VLA 与模型 · VLA and Foundation Models

Sparse language, dense vision, and continuous action alignment in VLA systems

Language, Vision, and Action Alignment in VLA Systems

更新于

30 秒结论

先读这个判断

本文综述一个正在 VLA / 机器人基础模型中变得清晰的问题:语言、视觉和动作并不是三种可直接拼接的 token 流。语言通常以任务级或阶段级指令出现,语义稀疏、时间标注粗;图像/视频是高维稠密信号,容易主导训练;动作是连续、闭环、受本体和控制器约束的物理量。近六个月的论文显示,失败常不是某个 backbone 不够大,而是语言-动作、视觉-动作、动作-控制器三条接口没有被显式对齐。

精读论文
15
正式证据事件
15
知识单元
EA-MODEL, EA-SENSOR, EA-XEMBODIMENT, EA-ALIGN
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

本文综述一个正在 VLA / 机器人基础模型中变得清晰的问题:语言、视觉和动作并不是三种可直接拼接的 token 流。语言通常以任务级或阶段级指令出现,语义稀疏、时间标注粗;图像/视频是高维稠密信号,容易主导训练;动作是连续、闭环、受本体和控制器约束的物理量。近六个月的论文显示,失败常不是某个 backbone 不够大,而是语言-动作、视觉-动作、动作-控制器三条接口没有被显式对齐。

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。

  • 检索时间:2026 年 1 月 14 日至 7 月 14 日;经典论文只作背景基线,不被包装成最新发现。
  • 覆盖规模:924 条去重候选、131 篇可用全文、24 篇直接相关论文;检索覆盖语言—动作、视觉—动作、动作接口、跨本体迁移与接触边界。
  • 停止条件:候选量、全文量、关键维度和连续两轮低新增率均已满足;核心引用保持克制,不等于只检索了文中出现的十余篇论文。
  • 限制:本综述聚焦机器人/VLA,不覆盖一般多模态大模型的全量图文对齐文献;2026 年论文多为预印本,结论应按“当前证据”而非领域共识处理。

背景脉络

RT-1 把大规模真实机器人数据、语言指令和视觉观测接到 Transformer 控制策略上,提出机器人也可能走 scaling 路线。RT-2 进一步把 VLM/VQA 风格的互联网语义知识迁移到机器人动作输出。Open X-Embodiment / RT-X、Octo、OpenVLA 则把跨机器人数据、通用策略和开源 VLA 推向主流。

与此同时,Diffusion Policy 和 π0 代表了另一条线:用扩散或 flow matching 处理连续、高维、可多峰的动作分布。

这条历史脉络解释了为什么 2026 年的新论文开始集中处理“接口”问题:VLM 提供语义和视觉先验,但机器人动作不是语言 token,也不是图像 patch;动作要通过本体、控制器、状态估计和接触反馈落到物理世界。

中心判断

语言稀疏

这带来一个重要推论:如果语言只在输入端作为 prompt 出现,而没有阶段级、动作级或因果级监督,它很容易成为装饰性条件。ZR-0 的 dense embodied chain-of-thought 方向可以理解为把稀疏语言“加密”:把场景描述、任务进度、未来计划、原子子任务、目标框和离散动作 token 作为 dense supervision,使高层认知过程更细粒度地对齐到动作专家(Training Vision-Language-Action Models with Dense Embodied Chain-of-Thought Supervision)。

动作连续

动作不是自然语言词表。离散 action token 对 autoregressive VLA 很友好,但会带来压缩和解码问题:同一个 token 在不同关节状态、物体位姿、接触条件下不应解码成同一个连续控制量。SA-VLA 因此把机器人状态注入 action tokenizer,试图缩小离散 token 到连续控制之间的 compression 空白(SA-VLA: State-aware tokenizer for improving Vision-Language-Action Models' performance)。

另一类论文绕开“先离散再还原”的瓶颈,用 flow matching / diffusion 直接学习连续动作流。Learning Action Priors for Cross-embodiment Robot Manipulation 的关键判断是:VLA 从 VLM 继承了视觉和语言先验,但 action module 往往从零学习物理运动;

视觉稠密

SSI-Policy 则构造 RGB-only structured scene interface,把 monocular depth、language-grounded object layout 和 instruction-conditioned 2D motion trajectory 放到中间层,再交给 diffusion action planner;这等于在视觉和控制之间加了一个可解释、任务对齐的空间接口(SSI-Policy: Learning Structured Scene Interfaces for Vision-Language Robotic Manipulation)。

物理闭环

真实机器人中的 action 不是抽象标签,而是控制器输入。SPACE 指出,不同机器人甚至同一型号不同硬件单元中,同一 action command 都可能产生不同 motion;因此它用 Cartesian state delta 做共享动作表示,再用 Action Adapter 转成具体机器人控制命令(SPACE: Enabling Learning from Cross-Robot Data Toward Generalist Policies)。

核心机制

1. 对齐难题本质上是粒度错配

语言粒度粗,视觉粒度细,动作粒度连续。把它们统一成 token 后,表面上进入同一个模型,实际上仍有三个未解决的 mapping:语言目标如何落到动作阶段,视觉几何如何落到可执行位姿,连续控制如何适配本体和控制器。

当前论文的共同趋势是为每个 mapping 引入更强的中间约束,例如 dense ECoT、语言—动作 pretraining、state-aware action tokenizer、action prior、structured scene interface、Cartesian state delta。

2. 视觉不是越稠密越好,关键是稠密信号是否被动作约束

图像 patch、视频帧和多视角输入可以提供大量信息,但也会制造捷径。Sparse2Act 和 SSI-Policy 的价值在于把视觉压缩成 action-aligned geometry 或 task-aligned scene interface,而不是让视觉 token 自己决定什么重要。换句话说,机器人视觉表征的目标不只是识别对象,而是把对象、空间关系和动作可达性放到同一坐标系里。

3. 动作空间是 VLA 与普通 VLM 的最大分水岭

普通 VLM 可以用文本 token 作为输出;VLA 的输出必须被控制器执行。动作可以被离散化,但离散 token 最后仍要还原到连续控制;动作可以直接用 flow/diffusion 生成,但还要匹配本体、频率、延迟、坐标系和安全约束。SA-VLA、Action Prior、SPACE、Rethinking VLA Scaling 都在从不同角度回答同一个问题:动作表征应该以物理状态变化和控制可执行性为中心,而不是以模型输出便利性为中心。

4. 跨本体 scaling 的瓶颈不是数据量,而是数据兼容性

Open X-Embodiment/RT-X 让“跨机器人大数据”成为可能,但 2026 年的证据明显更谨慎。Rethinking VLA Scaling 指出,机器人数据在本体、传感器、频率和动作空间上异构,朴素混合可能负迁移;SPACE 进一步说明 action command 本身不是通用标签。由此可推断:未来的数据 scaling 需要 metadata、action adapter、坐标系统一、质量审计和任务族分层,而不是简单合并轨迹。

5. 接触状态是视觉-语言-动作三元组的盲区

对于开放空间抓放,RGB + language + action 可能足够;但对滑移、柔顺贴合、灵巧抓握、插入和易碎物,关键状态常在视觉遮挡后发生。Transferring Contact, Not Just Motion 的证据说明,motion 潜在状态 或 pose retargeting 不能替代 force/contact 表示。EA-SENSOR 的主题卡判断也支持这一点:视觉负责全局语义和接触前规划,触觉/力负责接触后的局部闭环。

争议与条件

  • Dense ECoT 和 语言—动作 pretraining 都试图增强语言监督,但前者保留视觉并增加 dense reasoning,后者刻意拿掉视觉以避免捷径;两者不是互斥,而是分别处理“语言太粗”和“视觉太强”。
  • Action tokenizer 与 continuous action expert 不是简单优劣关系。tokenizer 方便接入 LLM/VLM 的自回归范式,continuous expert 更贴近控制分布;关键取决于是否保留状态、接触和本体条件。
  • Structured RGB interface 能缓解视觉对齐,但论文也报告失败来自感知噪声、单目深度不稳定和接触不足。因此它适合做 perception-control interface,不应被当作万能替代触觉/力控。
  • 大数据混合只有在动作表示、坐标系、采样频率、质量和任务语义可比时才可能带来正迁移;否则扩数据会扩噪声。

对后续研究的启发

  1. 数据采集应增加阶段级语言、关键动作片段、失败恢复、坐标系、控制频率、接触/力状态等元数据,减少“语言只在轨迹头部出现”的稀疏监督。
  2. 模型结构上可考虑三层接口:语言/任务层负责 goal 与 subtask,结构化视觉层负责 object/geometry/affordance,动作层负责 continuous trajectory、controller adapter 和 safety constraint。
  3. 评测不应只看 open-loop action prediction。需要闭环成功率、action-token 解码误差、时序对齐误差、坐标系错误、模态 dropout、接触遮挡恢复、跨本体 transfer matrix。
  4. 对工业或接触丰富任务,应把触觉/力控从“附加模态”提升为对齐对象:视觉看不见的接触状态必须有独立监督和闭环接口。
  5. 未来 topic-card 可加入一条工作记忆:VLA 对齐难题主要来自信号粒度、物理动作空间和系统闭环接口不一致,而不是单纯缺少更大 VLM。

系统设计含义

把对齐问题视为接口问题后,模型可以分成三层。任务层把语言目标拆为阶段和约束;状态层从稠密视觉中提取对象、几何与可供性;控制层生成连续轨迹,并依据机器人本体、频率和接触状态完成适配。三层之间需要可测的中间变量,而不是只靠端到端损失隐式连接。

这套分层的价值还在于允许逐层证伪。若给定正确阶段标签后动作仍错,问题不在语言稀疏;若给定真实三维状态后恢复,说明视觉接口是瓶颈;若给定正确轨迹仍执行失败,则应检查控制器、坐标系和频率。把这些干预写进实验设计,才能避免把所有增益都归功于更大的主干模型。

数据也应围绕接口补齐。语言不能只出现在轨迹开头,应标出阶段转换和失败原因;视觉应保存关键对象与空间参照;动作要携带坐标系、控制周期和执行状态;接触任务还要记录力与滑移。这样才能区分语义未落地、视觉未对齐和控制器未执行三种不同失败。

评测时可采用逐层替换:给定正确阶段标签测试状态接口,再给定真实状态测试动作生成,最后用正确动作测试控制器。若端到端失败而某一替换显著恢复性能,就能定位瓶颈。相比只比较最终成功率,这种干预更能指导架构与数据修改。

精读复核后的对齐框架

移出无法获得完整正文的论文后,论证仍然指向四个可分开验证的接口。第一,语言需要被翻译为动作相关中间量:稠密具身 CoT 和 ERVLA 都把高层语义与末端运动、图像空间轨迹等表示相连,后者还显示把长 CoT 当动作前缀会累积推理误差。

第二,连续动作不能被当成与状态无关的离散单词。SA-VLA 把问题定位为固定 action token 回解码到连续控制时的状态依赖;SPACE 则表明同一记录命令在不同本体上未必对应同一物理运动。因此,动作对齐必须包含坐标系、本体和时序语义。

第三,稠密视觉应被压缩为结构化、动作可查询的接口。SSI-Policy 将 RGB 的几何和任务 grounding 与运动轨迹分层,减轻了视觉 token 直接淹没控制信号的风险。第四,接触不能继续停留在 RGB 之外:TACO 把未来视频和力序列联合预测,再用失败附近状态生成纠正动作。这些结果共同将“多模态对齐”改写为四个不同频、不同粒度的接口设计问题。

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. ActionReasoning: Robot Action Reasoning in 3D Space with LLM for Robotic Brick StackingGuangming Wang, Qizhen Ying, Yixiong Jing, Olaf Wysocki, Brian Sheil · 2026
  3. 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
  4. Fast-ThinkAct: Efficient Vision-Language-Action Reasoning via Verbalizable Latent PlanningChi-Pin Huang, Yunze Man, Zhiding Yu, Min-Hung Chen, Jan Kautz, Yu-Chiang Frank Wang, Fu-En Yang · 2026
  5. 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
  6. ReCoVLA: VLM-Guided Reward Compilation for Failure Recovery in Vision-Language-Action PoliciesHaodi Hu, Chung-Ta Huang, Jing Liu, Ye Wang, Kei Suzuki, Matthew Brand, Toshiaki Koike-Akino · 2026
  7. Revisiting Embodied Chain-of-Thought for Generalizable Robot ManipulationNan Sun, Yuan Zhang, Yongkun Yang, Wentao Zhao, Peiyan Li, Jun Guo, Wenxuan Song, Pengxiang Ding et al. · 2026
  8. SA-VLA: State-aware tokenizer for improving Vision-Language-Action Models' performanceTengyue Jiang, Chunpu Xu, Jiayue Kang, Yao Mu · 2026
  9. SPACE: Enabling Learning from Cross-Robot Data Toward Generalist PoliciesHaeone Lee · 2026
  10. SSI-Policy: Learning Structured Scene Interfaces for Vision-Language Robotic ManipulationKaijun Wang, Zikai Ouyang, Xuping Wu, Jinyi Hong, Wei Pan, Haibo Lu, Jia Pan, Wei Zhang et al. · 2026
  11. 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
  12. 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
  13. Training Vision-Language-Action Models with Dense Embodied Chain-of-Thought SupervisionHaoyang Li, Guanlin Li, Youhe Feng, Chen Zhao, Zhuoran Wang, Yang Li, Qizhe Wei, Shifeng Bao et al. · 2026
  14. Vision-Language-Action Models: Experimental Insights from a Real-World UR5 PlatformMathilde Hochedel, Marc Lalonde · 2026
  15. World Model for Robot Learning: A Comprehensive SurveyBohan Hou, Gen Li, Jindou Jia, Tuo An, Xinying Guo, Sicong Leng, Haoran Geng, Yanjie Ze et al. · 2026
完整证据附录(15 条正式证据事件)
  • Time range: 2026-01-14..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。

EA-ALIGN-READ-0013

  • Claim: 长程规划、失败后自我纠正、少样本适应属于认知层能力;显式 CoT 能提升它们但推理延迟高,认知处理可以压缩为 latent 推理而保持规划与恢复能力。
  • Stance: support | Confidence: direct
  • Paper: 2601.09708 Fast-ThinkAct: Efficient Vision-Language-Action Reasoning via Verbalizable Latent Planning
  • Locator: 5 Conclusion
  • Evidence: 论文指出 VLA 靠动作监督擅长基本技能,但在长程规划、失败自我纠正、新场景适应上泛化差;Fast-ThinkAct 用 preference-guided 蒸馏把冗长文本推理压缩为紧凑 latent CoT,在保持 long-horizon planning、few-shot adaptation 和 failure recovery 的同时推理延迟最多降 89.3%。
  • Quote: “By distilling lengthy textual reasoning into compact latent representations via preference-guided distillation and visual trajectory alignment, our approach bridges high-level embodied reasoning with low-level action execution through reasoning-enhanced policy learning. Extensive experiments across diverse robotic manipulation and embodied reasoning benchmarks demonstrate that Fast-ThinkAct achieves strong performance with significantly reduced inference latency while enabling effective long-hor”
  • Authors: chi-pin-huang; yunze-man; zhiding-yu; et al.

EA-ALIGN-READ-0014

  • Claim: 纯反应式 VLA 在复杂物理环境中仍受长时程推理、时序归因和误差累积限制,这构成引入显式预测结构的主要动机。
  • Stance: support | Confidence: direct
  • Paper: 2605.00080 World Model for Robot Learning: A Comprehensive Survey
  • Locator: 1 Introduction
  • Evidence: 引言直接将纯反应 VLA 的三类困难列为长时程推理、temporal credit assignment 与 compounding errors。
  • Quote: “Yet despite strong scaling trends ( Xiao et al. , 2025 ; Li et al. , 2025b ; Zhu et al. , 2026 ) , purely reactive VLA policies remain limited in complex physical environments, where they often struggle with long-horizon reasoning, temporal credit assignment, and robustness under compounding errors. A growing body of work argues that these limitations stem not only from insufficient action prediction capacity ( Ye et al. , 2026b ; Dang et al. , 2026 ) , but also from the lack of explicit predict”
  • Authors: bohan-hou; gen-li; jindou-jia; 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-0015

  • Claim: 失败恢复可以把认知层(failure mode/recovery stage 判断与 reward 选择)与控制层(residual 纠正)分开;VLM 的失败分类错误由此成为与感知误差独立的认知误差来源。
  • Stance: support | Confidence: direct
  • Paper: 2606.09630 ReCoVLA: VLM-Guided Reward Compilation for Failure Recovery in Vision-Language-Action Policies
  • Locator: 1 Introduction
  • Evidence: ReCoVLA 用外部 VLM 只推断 failure type、recovery stage、active entities、confidence 和 reward mask,不直接生成动作;确定性 reward compiler 做实体 grounding 与 stage gates,residual policy 在冻结 VLA latents 上学纠正。Limitations 明确列出 VLM failure-classification mistakes 与 perception errors、sim-to-real mismatch 并列为失败来源。
  • Quote: “Instead, it produces a structured recovery descriptor containing the failure type, recovery stage, active entities, confidence, and reward mask.”
  • Authors: haodi-hu; chung-ta-huang; jing-liu; et al.

EA-ALIGN-READ-0005

  • Claim: Cross-embodiment VLA alignment is difficult partly because shared high-level task cognition must be connected to platform-specific low-level state and action spaces.
  • Stance: support | Confidence: direct
  • Paper: 2606.30552 Training Vision-Language-Action Models with Dense Embodied Chain-of-Thought Supervision
  • Locator: Abstract (full-text section)
  • Evidence: The paper frames low-level state/action heterogeneity as a core cross-embodiment challenge, then uses dense embodied chain-of-thought supervision in the VLM stream and a flow-matching action expert that outputs continuous action chunks.
  • Quote: “Abstract Cross-embodiment transfer in vision-language-action (VLA) models remains challenging because low-level state and action spaces differ fundamentally across robot platforms. We observe that the high-level cognitive process underlying manipulation, including scene perception, object identification, task planning, and sub-task decomposition, is largely shared across embodiments. Based on this observation, we present ZR-0 , a 2.6 billion parameter end-to-end VLA model that uses dense Embodie”
  • Authors: haoyang-li; guanlin-li; youhe-feng; et al.

EA-ALIGN-READ-0010

  • Claim: ActionReasoning假设感知已由视觉算法可靠提供,将 LLM 的任务收窄为 3D 动作推理;作者认为这种解耦可降低端到端训练的数据需求。
  • Stance: conditional | Confidence: direct
  • Paper: 2602.21161 ActionReasoning: Robot Action Reasoning in 3D Space with LLM for Robotic Brick Stacking
  • Locator: II-B LLM/VLM Based Robotic Operation
  • Evidence: 相关工作段明确提出解耦视觉部件,让 LLM 在已知感知状态上做 3D 物理与动作推理。
  • Quote: “To address this, we argue for decoupling the visual component: assuming the robot already has sufficiently accurate perceptual information via computer vision algorithms, and the LLMs are asked to focus only on action reasoning. This approach significantly reduces data requirements while also leveraging the wealth of existing research in computer vision. For example, the ReKep series [ 11 , 16 ] use vision-language models to identify keypoints for robotic manipulation, which significantly reduce”
  • Authors: guangming-wang; qizhen-ying; yixiong-jing; et al.

EA-ALIGN-READ-0011

  • 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.

EA-ALIGN-READ-0006

  • Claim: ERVLA 的大规模实验表明,有效的具身 CoT 需要落到末端执行器运动或图像空间轨迹等动作相关表示,而将显式 CoT 作为自回归动作前缀会在推理时累积误差。
  • Stance: conditional | Confidence: direct
  • Paper: 2606.03784 Revisiting Embodied Chain-of-Thought for Generalizable Robot Manipulation
  • Locator: Abstract (full-text section)
  • Evidence: 摘要同时给出了动作相关 grounding 的有效性与 autoregressive action prefix 的 compounding-error 限制。
  • Quote: “Abstract Embodied chain-of-thought (CoT) aims to bridge linguistic reasoning with robotic control, yet its effective form and integration remain underexplored. In this paper, we revisit embodied CoT for robotic control at an unprecedented scale. We curate the largest embodied CoT corpus to date, comprising 978,743 trajectories, 226.3M samples, and 2592.5 hours of data. Through extensive experiments, we show that effective CoT must ground high-level semantic understaning in concrete linguistic ac”
  • Authors: nan-sun; yuan-zhang; yongkun-yang; et al.

EA-ALIGN-READ-0007

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

EA-ALIGN-READ-0008

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

EA-ALIGN-READ-0002

  • Claim: A structured intermediate visual interface can reduce the alignment burden by separating RGB-based geometric/task grounding from embodiment-specific action control.
  • Stance: conditional | Confidence: direct
  • Paper: 2606.26800 SSI-Policy: Learning Structured Scene Interfaces for Vision-Language Robotic Manipulation
  • Locator: Abstract (full-text section)
  • Evidence: SSI-Policy builds an RGB-only structured scene interface encoding monocular depth features, language-grounded layouts, and instruction-conditioned 2D motion trajectories; it reports few-shot gains but notes failures from perception noise and contact limitations.
  • Quote: “Abstract Real-world robotic manipulation demands spatial grounding, task-aware reasoning, and precise control. Learning such capabilities becomes particularly challenging in the low-data regime. Prior methods often trade off scalable task-level reasoning and explicit physical structure: video-based approaches can drift geometrically over long horizons, 3D approaches often require depth sensing, and many flow/trajectory interfaces emphasize motion without an explicit RGB-only geometric representa”
  • Authors: kaijun-wang; zikai-ouyang; xuping-wu; 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-ALIGN-READ-0003

  • Claim: Discrete action tokenization is a compact interface for autoregressive VLA policies, but decoding fixed tokens back into continuous robot controls is a bottleneck when the same token must mean different controls under different robot states and contacts.
  • Stance: limit | Confidence: direct
  • Paper: 2606.30113 SA-VLA: State-aware tokenizer for improving Vision-Language-Action Models' performance
  • Locator: Abstract (full-text section)
  • Evidence: SA-VLA conditions action-token decoding on proprioceptive state via adapters or cross-attention, reporting improved RoboTwin and zero-shot sim-to-real success over tokenizer baselines.
  • Quote: “Abstract Discrete action tokenization provides a compact interface for autoregressive VLA policies, but accurately recovering continuous robot actions from discrete codes remains challenging. Existing tokenizers typically map each discrete code to a fixed continuous action prototype, ignoring the robot’s current proprioceptive state. This limitation is particularly pronounced in manipulation, where the same action token may require different continuous controls under different joint configuratio”
  • Authors: tengyue-jiang; chunpu-xu; jiayue-kang; et al.

EA-ALIGN-READ-0004

  • Claim: Offline VLA indicators can fail to transfer to stable real-robot behavior when action semantics, coordinate frames, temporal modality alignment, image preprocessing, and dataset coverage are not controlled together.
  • Stance: limit | Confidence: direct
  • Paper: 2606.30456 Vision-Language-Action Models: Experimental Insights from a Real-World UR5 Platform
  • Locator: Abstract (full-text section)
  • Evidence: The UR5 study reports a gap between offline indicators and unstable closed-loop physical behavior, attributing it to data-model-control pipeline consistency rather than model capacity alone.
  • Quote: “Instead, it is strongly influenced by a combination of factors, including action semantics, coordinate frame conventions, temporal alignment between modalities, image preprocessing consistency, and dataset coverage and quality. These observations lead to a key interpretation: the successful deployment of VLA systems in real-world settings depends less on incremental improvements in model capacity and more on precise control of the entire data–model–control pipeline.”
  • Authors: mathilde-hochedel; marc-lalonde

EA-ALIGN-READ-0009

  • 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。