世界模型与评测 · World Models and Evaluation

ACT及动作分块策略在RoboTwin 2.0中的接入、训练与闭环评测

ACT and Action Chunking in RoboTwin 2.0

更新于

30 秒结论

先读这个判断

本备忘录关注的不是“怎样把一个 ACT 检查点跑起来”,而是更完整的工程问题:动作分块策略进入 RoboTwin 2.0 后,怎样定义动作表示、组织多任务训练、选择块内执行长度,并用闭环实验区分模型能力与执行器调度能力。检索时间为 2026 年 2 月 7 日至 8 月 7 日;四轮检索获得 228 篇去重候选,恢复 39 篇完整非 OCR 正文,最终深读并审计 15 篇。它们覆盖 ACT 变体、扩散或流匹配动作头、自适应执行视界、世界动作模型与双臂基准。多数工作仍是 2026 年预印本,跨论文的成功率不能直接横比…

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

研究边界与证据范围

本备忘录关注的不是“怎样把一个 ACT 检查点跑起来”,而是更完整的工程问题:动作分块策略进入 RoboTwin 2.0 后,怎样定义动作表示、组织多任务训练、选择块内执行长度,并用闭环实验区分模型能力与执行器调度能力。检索时间为 2026 年 2 月 7 日至 8 月 7 日;四轮检索获得 228 篇去重候选,恢复 39 篇完整非 OCR 正文,最终深读并审计 15 篇。它们覆盖 ACT 变体、扩散或流匹配动作头、自适应执行视界、世界动作模型与双臂基准。多数工作仍是 2026 年预印本,跨论文的成功率不能直接横比。

中心判断

ACT 接入 RoboTwin 2.0 的核心不在于选定一个 chunk_size,而在于拆开三个原本容易混淆的时间尺度:模型一次预测多长的动作序列、控制器一次真正执行多少步、环境在何时提供新观测并触发重规划。预测视界决定学习目标,执行视界决定开环暴露长度,重规划频率决定反馈带宽。把三者绑定成同一个常数,虽然实现简单,却会把任务阶段、接触风险和推理成本都藏进一个经验超参数。

因此,较稳妥的项目方案应采用“两层基线、三类增强、四组对照”。两层基线是原始 ACT 训练基线与固定视界闭环基线;三类增强分别针对多任务表示、执行时机和跨块状态;四组对照则同时覆盖干净与随机场景、固定与自适应视界、等推理预算与等执行步数、仿真与真实机器人。这样才能判断增益来自更好的动作生成,还是来自更频繁地重新调用策略。

一、接入层:先统一动作与数据接口,再讨论模型结构

RoboTwin 2.0 的双臂任务要求策略同时处理多视角图像、本体状态、语言任务以及左右臂协调动作。最小可复现实验应先冻结一套明确的数据契约:观测视角及时间同步、关节或末端位姿表示、左右臂和夹爪维度顺序、动作归一化范围、控制频率、episode 截断规则以及成功判定。若这些接口在训练和评测间发生漂移,模型比较会被坐标系或执行器差异污染。

MoE-ACT 给出了最接近“ACT 多任务化”的直接证据:它保留动作分块解码框架,在编码器中加入稀疏专家路由,并用语言条件调制和多尺度交叉注意力减少任务间动作分布缠结。这提示项目可以把原始 ACT 作为可解释基线,再逐项加入任务条件、专家路由和视觉融合,而不是一开始就更换整个策略族。与此同时,Action-Geometry Prediction 表明,双臂动作块还可以和未来三维几何潜变量联合预测;这种设计适合检验“只靠二维图像特征是否足以处理遮挡、接触和双臂协同”。

训练目标也不必局限于对单条专家动作做点估计。VQActFlow 将动作块编码为离散代码,使语言指令对应的任务模式和场景可行性能够分别引导;Set-Supervised Diffusion Policy 则把人工纠错扩展成动作块的期望集合,而不是把一次纠错当作唯一正确标签。两者共同说明,多任务失败可能来自“动作模式选择错误”和“监督目标过窄”,这两种问题不应都归因于主干容量不足。

二、执行层:动作块长度不等于执行视界

最强的一致信号来自执行时机研究。固定执行视界让每个动作块在若干步内开环运行:视界太短会频繁打断连贯动作并增加计算,视界太长则会在接触、对齐和插入阶段继续执行过时预测。PACE 在 RoboTwin 2.0 上观察到,成功率随执行视界变化具有任务依赖和非单调性;它使用动作速度曲线的低速谷值寻找阶段边界。这意味着项目不应只报告一个人工选出的最佳视界,而应至少做视界扫描并公开任务级曲线。

自适应方法可以按信号来源分为三组。第一组直接读动作分布:AAC 用动作熵决定块长,DVAC 用流匹配末端去噪过程中干净动作估计的方差识别不稳定未来动作。第二组学习一个调度器:DEHP 在冻结基础策略的前提下用在线强化学习训练视界预测头,BCP 把选择分解为有序的“继续或重规划”决策,并把成功和调用效率同时写入奖励。第三组验证预测世界与真实世界是否仍一致:FFDC-WAM 在执行过程中比较预测视觉动态、动作与新观测,在想象失真时提前重规划。

这些方法不应被简单排成单一排行榜。基于速度或去噪方差的规则训练成本低,但信号只适用于特定策略结构;学习式调度器更灵活,却需要在线交互和额外训练;未来—现实验证能直接利用世界模型,却受视觉预测质量影响。合理的工程顺序是先建立固定视界扫描,再加入无需训练的调度方法,最后才评估强化学习或世界模型调度器是否值得其额外复杂度。

三、跨块状态:闭环不只是更频繁地看图

缩短执行视界只能更早获得新观测,不能保证策略知道自己的动作怎样改变了场景。EvoScene-VLA 为每个动作块维护由动作更新、再由新观测校正的循环场景前缀;AHA-WAM 则把低频视频规划器和高频动作执行器拆开,让短动作块复用长时世界上下文。这两项工作指向同一结构:慢速场景推理和快速闭环控制可以异步协作,不必共享一个时间分辨率。

另一条路线是把未来预测保留在训练侧。GigaWorld-Policy 联合学习动作与未来视频,但通过因果掩码使部署时可以只解码动作;DREAM-Chunk 则在测试时采样多个动作块,用轻量潜在世界模型筛选与真实观测最一致的候选。前者优先降低在线时延,后者用额外计算换取随机扰动下的块内反应性。项目需要明确选择的是“训练期辅助监督”“跨块记忆”还是“测试时多候选搜索”,因为三者的算力位置和失败方式不同。

四、闭环评测:把模型、调度器和计算预算分开

建议把 RoboTwin 2.0 评测矩阵分成五个轴。第一,场景轴同时报告干净和随机设置,并保留每任务结果,避免平均分掩盖少数高风险任务。第二,时间轴扫描预测视界与执行视界,至少包含单步、短块、完整块和一个自适应方案。第三,预算轴分别做等策略调用次数、等墙钟时间和等执行步数比较,防止“更高成功率”只是更多推理。第四,扰动轴加入动作噪声、视觉延迟、遮挡和物体位姿偏移,观察重规划是否真的发生在反馈敏感阶段。第五,迁移轴在同一控制接口下复核少量实机双臂任务,记录失败类型而不只记录成功率。

最关键的消融应围绕因果拆分:基础策略冻结时只改变执行调度;调度固定时只改变动作表示;主干、数据和训练轮数保持一致时再比较结构模块。对自适应方法还应报告执行视界分布、策略调用次数、控制频率和阶段级可视化。否则,很难区分“更聪明地决定何时看一眼”与“只是看得更勤”。

条件、限制与未决问题

现有证据存在四项共同边界。其一,许多结果使用不同预训练数据、基础策略和任务子集,数值并非同一赛道。其二,速度、动作熵、去噪方差和预测—观测误差都只是风险代理,没有哪个信号被证明在所有任务上可靠。其三,学习式视界调度依赖在线强化学习,而冻结基础策略会限制性能上限。其四,世界模型和多候选搜索会改变算力、延迟与显存位置,仿真收益未必能在边缘设备复现。

还有一个更根本的空白:当前多数工作优化“何时重规划”,但很少统一处理重规划之后如何连续拼接新旧动作块、如何把安全约束写入执行器、以及如何在任务阶段发生误识别时恢复。动作块调度应被视为闭环控制系统的一部分,而不是独立的推理技巧。

结论与项目启发

对本项目最可执行的路线是:先把原始 ACT 做成数据、动作和评测接口都稳定的 RoboTwin 2.0 基线;随后加入语言条件和多任务解耦,建立固定视界扫描;再以 PACE 或 DVAC 这类无需重训的方案验证自适应执行价值;最后根据失败集中在“场景状态遗失”还是“动作候选不足”,选择循环场景状态、异步世界规划或测试时多候选。每次增强都必须在冻结其他变量的闭环对照中验证。

由此得到的判断不是“ACT 已经过时”,而是 ACT 只定义了动作块如何生成,尚未完整定义动作块如何被信任、执行、打断和评测。RoboTwin 2.0 的价值,正是把这几层放进同一双臂闭环系统中逐层检验。

References

  • MoE-ACT:ACT 的语言条件多任务稀疏专家改造。
  • Action-Geometry Prediction:动作块与未来三维几何联合预测。
  • PACE:基于运动相位的测试时执行视界选择。
  • AAC:以动作熵调节块长。
  • DVAC:以去噪方差触发重规划。
  • DEHP:冻结策略上的强化学习视界预测。
  • BCP:有序继续/重规划策略。
  • FFDC-WAM:基于未来—现实一致性的自适应执行。
  • AHA-WAM:异步世界规划与高频动作执行。
  • EvoScene-VLA:跨块循环场景状态。
  • DREAM-Chunk:潜在世界模型支持的测试时多候选执行。
  • GigaWorld-Policy:训练期视频监督与可选视频解码。

Accepted evidence

去重论文引用

  1. Action-Geometry Prediction with 3D Geometric Prior for Bimanual ManipulationChongyang Xu, Haipeng Li, Shen Cheng, Jingyu Hu, Haoqiang Fan, Ziliang Feng, Shuaicheng Liu · 2026
  2. Adaptive Action Chunking at Inference-time for Vision-Language-Action ModelsYuanchang Liang, Xiaobo Wang, Kai Wang, Shuo Wang, Xiaojiang Peng, Haoyu Chen, David Kim Huat Chua, Prahlad Vadakkepat · 2026
  3. AHA-WAM:Asynchronous Horizon-Adaptive World-Action Modeling with Observation-Guided Context RoutingJisong Cai, Long Ling, Shiwei Chu, Zhongshan Liu, Jiayue Kang, Zhixuan Liang, Wenjie Xu, Yinan Mao et al. · 2026
  4. Continue or Replan? Bernoulli-Continuation Policy Learning for Adaptive Horizon ExecutionWeichen Xu, Zhenhua Liu, Lin Luo, Yaobo Liang, Chengtang Yao, Qingyu Mei, Jian Cao, Xixin Cao et al. · 2026
  5. Denoising Tells When to Replan: Denoising-Variance Adaptive Chunking for Flow-Based Robot PoliciesXiangdong Feng, Yuxuan Cheng, Chen Shi, Boyao Han, Yuxuan Yan, Yitong Hong, Zhuotao Tian, Li Jiang · 2026
  6. DREAM-Chunk: Reactive Action Chunking with Latent World ModelWenxi Chen, Kaidi Zhang, Chi Lin, Zhiyuan Zhang, Yu She, Yuejiang Liu, Raymond A. Yeh, Shaoshuai Mou et al. · 2026
  7. Dynamic Execution Horizon Prediction for Chunk-based Robot PoliciesYuchi Zhao, Miroslav Bogdanovic, Arjun Sohal, Liyu Tao, Kourosh Darvish, Alán Aspuru-Guzik, Florian Shkurti, Animesh Garg · 2026
  8. EvoScene-VLA: Evolving Scene Beliefs Inside the Action Decoder for Chunked Robot ControlChushan Zhang, Ruihan Lu, Jinguang Tong, Xuesong Li, Yikai Wang, Hongdong Li · 2026
  9. FabriVLA: A Lightweight Vision-Language-Action Model with Conformal Action Chunk UncertaintyShiyuan Yang, Borong Zhang, Jizheng Zhang, Zhijia Tao, Junfei Guo, Donglai Ran, Xu Bian, Qingbiao Li · 2026
  10. GigaWorld-Policy: An Efficient Action-Centered World--Action ModelAngen Ye, Boyuan Wang, Chaojun Ni, Guan Huang, Guosheng Zhao, Hao Li, Hengtao Li, Jie Li et al. · 2026
  11. MoE-ACT: Scaling Multi-Task Bimanual Manipulation with Sparse Language-Conditioned Mixture-of-Experts TransformersKangjun Guo, Haichao Liu, Yanji Sun, Ruhan Zhao, Jinni Zhou, Jun Ma · 2026
  12. PACE: Phase-Aware Chunk Execution for Robot Policies with Action ChunkingJunnan Nie, Jiayi Li, Jiachen Zhang, Junyi Lao, Chenghao Liu, Tianle Zhang, Liang Lin, Songfang Huang · 2026
  13. Set-Supervised Diffusion Policy: Learning Action-Chunking Diffusion through CorrectionsZhaoting Li, Gang Chen, Javier Alonso-Mora, Cosimo Della Santina, Jens Kober · 2026
  14. VQActFlow: Vector-Quantized Action Mode Steering for Multi-Task Robot ManipulationZhigen Zhao, Mark Leggiero, Yipu Chen, Haoran Liu, Yifan Wu, Huishu Xue, Sirui Zhan, Ye Zhao · 2026
  15. When to Trust Imagination: Adaptive Action Execution for World Action ModelsRui Wang, Yue Zhang, Jiehong Lin, Kuncheng Luo, Jianan Wang, Zhongrui Wang, Xiaojuan Qi · 2026
完整证据附录(15 条正式证据事件)
  • Time range: 2026-02-07..2026-08-07
  • 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-ACTRT-2026-0003

  • Claim: 双臂动作分块可以与未来几何预测联合训练:策略融合 3D 几何潜变量、2D 语义和本体状态,并由扩散模型同时预测未来动作块与可解码的未来 3D 场景潜变量。
  • Stance: support | Confidence: direct
  • Paper: 2602.23814 Action-Geometry Prediction with 3D Geometric Prior for Bimanual Manipulation
  • Locator: Abstract (full-text section)
  • Evidence: 该方法把预训练 3D 几何基础模型作为 RGB 输入的空间先验,使用统一状态表示联合生成动作块和未来稠密点图,使动作学习显式感知双臂协调与场景几何演化。
  • Quote: “Abstract Bimanual manipulation requires policies that can reason about 3D geometry, anticipate how it evolves under action, and generate smooth, coordinated motions. However, existing methods typically rely on 2D features with limited spatial awareness, or require explicit point clouds that are difficult to obtain reliably in real-world settings. At the same time, recent 3D geometric foundation models show that accurate and diverse 3D structure can be reconstructed directly from RGB images in a”
  • Authors: chongyang-xu; haipeng-li; shen-cheng; et al.

EA-ACTRT-2026-0002

  • Claim: 把 ACT 接入多任务双臂场景时,可在编码器加入稀疏专家路由,并用语言条件 FiLM 与多尺度交叉注意力约束动作解码,从而降低不同任务动作分布的相互干扰。
  • Stance: support | Confidence: direct
  • Paper: 2603.15265 MoE-ACT: Scaling Multi-Task Bimanual Manipulation with Sparse Language-Conditioned Mixture-of-Experts Transformers
  • Locator: Abstract (full-text section)
  • Evidence: MoE-ACT 保留 ACT 的动作分块骨架,以稀疏 MoE 分解多任务潜在动作分布,再用语言条件调制和多尺度视觉融合生成动作序列,并在 RoboTwin 2.0 与真实双臂设置中评测。
  • Quote: “Abstract The ability of robots to handle multiple tasks under a unified policy is critical for deploying embodied intelligence in real-world household and industrial applications. However, out-of-distribution variation across tasks often causes severe task interference and negative transfer when training general robotic policies. To address this challenge, we propose a lightweight multi-task imitation learning framework for bimanual manipulation, termed Mixture-of-Experts-Enhanced Action Chunkin”
  • Authors: kangjun-guo; haichao-liu; yanji-sun; et al.

EA-ACTRT-2026-0005

  • Claim: 动作块越长并不总是越好:长块降低闭环反应性,短块增加跨块模式跳变;AAC 用当前预测的动作熵在线调节执行块长度,以平衡一致性与反应性。
  • Stance: support | Confidence: direct
  • Paper: 2604.04161 Adaptive Action Chunking at Inference-time for Vision-Language-Action Models
  • Locator: Abstract (full-text section)
  • Evidence: AAC 针对带扩散动作头的 VLA 并行采样候选动作,用动作熵作为不确定性线索,在高不确定阶段缩短块、在低不确定阶段延长块。
  • Quote: “Abstract In Vision-Language-Action (VLA) models, action chunking ( i.e. , executing a sequence of actions without intermediate replanning) is a key technique to improve robotic manipulation abilities. However, a large chunk size reduces the model’s responsiveness to new information, while a small one increases the likelihood of mode-jumping, jerky behavior resulting from discontinuities between chunks. Therefore, selecting the optimal chunk size is an urgent demand to balance the model’s reactiv”
  • Authors: yuanchang-liang; xiaobo-wang; kai-wang; et al.

EA-ACTRT-2026-0006

  • Claim: 对流匹配动作策略,末端去噪步骤中干净动作估计的方差可作为重规划信号:DVAC 执行低方差稳定前缀,并在高方差未来动作被提交前重新观测与规划。
  • Stance: support | Confidence: direct
  • Paper: 2606.03847 Denoising Tells When to Replan: Denoising-Variance Adaptive Chunking for Flow-Based Robot Policies
  • Locator: Abstract (full-text section)
  • Evidence: DVAC 不增加训练模块,而是利用推理过程已有的去噪轨迹估计未来动作稳定性,并用局部滚动尺度校准阈值,在 RoboTwin、LIBERO、CALVIN 与实机任务中比较成功率和重规划开销。
  • Quote: “Abstract Action chunking has become a common inference strategy for flow-based robot policies, improving action coherence by modeling multi-step temporal dependencies in demonstrations. However, the execution horizon is still typically set as an empirical fixed value, overlooking that predictable free-space motions and precision-critical interaction phases often require different replanning frequencies. In this work, we first show that the denoising process of flow-based policies contains an int”
  • Authors: xiangdong-feng; yuxuan-cheng; chen-shi; et al.

EA-ACTRT-2026-0007

  • Claim: 世界预测与动作执行不必保持同一时间分辨率;AHA-WAM 将低频视频规划器与高频短动作块执行器解耦,并通过观测引导的上下文路由在闭环执行时复用长时世界上下文。
  • Stance: support | Confidence: direct
  • Paper: 2606.09811 AHA-WAM:Asynchronous Horizon-Adaptive World-Action Modeling with Observation-Guided Context Routing
  • Locator: Abstract (full-text section)
  • Evidence: 双 DiT 架构让视频分支维护滚动记忆,动作分支以较高频率查询该上下文并输出短动作块;视界偏移训练支持规划器与执行器不同步。
  • Quote: “Abstract World-action models have emerged as a promising paradigm for robot manipulation, jointly modeling visual scene dynamics and actions to inject physical priors into policy learning. However, existing world-action models couple world prediction and action execution at the same temporal resolution, forcing the world branch to model near-term frame variations that are redundant and weakly informative. We posit that strictly binding world prediction and action execution to the same temporal r”
  • Authors: jisong-cai; long-ling; shiwei-chu; et al.

EA-ACTRT-2026-0012

  • Claim: 固定执行视界会把重规划变成与任务阶段无关的周期调度;BCP 把视界选择分解为有序的继续/重规划决策,并用轨迹级强化学习同时优化成功与 VLA 调用效率。
  • Stance: support | Confidence: direct
  • Paper: 2608.03483 Continue or Replan? Bernoulli-Continuation Policy Learning for Adaptive Horizon Execution
  • Locator: Abstract (full-text section)
  • Evidence: BCP 在冻结 VLA 上附加轻量 continuation head,复用动作块与视觉语言表征;其有序前缀结构与效率奖励避免把所有候选视界当成互不相关类别或退化为频繁重规划。
  • Quote: “Abstract Existing chunk-based Vision-Language-Action (VLA) models execute a fixed number of actions (i.e., execution horizon) before replanning, turning replanning into a task-agnostic periodic schedule that is independent of task progress. As a result, when no replanning boundary falls before a critical manipulation stage, it is executed from a stale chunk rather than a freshly replanned one. To address this limitation, we propose B ernoulli- C ontinuation P olicy ( BCP ), a lightweight, plug-a”
  • Authors: weichen-xu; zhenhua-liu; lin-luo; et al.

EA-ACTRT-2026-0015

  • Claim: 世界动作模型可以在训练时联合学习动作块与未来视频、在部署时只解码动作:GigaWorld-Policy 用因果掩码阻止未来视频 token 反向影响动作 token,使视频生成成为可选推理分支。
  • Stance: conditional | Confidence: direct
  • Paper: 2603.17240 GigaWorld-Policy: An Efficient Action-Centered World--Action Model
  • Locator: Abstract (full-text section)
  • Evidence: 该动作中心 WAM 以动作预测和视觉动力学共同监督共享模型,同时通过因果设计保留低时延动作解码路径,并在 RoboTwin 2.0 与真实机器人上评估。
  • Quote: “Abstract World–Action Models (WAM) initialized from pre-trained video generation backbones have demonstrated remarkable potential for robot policy learning. However, existing approaches face two critical bottlenecks that hinder performance and deployment. First, jointly reasoning over future visual dynamics and corresponding actions incurs substantial inference overhead. Second, joint modeling often entangles visual and motion representations, making motion prediction accuracy heavily dependent”
  • Authors: angen-ye; boyuan-wang; chaojun-ni; et al.

EA-ACTRT-2026-0013

  • Claim: 世界动作模型的执行长度可以由“想象—现实一致性”决定:FFDC 联合读取预测动作、预测视觉动态、真实观测和语言,继续执行可信片段并在偏差出现时提前重规划。
  • Stance: conditional | Confidence: direct
  • Paper: 2605.06222 When to Trust Imagination: Adaptive Action Execution for World Action Models
  • Locator: Abstract (full-text section)
  • Evidence: FFDC 将自适应执行表述为未来—现实验证问题,使长视界效率与接触阶段反应性不再依赖一个固定块长度,并在 RoboTwin 与实机中同时评估成功、调用次数和时延。
  • Quote: “Abstract World Action Models (WAMs) have recently emerged as a promising paradigm for robotic manipulation by jointly predicting future visual observations and future actions. However, current WAMs typically execute a fixed number of predicted actions after each model inference, leaving the robot blind to whether the imagined future remains consistent with the actual physical rollout. In this work, we formulate adaptive WAM execution as a future–reality verification problem: the robot should exe”
  • Authors: rui-wang; yue-zhang; jiehong-lin; et al.

EA-ACTRT-2026-0014

  • Claim: 跨动作块闭环控制需要保留由机器人自身动作更新的场景状态;EvoScene-VLA 让动作解码器同时输出动作块和紧凑场景更新,并在下一次视觉调用中用新观测校正这一先验。
  • Stance: conditional | Confidence: direct
  • Paper: 2605.21862 EvoScene-VLA: Evolving Scene Beliefs Inside the Action Decoder for Chunked Robot Control
  • Locator: Abstract (full-text section)
  • Evidence: 循环场景前缀把上一块动作导致的接触、遮挡和物体移动带入下一控制调用;训练时用几何锚点与未来场景目标监督,部署时移除教师模块。
  • Quote: “Abstract Chunked vision-language-action (VLA) policies predict multi-step robot controls, conditioning each update on the current visual observation alone. Yet robot actions cause contact, occlusion, and object motion, and the geometry that later decisions depend on can change before the next visual update arrives. Spatial VLAs improve current-frame geometry. Temporal VLAs aggregate past frames. Neither maintains an action-updated scene prior across chunks. We argue for a persistent action-updat”
  • Authors: chushan-zhang; ruihan-lu; jinguang-tong; et al.

EA-ACTRT-2026-0010

  • Claim: 交互纠错可从单步监督扩展到动作块集合监督:Set-Supervised Diffusion Policy 将正负动作块逐时刻构造成期望动作集合,再训练扩散策略生成落入该集合的动作块。
  • Stance: conditional | Confidence: direct
  • Paper: 2606.01865 Set-Supervised Diffusion Policy: Learning Action-Chunking Diffusion through Corrections
  • Locator: IV-B Extending Desired Action Set to Action-Chunks
  • Evidence: 该方法不把一次人工纠错当作唯一精确标签,而是由正负动作对定义可接受动作区域,并用多次纠错逐步收紧动作块监督。
  • Quote: “CLIC [ 19 ] shows that imitating a set of desired actions, constructed by a pair of positive and negative actions, achieves more robust behavior than imitating only the positive action. Building on this idea, we generalize the single-step desired action set to action-chunks. At step , under observation , we denote a positive and negative action-chunk as Given and the preceding positive actions , the action is better than for all . As we consider demonstration data or intervention feedback, to ma”
  • Authors: zhaoting-li; gang-chen; javier-alonso-mora; et al.

EA-ACTRT-2026-0009

  • Claim: 执行视界可以由冻结动作策略之外的轻量模块学习:DEHP 用在线强化学习根据当前观测和预测动作块决定何时重规划,并在精细阶段缩短视界、自由空间运动中延长视界。
  • Stance: conditional | Confidence: direct
  • Paper: 2606.11408 Dynamic Execution Horizon Prediction for Chunk-based Robot Policies
  • Locator: Abstract (full-text section)
  • Evidence: DEHP 将动作生成与执行时机分离,只训练视界预测分支,从而能在不改基础行为克隆策略权重的情况下检验动态重规划的增益。
  • Quote: “Abstract Action chunking has become a standard design in modern robot policies, from diffusion/flow policies to vision-language-action models, where the policy predicts a sequence of actions and executes a fixed number of them instead of acting one step at a time. However, this paradigm relies on a key assumption: a fixed execution horizon. During chunk execution, the policy operates open-loop, which is particularly problematic for fine-grained manipulation tasks that require frequent replanning”
  • Authors: yuchi-zhao; miroslav-bogdanovic; arjun-sohal; et al.

EA-ACTRT-2026-0011

  • Claim: 长动作块的开环脆弱性可以在块内缓解:DREAM-Chunk 并行采样多个候选动作块,用轻量潜在世界模型预测各自未来,并依据预测潜状态与真实观测的匹配在线选择动作。
  • Stance: conditional | Confidence: direct
  • Paper: 2606.18589 DREAM-Chunk: Reactive Action Chunking with Latent World Model
  • Locator: Abstract (full-text section)
  • Evidence: DREAM-Chunk 不微调基础策略,而用额外测试时计算维持多个可能轨迹,使机器人在随机动力学、执行误差和部分可观测条件下能够块内反应。
  • Quote: “Abstract Action chunking has become a common interface for vision-language-action (VLA) models, enabling low-frequency policy inference to drive high-frequency robot execution. However, once a chunk is committed, its open-loop execution can be brittle under stochastic dynamics, hardware execution errors, and partially observed state. We propose DREAM-Chunk , a test-time scaling method that augments chunking-based policies with a lightweight latent world model without requiring additional policy”
  • Authors: wenxi-chen; kaidi-zhang; chi-lin; et al.

EA-ACTRT-2026-0004

  • Claim: 多任务动作分块不必只在连续空间直接回归;VQActFlow 先把动作块编码为离散代码,再用变分流匹配生成代码序列,使语言任务模式与场景可行性可以在推理时分别施加引导。
  • Stance: conditional | Confidence: direct
  • Paper: 2606.21600 VQActFlow: Vector-Quantized Action Mode Steering for Multi-Task Robot Manipulation
  • Locator: Abstract (full-text section)
  • Evidence: 离散代码簿显式分离多模态动作模式,语言条件的 classifier-free guidance 选择任务模式,代码簿 critic 补充场景可行性信号。
  • Quote: “Abstract Multi-task robot manipulation policies are challenging to learn from demonstration because traditionally a single network must select among qualitatively different action modes from a multimodal demonstration distribution, conditioned on language and visual context. A wrong mode selection means executing the wrong task or an action infeasible in the scene. Tokenizing continuous actions into a learned discrete codebook separates these modes at the representation level, offering structura”
  • Authors: zhigen-zhao; mark-leggiero; yipu-chen; et al.

EA-ACTRT-2026-0008

  • Claim: 动作块的不确定性可以按“整个将执行前缀”联合校准,而非只估计单步误差;JCAC 为冻结策略增加轻量残差尺度头并输出具有用户指定覆盖水平的动作块集合。
  • Stance: conditional | Confidence: direct
  • Paper: 2607.08575 FabriVLA: A Lightweight Vision-Language-Action Model with Conformal Action Chunk Uncertainty
  • Locator: Abstract (full-text section)
  • Evidence: FabriVLA 的 JCAC 使用单次策略查询生成联合共形集合,可在动作执行前对整段前缀进行风险排序,为闭环部署增加可审计的不确定性接口。
  • Quote: “Abstract Vision-Language-Action (VLA) models have become a leading paradigm for general purpose robotic manipulation, but their computational cost and limited uncertainty awareness hinder practical deployment. We present FabriVLA, a lightweight VLA that fuses shallow and intermediate VLM layers to preserve fine-grained visual features, and gates self-attention among action tokens so that its flow matching head admits inter step structure only as far as training warrants. Trained end-to-end in a”
  • Authors: shiyuan-yang; borong-zhang; jizheng-zhang; et al.

EA-ACTRT-2026-0001

  • Claim: 动作分块策略的预测长度与实际执行长度应分开配置;在 RoboTwin 2.0 等闭环任务中,固定执行视界具有任务依赖且非单调的风险,PACE 以动作速度曲线中的低速相位边界在线选择重规划点。
  • Stance: limit | Confidence: direct
  • Paper: 2606.00537 PACE: Phase-Aware Chunk Execution for Robot Policies with Action Chunking
  • Locator: Abstract (full-text section)
  • Evidence: PACE 只读取策略已经生成的动作块,以低速谷值标记阶段切换,在连贯运动中保持较长执行、在接触或阶段转换前更早重规划,不修改也不重训基础策略。
  • Quote: “Abstract Recent vision–language–action and diffusion-based robot policies often use action chunking, where each policy query predicts a sequence of future actions and the robot executes an open-loop prefix before re-querying. While this interface improves local motion continuity, deployment still requires choosing the execution horizon : how much of each predicted chunk should be executed before acquiring a new observation. However, our experiments show that success is strongly task-dependent an”
  • Authors: junnan-nie; jiayi-li; jiachen-zhang; 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。