VLA [Tsinghua University]
UMI-Bridge: Action-Anchored Latent Alignment across Human and Robot Manipulation Data
Haiyi Liu et al. · 以动作作为锚点在隐空间对齐人类手持 UMI/第一视角视频与真机演示，缓解视角与具身差异，让无机器人采集的人类数据能直接用于训练操作策略，是扩大 VLA 数据来源的可复用方案。
cs.RO
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
- Energy-Regularized Imitation Learning for Force- and Work-Aware Robotic Manipulation 从关节力矩与角位移构造机械功代理量，训练可微能量预测器来正则化模仿学习，使策略具备力/功耗意识；力觉信号设计有参考价值，但未涉及 VLA 架构或明显的真机对比。 HF-PAPER
- rMuscle: Robotic Muscle Memory for Efficient Vision-Language-Action Model Inference Factory work is a promising early scenario for embodied AI: assigning repetitive manual jobs to robots has clear economic payoff, and a structured station keeps the jobs tractable for current policies. Vision-Language-Action (VLA) models now dominate as the policy paradigm for these robots. The inference latency of VLA models directly affects robot responsiveness and motion smoothness. However, existing VLA inference frameworks do not fully exploit the characteristics of embodied workloads or ac HF-PAPER
- PointZero: 3D Point Track Completion for Learning Transferable 3D Dynamics 用 3D 点轨迹补全学习可迁移的 3D 动力学世界模型，无需机器人动作标签即可预训练；对 world-action model 的动作条件建模有借鉴意义，但与 VLA 策略本身无直接接口。 HF-PAPER
- Acting in Meters: Learning Metric Interactions for Precise Robotic Manipulation 指出 VLA/世界动作模型把动作与物体、场景几何的度量关系隐含化，提出显式度量交互表示以提升精细操作精度；方向对口，但摘要未给出可复用的具体模块或结果。 HF-PAPER
- Vision-Language Grounded Task-Context-Aware Imitation Learning for Robotic Disassembly [Viterbi School of Engineering, University of Southern California] Jeon Ho Kang et al. · 针对多零件长时序拆解任务，用视觉-语言 grounding 提供任务上下文来区分多个有效目标，属特定工业应用场景的模仿学习，无 VLA 架构侧贡献。 CS.RO
- Causal-History Test-Time Scaling for Failure Recovery in Autoregressive World-Action Models [General Intelligence Machine] Lin Li et al. · 针对只在成功轨迹上训练的 world-action model，用因果历史做测试时扩展以实现失败恢复；补上 WAM 失败恢复这一环，但尚未见 VLA 策略上的迁移证据。 CS.RO
- Not All Layers Need Tuning: Diagnosing and Directing Adaptation in Vision-Language-Action Models [Robotics Institute, Carnegie Mellon University] Shahram Najam Syed et al. · arXiv:2609.18084v1 Announce Type: new Abstract: Fine-tuning a Vision-Language-Action (VLA) model for a new deployment environment is expensive, yet most methods apply uniform-capacity adapters to every network region as if every region requires equal adjustment. This paper tests that assumption on five architecturally diverse VLAs (OpenVLA-OFT, $\pi_0$, SmolVLA, DTP, Octo; 93M-7B parameters). Measuring per-region adaptation cost as normalized parameter displacement under region-isolated fine-tun CS.RO
- Technical Report: One-Step Drifting Action Heads for GR00T N1.7 [ZJU-UIUC Institute, Zhejiang University] Xihe Shao · 技术报告：把 GR00T N1.7 的迭代式动作头换成一步生成（drifting）动作头，考察闭环任务成功率是否受损；结论对 VLA 推理加速有参考，但为单作者报告、实验规模有限。 CS.RO
- A Comprehensive Review of Generative Physical Artificial Intelligence [Qualcomm (United States)] Satyam Gaba et al. · 综述类文章，梳理生成式基础模型与物理具身结合（GPAI）的感知-推理-行动范式，可作为选题背景，但无新方法与实验。 CS.RO
- Reinforcement Learning for Real-Time Vision-Language-Action Policies [Stanford|Finn] Perry Dong et al. · arXiv:2609.18207v1 Announce Type: new Abstract: Reinforcement learning fine-tuning on top of large, pretrained Vision-Language-Action (VLA) models offers promise for highly reliable robot deployment. However, because of their scale, modern VLA models suffer from high inference latency, so the observation used to select an action is often stale by execution time, creating a distribution shift that can substantially degrade reliability and performance. Prior work has explored asynchronous policy e CS.RO
- UMI-Bridge: Action-Anchored Latent Alignment across Human and Robot Manipulation Data [Tsinghua University] Haiyi Liu et al. · 以动作作为锚点在隐空间对齐人类手持 UMI/第一视角视频与真机演示，缓解视角与具身差异，让无机器人采集的人类数据能直接用于训练操作策略，是扩大 VLA 数据来源的可复用方案。 CS.RO
- ForceDelta-VLA: Distilling Force-Conditioned ActionCorrections for Contact-Rich Manipulation [TAMS (Technical Aspects of Multimodal Systems), Department of Informatics, University of Hamburg] Ju Dong et al. · arXiv:2609.18242v1 Announce Type: new Abstract: Force-aware Vision-Language-Action (VLA) policies improve contact-rich manipulation, but typically combine task-level motion and contact-dependent adjustment in a single action prediction. Demonstrations provide no explicit labels for decomposing that prediction into a reusable reference action and a correction. We present ForceDelta-VLA, a correction-distillation framework that constructs an explicit force-correction target using paired prediction CS.RO
VLA 研究日報VLA 研究日报
共 28 篇
🔧 技術技术
VLA [Huazhong University of Science and Technology]
ActionPiece: Rethinking Action Tokenization for Autoregressive Vision-Language-Action Models
Shijie Lian et al. · 指出动作 tokenizer 仅用逐点误差评估的缺陷，重新设计自回归 VLA 的动作分词目标与可执行指令还原方式，直接关系 AR-VLA 的轨迹保真度，做自回归策略的团队可马上对比采用。
cs.RO
閱讀原文
📖 背景閱讀背景阅读
VLA
Energy-Regularized Imitation Learning for Force- and Work-Aware Robotic Manipulation
从关节力矩与角位移构造机械功代理量，训练可微能量预测器来正则化模仿学习，使策略具备力/功耗意识；力觉信号设计有参考价值，但未涉及 VLA 架构或明显的真机对比。
hf-papers
閱讀原文
VLA
rMuscle: Robotic Muscle Memory for Efficient Vision-Language-Action Model Inference
Factory work is a promising early scenario for embodied AI: assigning repetitive manual jobs to robots has clear economic payoff, and a structured station keeps the jobs tractable for current policies. Vision-Language-Action (VLA) models now dominate as the policy paradigm for these robots. The inference latency of VLA models directly affects robot responsiveness and motion smoothness. However, existing VLA inference frameworks do not fully exploit the characteristics of embodied workloads or ac
hf-papers
閱讀原文
VLA
PointZero: 3D Point Track Completion for Learning Transferable 3D Dynamics
用 3D 点轨迹补全学习可迁移的 3D 动力学世界模型，无需机器人动作标签即可预训练；对 world-action model 的动作条件建模有借鉴意义，但与 VLA 策略本身无直接接口。
hf-papers
閱讀原文
VLA
Acting in Meters: Learning Metric Interactions for Precise Robotic Manipulation
指出 VLA/世界动作模型把动作与物体、场景几何的度量关系隐含化，提出显式度量交互表示以提升精细操作精度；方向对口，但摘要未给出可复用的具体模块或结果。
hf-papers
閱讀原文
VLA [Viterbi School of Engineering, University of Southern California]
Vision-Language Grounded Task-Context-Aware Imitation Learning for Robotic Disassembly
Jeon Ho Kang et al. · 针对多零件长时序拆解任务，用视觉-语言 grounding 提供任务上下文来区分多个有效目标，属特定工业应用场景的模仿学习，无 VLA 架构侧贡献。
cs.RO
閱讀原文
VLA [General Intelligence Machine]
Causal-History Test-Time Scaling for Failure Recovery in Autoregressive World-Action Models
Lin Li et al. · 针对只在成功轨迹上训练的 world-action model，用因果历史做测试时扩展以实现失败恢复；补上 WAM 失败恢复这一环，但尚未见 VLA 策略上的迁移证据。
cs.RO
閱讀原文
VLA [Robotics Institute, Carnegie Mellon University]
Not All Layers Need Tuning: Diagnosing and Directing Adaptation in Vision-Language-Action Models
Shahram Najam Syed et al. · arXiv:2609.18084v1 Announce Type: new Abstract: Fine-tuning a Vision-Language-Action (VLA) model for a new deployment environment is expensive, yet most methods apply uniform-capacity adapters to every network region as if every region requires equal adjustment. This paper tests that assumption on five architecturally diverse VLAs (OpenVLA-OFT, $\pi_0$, SmolVLA, DTP, Octo; 93M-7B parameters). Measuring per-region adaptation cost as normalized parameter displacement under region-isolated fine-tun
cs.RO
閱讀原文
VLA [ZJU-UIUC Institute, Zhejiang University]
Technical Report: One-Step Drifting Action Heads for GR00T N1.7
Xihe Shao · 技术报告：把 GR00T N1.7 的迭代式动作头换成一步生成（drifting）动作头，考察闭环任务成功率是否受损；结论对 VLA 推理加速有参考，但为单作者报告、实验规模有限。
cs.RO
閱讀原文
VLA [Qualcomm (United States)]
A Comprehensive Review of Generative Physical Artificial Intelligence
Satyam Gaba et al. · 综述类文章，梳理生成式基础模型与物理具身结合（GPAI）的感知-推理-行动范式，可作为选题背景，但无新方法与实验。
cs.RO
閱讀原文
VLA [Stanford|Finn]
Reinforcement Learning for Real-Time Vision-Language-Action Policies
Perry Dong et al. · arXiv:2609.18207v1 Announce Type: new Abstract: Reinforcement learning fine-tuning on top of large, pretrained Vision-Language-Action (VLA) models offers promise for highly reliable robot deployment. However, because of their scale, modern VLA models suffer from high inference latency, so the observation used to select an action is often stale by execution time, creating a distribution shift that can substantially degrade reliability and performance. Prior work has explored asynchronous policy e
cs.RO
閱讀原文
VLA [TAMS (Technical Aspects of Multimodal Systems), Department of Informatics, University of Hamburg]
ForceDelta-VLA: Distilling Force-Conditioned ActionCorrections for Contact-Rich Manipulation
Ju Dong et al. · arXiv:2609.18242v1 Announce Type: new Abstract: Force-aware Vision-Language-Action (VLA) policies improve contact-rich manipulation, but typically combine task-level motion and contact-dependent adjustment in a single action prediction. Demonstrations provide no explicit labels for decomposing that prediction into a reusable reference action and a correction. We present ForceDelta-VLA, a correction-distillation framework that constructs an explicit force-correction target using paired prediction
cs.RO
閱讀原文
VLA [The Hong Kong Polytechnic University]
${M}^2$Tok: Multi-head Multi-codebook Discrete Action Tokenization for Vision-Language-Action Models
Chunpu Xu et al. · 提出多头多码本离散动作 tokenizer，缓解连续动作离散化误差；与今日 ActionPiece 属同一动作分词方向的并行工作，创新点相对局部。
cs.RO
閱讀原文
VLA [The Hong Kong University of Science and Technology (Guangzhou)]
Function-Preserving Data Generation for Zero-Shot Real-to-Sim-to-Real Manipulation
Tianyi Xiang et al. · 在 real-to-sim-to-real 流程中生成几何多样且物理有效的接触丰富任务数据，强调保持功能语义；对数据合成思路有启发，但未与 VLA 策略训练打通。
cs.RO
閱讀原文
VLA [Institute for Material Handling and Logistics (IFL), Karlsruhe Institute of Technology]
RAFAIL: Relationship-Aware Failure Detection for Robotic Manipulation
Loris Schneider et al. · 用关系感知的失败检测替代逐帧 VLM 语义判断与 OOD 检测，降低运行时开销并避免对无害场景变化误报；属 VLM 侧监控模块，非 VLA 策略本体。
cs.RO
閱讀原文
VLA [上交]
GraphPoint: Semantic Entity Graphs and Point Trajectories for Compositional Robot Manipulation
Kang Luo et al. · 用语义实体图加物体点轨迹解耦语言与场景的伪相关，提升新指令下的组合泛化；泛化分析有价值，但属通用策略表示，未落到 VLA 主干上。
cs.RO
閱讀原文
VLA [Department of Computer Science, Yale University]
Decoupling Vision, Language, and Action for Efficient Multi-Task Robot Policies
Xiatao Sun et al. · 指出每个控制步都调用十亿参数 VLM 主干是浪费，把视觉-语言与动作解耦、以低频提供语义、高频跑轻量策略；思路与今日多篇 VLA 推理加速工作重合。
cs.RO
閱讀原文
VLA [Oxford Robotics Institute, University of Oxford]
DistAL: Distance-based Advantage Learning for VLA Fine-Tuning
Reece O'Mahoney et al. · 把 advantage conditioning 用于 flow-matching VLA 的微调，用基于距离的优势估计引导策略；与今日 RL 微调 VLA 主题重叠，方法增量偏小。
cs.RO
閱讀原文
VLA [Purdue University West Lafayette]
DetAug: Obstacle-Blind Trajectory Augmentation for Zero-shot Obstacle Avoidance
Reece O'Mahoney et al. · arXiv:2609.18395v1 Announce Type: new Abstract: Policies for robotic manipulation are produced by training on large teleoperated datasets. These datasets typically consist of free-space trajectories, making them difficult to transfer to test-time environments with obstacles. Previous methods for closing this gap have largely fallen into two groups. Dataset augmentation addresses it at training time but needs obstacle geometry in advance, whereas steering an existing checkpoint at inference time
cs.RO
閱讀原文
VLA [Technical University of Darmstadt]
ForwardDLO: Model-Based Bimanual Shape Matching of Unconstrained Deformable Linear Objects
Tim Missal et al. · 面向绳索/线缆等可变形线性物体的双臂形状匹配，属模型-based 的专用操作方法；与 VLA 学习范式关系弱。
cs.RO
閱讀原文
VLA [China Mobile (Hangzhou) Information Technology Co., Ltd]
TAO-Force: Unifying Force-Aware Perception and Fast-Slow Control for Contact-Rich Manipulation
Bohan Gan et al. · 指出 VLA 以视觉为主、位置控制的范式难以胜任接触丰富任务，提出融合力感知与快慢双环控制的方案；与今日 ForceDelta-VLA 同属力觉 VLA 方向，可选择其一精读。
cs.RO
閱讀原文
VLA [University of Science and Technology of China]
InterMASH: A Unified Geometric Representation for Grasp Synthesis
Xuanze Yang et al. · 提出跨人类手与机器人手的统一几何表示以生成稳定抓取，属抓取合成基础问题，与 VLA 策略训练无直接接口。
cs.RO
閱讀原文
VLA [Robotics Institute, Carnegie Mellon University]
ActiveScale: Scaling Active Perception for Robots across Model, Data, and Hardware
Shuai Zhou et al. · 让 VLA 在视角变化下主动获取被遮挡的任务相关信息，并沿模型/数据/硬件三轴扩展主动感知；方向有新意，但摘要未给出可复用的具体机制与量化对比。
cs.RO
閱讀原文
VLA [Nankai University]
DeformSmith: Physics Harness-Guided Hierarchical Generation of Deformable Assets for Robot Manipulation
Can Li et al. · 分层生成可变形物体的几何、外观与物理属性资产，用于机器人操作仿真；属仿真资产生成工具链，离 VLA 训练较远。
cs.RO
閱讀原文
VLA [北大|Wang]
From Gameplay to Policy: Towards Scalable Robot Data Collection via Gamified Robot-Free Interaction
Zheng Li et al. · 用游戏化的人机无关交互采集大规模多样化操作数据，替代昂贵的真机演示；数据规模化思路与 UMI 一类工作互补，但未涉及 VLA 训练细节。
cs.RO
閱讀原文
VLA [The University of Tokyo Institute of Science Tokyo]
VLA-ULAP: Interleaving Cloud VLA Calls with Ultra-Lightweight Local Action Prediction at the Edge
Deyu Cao et al. · 边缘部署方案：把云端大 VLA 调用与本地超轻量动作预测交错执行，兼顾功耗与响应时延；工程性强，但属系统层优化而非模型能力提升。
cs.RO
閱讀原文
VLA [University of Engineering and Technology, Vietnam National University, Hanoi]
Calibrated Probabilistic Obstruction Reasoning with Vision-Language Models for Grasping in Clutter
Thanh-Tuan Tran et al. · 把杂乱场景抓取中的遮挡推理建模为带校准的概率决策，在抓目标、移除遮挡物与延后之间选择；属 VLM 规划层工作，与 VLA 动作生成链路无关。
cs.RO
閱讀原文