object slip during the task. Since the gripper width remains fixed, the policy cannot actively suppress slip once the external force changes. Table 4: Deployment success rate comparison with baseli”
  • Authors: muqun-hu; yuhao-zhou; kabir-ray-malik; et al.

EA-VTTRAIN-2026-0003

  • Claim: OmniViTac 将视觉—触觉—动作联合训练的数据底座扩展到 21,879 条轨迹、86 个任务和 126 个物体,覆盖六类接触交互模式。
  • Stance: support | Confidence: direct
  • Paper: 2603.19201 OmniVTA: Visuo-Tactile World Modeling for Contact-Rich Robotic Manipulation
  • Locator: III-A 2 Dual-Embodiment Data Collection
  • Evidence: Table I 报告任务、轨迹、物体、视觉、触觉和采集设置的规模。
  • Quote: “OmniViTac (Ours) 86 21,879 126 RealSense/Fisheye Xense/Gelsight Mini/DM-Tac/Tac3D 3 30 60 ✓ UMI-style/Kinesthetic Teaching”
  • Authors: yuhang-zheng; songen-gu; weize-li; et al.

EA-VTTRAIN-2026-0004

  • Claim: 在 OmniVTA 的真实机器人评测中,60 Hz 触觉反射闭环相对移除该控制器的开放环版本,把擦拭成功率从 0.66 提到 0.80、切割从 0.50 提到 0.85;预测触觉只有连接到快速纠偏才形成完整控制收益。
  • Stance: support | Confidence: direct
  • Paper: 2603.19201 OmniVTA: Visuo-Tactile World Modeling for Contact-Rich Robotic Manipulation
  • Locator: V-B Overall Performance
  • Evidence: Table III 的 object-diversity 列隔离了移除 RLTC 的开放环变体,两个代表任务均显著提升。
  • Quote: “OmniVTA w/o RLTC 0.66 0.40 0.25 0.40 0.30 0.20 0.50 0.50 0.20 0.40 0.35 0.20 0.70 0.40 0.30 OmniVTA 0.80 0.58 0.60 0.55 0.48 0.63 0.85 0.83 0.60 0.60 0.50 0.40 0.90 0.65 0.65”
  • Authors: yuhang-zheng; songen-gu; weize-li; et al.

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-SENSORERR-READ-0010

  • Claim: 物体 6-DoF 位姿误差在遮挡、弱光、反光/透明表面下会让视觉方法失效;单次双触点触觉可作为视觉不可靠时的位姿观测补充。
  • Stance: support | Confidence: direct
  • Paper: 2606.28899 You Only Touch Once: 6-DoF Object Pose Estimation from Single Tactile Contact
  • Locator: Abstract (full-text section)
  • Evidence: 作者明确指出视觉位姿估计常在遮挡、差光照、反光或透明表面下失败,并提出 tactile-only pose estimation:把触觉接触表示成局部 3D 点云,结合校准传感器位姿恢复完整 6-DoF object pose;实验在视觉不可靠时优于视觉和几何基线。
  • Quote: “Abstract Accurate 6-DoF object pose estimation is fundamental to robotic manipulation, yet vision-based methods often fail under occlusion, poor lighting, and reflective or transparent surfaces.”
  • Authors: pengfei-ye; yuxiang-ma; haonan-chen; et al.

EA-VTTRAIN-2026-0006

  • Claim: 在相同 held-out-contact-sequence 测试池上,沿一次按压从浅到深均匀取 5 帧只用 7,055 个训练帧,却把触觉—文本 Recall@1 提到 68.18%,高于使用 22,576 帧完整按压的 62.44% 和只取最深 5 帧的 59.52%。
  • Stance: support | Confidence: direct
  • Paper: 2606.31694 RCT: A Robot-Collected Touch-Vision-Language Dataset for Tactile Generalization
  • Locator: 5 Experimental Evaluation, Table 5
  • Evidence: Uniform5 在相同测试池和模型配方下以更少帧取得更高检索性能。
  • Quote: “Training density Training frames Tac text R@1 Full press ( 16 frames/sequence) 22,576 62.44% Deep5 (five deepest frames) 7,055 59.52% Uniform5 (five frames across press) 7,055 68.18%”
  • Authors: jingbo-he; michael-frber; roberto-calandra

EA-SENSORERR-READ-0012

  • Claim: 触觉在灵巧操作中补足视觉/语言无法稳定观测的接触隐变量;滑移、力不匹配、接触稳定性等局部误差需要比语义规划更快的反馈通道。
  • Stance: support | Confidence: direct
  • Paper: 2607.07287 TouchWorld: A Predictive and Reactive Tactile Foundation Model for Dexterous Manipulation
  • Locator: Abstract (full-text section)
  • Evidence: 作者把日常灵巧操作的误差来源明确落在滑移、错位、不稳定抓取和力不匹配上,并指出视觉/语言不能可靠揭示力、滑移和接触稳定性;其分层策略将视觉语言子任务规划、触觉世界模型预测和高频触觉残差修正分开。
  • Quote: “Vision and language provide semantic and geometric guidance, but they cannot reliably reveal hidden contact states such as force, slip, and contact stability.”
  • Authors: jianyi-zhou; feiyang-hong; yunhao-li; et al.

EA-VTTRAIN-2026-0001

  • Claim: 在五项真实接触操作中,把未来触觉 latent 监督施加到动作专家中间层取得 74% 平均成功率,高于 VLM 侧的 58% 和最终动作态的 62%;联合训练的收益取决于监督接口是否与动作条件接触动力学对齐。
  • Stance: support | Confidence: direct
  • Paper: 2607.14609 Representation-Aligned Tactile Grounding for Contact-Rich Robotic Manipulation
  • Locator: 4.2 Main Results
  • Evidence: 同一触觉条件 VLA 下的接口控制比较表明,中间动作专家表示优于感知侧和最终动作侧。
  • Quote: “The key comparison is among future tactile prediction interfaces: VLM-level prediction and final-action-state prediction achieve 58% and 62% average success, whereas grounding the intermediate action-expert representation reaches 74%.”
  • Authors: ruilin-chen; jingkai-jia; tong-yang; et al.

EA-VTTRAIN-2026-0002

  • Claim: 触觉可以作为训练期 teacher 和安全奖励的监督,再蒸馏进只使用视觉与本体状态的部署 VLA;在三项易损物体任务、每项 20 次测试中,该无触觉 student 达到 86.7% 平均成功率。
  • Stance: conditional | Confidence: direct
  • Paper: 2603.15257 HapticVLA: Contact-Rich Manipulation via Vision-Language-Action Model without Inference-Time Tactile Sensing
  • Locator: V CONCLUSION
  • Evidence: Tactile Distillation 将触觉条件 teacher 的动作目标与真实示教混合训练 student,Table I 报告三任务平均成功率。
  • Quote: “To enable inference of VLA with SA-RWFM without tactile sensors, HapticVLA employs Tactile Distillation (TD), which distills tactile representations from vision and state inputs. Experiments across diverse real-world tasks show that HapticVLA achieves mean success rate of 86.7% consistently outperforming baselines in manipulating fragile and soft objects without any haptic feedback.”
  • Authors: konstantin-gubernatorov; mikhail-sannikov; ilya-mikhalchuk; 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.

EA-VTTRAIN-2026-0005

  • Claim: 视觉—触觉表示训练不能把单帧当独立数据单元:同一次机器人按压的相邻帧是近重复观测,frame-random 切分会把物理交互泄漏到训练和测试,因此至少应按 contact sequence、最好再按材料或传感器实例做 holdout。
  • Stance: limit | Confidence: direct
  • Paper: 2606.31694 RCT: A Robot-Collected Touch-Vision-Language Dataset for Tactile Generalization
  • Locator: 4 Evaluation Protocol
  • Evidence: RCT 与 TVL/HCT 审计都显示同一按压序列可同时出现在训练和测试,原始像素近邻也能利用该结构。
  • Quote: “Our evaluation asks whether tactile representations generalize across physical variation that matters in robot deployment. The central unit is the contact sequence : all frames recorded during one robot press on one material, with one sensor, at one contact position. Frames within a sequence are strongly correlated; in RCT, successive frames differ by only a indentation step. A frame-random split can therefore place near-duplicate observations of the same physical interaction in both training an”
  • Authors: jingbo-he; michael-frber; roberto-calandra

EA-SENSORERR-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.

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。