VLA [LIBERO Team]
LIBERO-VPro: Benchmarking Closed-Loop Visual Robustness of Robotic Foundation Models
在 LIBERO 上构建闭环视觉鲁棒性 benchmark，系统评测机器人基础模型在视觉扰动、延迟、观测不一致下的失效模式，可直接用来压测自家 VLA 的鲁棒性上限。
hf-papers
閱讀原文
VLA 研究日報 Pulsar
FEED
— AI 線暫無可用資料 —— AI 线暂无可用资料 —
VLA 線VLA 线 · cs.RO · cs.AI · cs.LG
- LIBERO-VPro: Benchmarking Closed-Loop Visual Robustness of Robotic Foundation Models [LIBERO Team] 在 LIBERO 上构建闭环视觉鲁棒性 benchmark，系统评测机器人基础模型在视觉扰动、延迟、观测不一致下的失效模式，可直接用来压测自家 VLA 的鲁棒性上限。 HF-PAPER
- StenoVLA-3D: 3D-Aware Reasoning VLA for Navigation Through Gastrointestinal Stenoses 面向胃肠道狭窄导航的 3D 感知推理 VLA，用几何推理补偿纹理贫乏的单目内窥观测并保留病灶证据；任务与医疗内镜强绑定，通用 VLA 复用路径有限。 HF-PAPER
- JAMB: Joint Action-Motion Diffusion for Bimanual Manipulation [CMU|Held] Chuyang Xiao et al. · 把双臂动作与未来场景几何变化放进同一扩散过程联合去噪，显式预测单臂动作对共享 3D 场景的影响；做双臂协作/扩散策略可直接借鉴其 action-motion 联合建模。 CS.RO
- Capability-Aware Arbitration for Semantic Intent-Based Shared Control [Colorado School of Mines] Zhaoda Du et al. · arXiv:2609.25369v1 Announce Type: new Abstract: Shared control often allocates robot authority based on confidence in inferred human intent, assuming reliable autonomous execution. When this assumption fails, high intent confidence can cause over-helping. We present a capability-aware shared-control framework in which a vision-language model (VLM) infers human intent and provides semantic-intent confidence, while a vision-language-action (VLA) policy generates autonomous actions. VLA capability CS.RO
- VLAQuantBench: Closed-Loop Evaluation of Post-Training Quantization for Vision-Language-Action Models [LIBERO Team] Jiuyi Xu et al. · 闭环评测 VLA 训练后量化中层范围、数值格式与校准数据三者的交互，给出精度选择指南；属评测类工作，但可作为部署压缩 VLA 的选型工具。 CS.RO
- RoboMP-DINOv2: Prompts, Not Filters for Robust Robot Manipulation [Harvard School of Engineering and Applied Sciences] Han Qi et al. · 用 DINOv2 的 prompt 机制替代分割掩码过滤，为操作策略提供跨视觉偏移仍保留场景上下文的通用视觉表征；可作 VLA 视觉 backbone 的替换参考。 CS.RO
- HABILIS Brain 0: Geometry-Change Supervision for Vision-Language-Action and Residual Flow Recovery [Robotics Research (United States)] Jinu Pahk et al. · arXiv:2609.25558v1 Announce Type: new Abstract: Vision-language-action policies benefit from geometric supervision, but current-frame geometry alone does not explicitly describe the changes associated with manipulation. This design is motivated by the goal of learning an embodiment-agnostic visual interface that can be pretrained across robot and egocentric video before robot-specific action alignment. We introduce Geometry-Change VLA (GC-VLA), which learns to predict multiview future-current ge CS.RO
- IndustrialVLA-Bench: A Traceable Multi-Axis Evaluation of Open Robot Policy Models [Griffith University] Yiqi Wang et al. · 对开源 VLA 与世界动作模型（WAM）两类范式做可追溯的多维度工业场景评测；评测类工作，可作工业落地前的模型选型参考。 CS.RO
- CableVLA: Simulation-Privileged Global-Local Representation Learning for Cable Routing [State Key Laboratory of Intelligent Manufacturing Equipment and Technology, School of Mechanical Science and Engineering, Huazhong University of Science and Technology] Zhifei Teng et al. · 把仿真特权监督蒸馏成可部署的全局线缆拓扑+局部接触表征，端到端完成线缆走线；任务专用，但 sim-to-real 特权表征迁移思路可借鉴。 CS.RO
- From Instrument-Mounted Demonstrations to In-Vivo Execution: Learning Bimanual Laparoscopic Appendectomy Without Robot-Collected Demonstrations [Department of Transdisciplinary Medicine, Seoul National University Hospital] Dongho Yee et al. · 从手持器械上采集的示范中学习双臂腹腔镜阑尾切除操作并迁移到体内执行，避开机器人遥操作采数；手术场景专用，泛化到通用双臂操作尚需验证。 CS.RO
- MachEmbodied-U0: Unified Understanding and Generation Model for Embodied Intelligence [Li Auto Inc] Haoran Wen et al. · arXiv:2609.25627v1 Announce Type: new Abstract: General-purpose robot control requires models to understand task intent, identify where to interact, capture how the scene evolves, and generate precise actions. Vision-language-action models provide strong semantic priors but typically do not explicitly model scene dynamics, while world-action models couple visual prediction with control without necessarily exposing the task-relevant semantic and spatial structure needed for fine-grained manipulat CS.RO
- RoboFollow: Unveiling the Instruction Following Mirage in Embodied Agents [AutoLab, School of Artificial Intelligence, Shanghai Jiao Tong University] Chang Guo et al. · arXiv:2609.25636v1 Announce Type: new Abstract: Modern embodied agents achieve impressive success rates, yet their actual instruction-following ability is far weaker than these numbers suggest. We trace this illusion to a structural property we term low scene entropy: when a visual scene admits only one valid task, language becomes redundant and a policy can score highly while barely using it. We introduce RoboFollow, a diagnostic benchmark with three principles: (1) High Scene Entropy: each tra CS.RO
VLA 研究日報VLA 研究日报
共 29 篇
🔧 技術技术
VLA [CMU|Held]
JAMB: Joint Action-Motion Diffusion for Bimanual Manipulation
Chuyang Xiao et al. · 把双臂动作与未来场景几何变化放进同一扩散过程联合去噪，显式预测单臂动作对共享 3D 场景的影响；做双臂协作/扩散策略可直接借鉴其 action-motion 联合建模。
cs.RO
閱讀原文
VLA [浙大|Si]
SafeLoop: Risk-Aware Rollback for Vision-Language-Action Manipulation
Zeyu Lou et al. · 给 VLA 长时程执行加风险感知回滚：在状态估计/控制误差累积成碰撞、掉物等不可逆失败前检测风险并回退；可移植到自家 VLA 执行栈提升长时程可靠性。
cs.RO
閱讀原文
📖 背景閱讀背景阅读
VLA
StenoVLA-3D: 3D-Aware Reasoning VLA for Navigation Through Gastrointestinal Stenoses
面向胃肠道狭窄导航的 3D 感知推理 VLA，用几何推理补偿纹理贫乏的单目内窥观测并保留病灶证据；任务与医疗内镜强绑定，通用 VLA 复用路径有限。
hf-papers
閱讀原文
VLA [Colorado School of Mines]
Capability-Aware Arbitration for Semantic Intent-Based Shared Control
Zhaoda Du et al. · arXiv:2609.25369v1 Announce Type: new Abstract: Shared control often allocates robot authority based on confidence in inferred human intent, assuming reliable autonomous execution. When this assumption fails, high intent confidence can cause over-helping. We present a capability-aware shared-control framework in which a vision-language model (VLM) infers human intent and provides semantic-intent confidence, while a vision-language-action (VLA) policy generates autonomous actions. VLA capability
cs.RO
閱讀原文
VLA [LIBERO Team]
VLAQuantBench: Closed-Loop Evaluation of Post-Training Quantization for Vision-Language-Action Models
Jiuyi Xu et al. · 闭环评测 VLA 训练后量化中层范围、数值格式与校准数据三者的交互，给出精度选择指南；属评测类工作，但可作为部署压缩 VLA 的选型工具。
cs.RO
閱讀原文
VLA [Harvard School of Engineering and Applied Sciences]
RoboMP-DINOv2: Prompts, Not Filters for Robust Robot Manipulation
Han Qi et al. · 用 DINOv2 的 prompt 机制替代分割掩码过滤，为操作策略提供跨视觉偏移仍保留场景上下文的通用视觉表征；可作 VLA 视觉 backbone 的替换参考。
cs.RO
閱讀原文
VLA [Robotics Research (United States)]
HABILIS Brain 0: Geometry-Change Supervision for Vision-Language-Action and Residual Flow Recovery
Jinu Pahk et al. · arXiv:2609.25558v1 Announce Type: new Abstract: Vision-language-action policies benefit from geometric supervision, but current-frame geometry alone does not explicitly describe the changes associated with manipulation. This design is motivated by the goal of learning an embodiment-agnostic visual interface that can be pretrained across robot and egocentric video before robot-specific action alignment. We introduce Geometry-Change VLA (GC-VLA), which learns to predict multiview future-current ge
cs.RO
閱讀原文
VLA [Griffith University]
IndustrialVLA-Bench: A Traceable Multi-Axis Evaluation of Open Robot Policy Models
Yiqi Wang et al. · 对开源 VLA 与世界动作模型（WAM）两类范式做可追溯的多维度工业场景评测；评测类工作，可作工业落地前的模型选型参考。
cs.RO
閱讀原文
VLA [State Key Laboratory of Intelligent Manufacturing Equipment and Technology, School of Mechanical Science and Engineering, Huazhong University of Science and Technology]
CableVLA: Simulation-Privileged Global-Local Representation Learning for Cable Routing
Zhifei Teng et al. · 把仿真特权监督蒸馏成可部署的全局线缆拓扑+局部接触表征，端到端完成线缆走线；任务专用，但 sim-to-real 特权表征迁移思路可借鉴。
cs.RO
閱讀原文
VLA [Department of Transdisciplinary Medicine, Seoul National University Hospital]
From Instrument-Mounted Demonstrations to In-Vivo Execution: Learning Bimanual Laparoscopic Appendectomy Without Robot-Collected Demonstrations
Dongho Yee et al. · 从手持器械上采集的示范中学习双臂腹腔镜阑尾切除操作并迁移到体内执行，避开机器人遥操作采数；手术场景专用，泛化到通用双臂操作尚需验证。
cs.RO
閱讀原文
VLA [Li Auto Inc]
MachEmbodied-U0: Unified Understanding and Generation Model for Embodied Intelligence
Haoran Wen et al. · arXiv:2609.25627v1 Announce Type: new Abstract: General-purpose robot control requires models to understand task intent, identify where to interact, capture how the scene evolves, and generate precise actions. Vision-language-action models provide strong semantic priors but typically do not explicitly model scene dynamics, while world-action models couple visual prediction with control without necessarily exposing the task-relevant semantic and spatial structure needed for fine-grained manipulat
cs.RO
閱讀原文
VLA [AutoLab, School of Artificial Intelligence, Shanghai Jiao Tong University]
RoboFollow: Unveiling the Instruction Following Mirage in Embodied Agents
Chang Guo et al. · arXiv:2609.25636v1 Announce Type: new Abstract: Modern embodied agents achieve impressive success rates, yet their actual instruction-following ability is far weaker than these numbers suggest. We trace this illusion to a structural property we term low scene entropy: when a visual scene admits only one valid task, language becomes redundant and a policy can score highly while barely using it. We introduce RoboFollow, a diagnostic benchmark with three principles: (1) High Scene Entropy: each tra
cs.RO
閱讀原文
VLA [School of Automation and Intelligent Sensing, Shanghai Jiao Tong University]
PhyVisGen: Physically and Visually High-Fidelity Robotic Manipulation Data Generation
Yu Zheng et al. · 生成物理与视觉双高保真的操作演示数据以缩小仿真到真实的差距，缓解真实采数昂贵问题；数据生成类工作，缺真实数据时可作为补充来源。
cs.RO
閱讀原文
VLA [Embodied Perception and Interaction Lab (EmPACT Lab), School of Electrical and Electronic Engineering, Nanyang Technological University]
MotionForge: A Data Generation Pipeline and Large-Scale Benchmark for Long-Horizon Manipulation of Dynamic Objects with Domain Shifts
Mohan Liu et al. · 面向动态物体长时程操作的数据生成管线与大规模 benchmark，显式引入域偏移；可作为动态操作评测与数据来源，但属数据集/评测类。
cs.RO
閱讀原文
VLA [Duke University]
The Cartesian Hand: In-Hand Manipulation with All-Linear Fingers
Boxi Xia et al. · arXiv:2609.25696v1 Announce Type: new Abstract: Robotic manipulation has increasingly pursued human-like dexterous hands with many articulated degrees of freedom, offering rich manipulation capabilities at the cost of mechanical and control complexity. At the other extreme, parallel grippers are simple and robust, but provide little ability to manipulate an object after grasping it. Operating articulated objects such as threaded containers, manufacturing tools, and laboratory instruments often r
cs.RO
閱讀原文
VLA [Hong Kong Embodied AI Lab]
Fisheye-VLA: Decoupling Coverage and Acuity for Manipulation with a Single Fisheye Camera
Ziang Ren et al. · 用单个鱼眼相机同时解耦全局覆盖与局部精细反馈，替代前视+腕视双相机配置；属 VLA 输入接口改造，对传感器受限平台有参考价值。
cs.RO
閱讀原文
VLA [Institute of Automation, Chinese Academy of Sciences]
MedVLA: A Hierarchical Vision-Language-Action Framework for Closed-Loop Precision Medical Robot Manipulation
Junjie Xie et al. · 面向精密医疗操作的分层 VLA，用高层推理解耦连续动作生成以兼顾安全与可解释；医疗场景专用，通用操作复用有限。
cs.RO
閱讀原文
VLA
VisForce: Visual Grounding of Current and Desired Forces for Goal-Conditioned Dexterous Manipulation
Jung-Woo Lee et al. · arXiv:2609.25785v1 Announce Type: new Abstract: Vision-Language-Action (VLA) models have emerged as general-purpose robotic manipulation policies. However, in dexterous hand manipulation, contact forces are typically provided as separate states or force-specific representations, making it difficult to explicitly represent the spatial correspondence between force and their corresponding visual locations. In this work, we propose VisForce, which visually grounds the current and desired forces at t
cs.RO
閱讀原文
VLA [Southwest Wisconsin Technical College]
Beyond Reconstruction Error: Analytical and Data-Driven Action Tokenization for Autoregressive Vision-Language-Action Models
Yuxin Yang et al. · 跳出重建误差，分析自回归 VLA 中离散动作 tokenizer 的哪些表示属性真正决定闭环控制性能；对做动作 token 的 VLA 设计有直接指导意义。
cs.RO
閱讀原文
VLA [The Hong Kong University of Science and Technology (Guangzhou)]
What is the Better Curriculum: Controller-Shaped Grasping Behavior for Contact Force-Sensitive Manipulation
Ziyan Feng et al. · 针对亚牛顿级易损物体，用控制器塑形的抓取行为构造课程，而非把触觉当额外策略输入；触觉力控操作方向可参考其课程设计。
cs.RO
閱讀原文
VLA [Xi'an Jiaotong University]
StrataVLA: Hierarchical and Efficient 3D Geometric Grounding for Vision-Language-Action Models
Jin Cui et al. · 分层高效地把 3D 几何接地进 VLA，无需显式深度或点云即可获得空间感知；3D 接地 VLA 已有大量同类工作，此处属增量改进。
cs.RO
閱讀原文
VLA [中科院]
GDLAM: Group-Disentangled Latent Action Model for Highly Disentangled Embodied Pretraining
Jiarui Yang et al. · 用分组解耦的隐动作模型，解决异构无动作视频在自监督未来预测中相互坍缩的问题，为具身预训练提供更干净的动作表征；可作 VLA 预训练表征参考。
cs.RO
閱讀原文
VLA [Peking University]
RoboTwin-Phys: Do WAMs and VLAs Understand the Physical World?
Jiaqi Zhang et al. · 在 RoboTwin 上加入物理条件多样性（质量/摩擦/刚度等），系统检验世界动作模型与 VLA 是否真正理解物理；可作为物理泛化评测的补充。
cs.RO
閱讀原文
VLA [School of Computer Science, The University of Sydney]
MAVP: Map-Aware Visuomotor Policies for Mobile Manipulation
Jinhe Tang et al. · 把地图信息显式注入视运动策略，缓解仅靠演示训练时底盘运动不准造成的空间错位；做移动操作可参考其地图条件化设计。
cs.RO
閱讀原文
VLA [Department of Mechanical and Industrial Engineering, University of Toronto]
RouteRLT: Learning When and Which RL Specialist Should Control a Vision-Language-Action Policy
Chongyu Zhu et al. · arXiv:2609.26467v1 Announce Type: new Abstract: Vision-language-action (VLA) models provide broad manipulation competence, but often struggle during the precision-critical stages that dominate contact-rich industrial tasks such as connector insertion and cable management. A common remedy is to refine a pretrained VLA with reinforcement learning (RL), enabling task-specific improvement beyond behavior cloning. However, how to preserve its generalist behavior while deciding when RL refinement is n
cs.RO
閱讀原文
VLA
MATE: Multi-Agent Virtual Teleoperation Platform for Humanoid Collaboration Data Collection
Yichuan Yu et al. · 多智能体虚拟遥操作平台，用 VR 采集人形机器人协作的移动操作数据，补足现有多机器人协作数据缺口；属数据采集工具类。
cs.RO
閱讀原文
VLA [Samsung Robotics eXperience]
Imperfection for Precision: Upcycling Imperfect Data for High-Precision Robotic Manipulation
Hao Wei et al. · arXiv:2609.26672v1 Announce Type: new Abstract: Training vision-language-action (VLA) models for high-precision manipulation typically requires task-specific, high-quality data (e.g., teleoperation), which is slow and expensive to collect. To reduce this burden without compromising manipulation precision, we propose $\varepsilon$4P (Imperfection for Precision), a simple yet effective method that "upcycles" two otherwise discarded data sources: (1) low-precision data from the target task and (2)
cs.RO
閱讀原文
VLA [Institute of Automation, Chinese Academy of Sciences]
World In Your Hands: A Large-Scale and Open-Source Ecosystem for Learning Human-Centric Manipulation in the Wild
Yupeng Zheng et al. · 开源 1000+ 小时野外采集、毫米级精度的第一人称人类操作数据生态与可穿戴采集套件；人类数据是 VLA 预训练的重要补充，属数据集类工作。
cs.RO
閱讀原文