{
  "project": "PaperRead",
  "export_date": "2026-10-09",
  "paper_count": 30,
  "uploaded_pdf_count": 29,
  "uploaded_unique_paper_count": 27,
  "official_supplement_count": 3,
  "han_characters_in_individual_summaries": 109495,
  "source_policy": "以附件版本为主体，官方资料核对版本与重要澄清；未重跑实验",
  "papers": [
    {
      "title": "Vision Language Action Models in Robotic Manipulation: A Systematic Review",
      "title_zh": "机器人操作中的视觉语言动作模型：系统综述",
      "year": "2025（所存预印本）",
      "arxiv": "2507.10672v1",
      "category": "综述与研究地图",
      "one_sentence_contribution": "汇总模型、数据集、仿真平台与泛化评价，并提出数据集描述分数。",
      "main_limit": "定性等级不是统一重跑；部分平台、频率和计数条目需回查一手论文。",
      "official_urls": [
        "https://arxiv.org/abs/2507.10672v1",
        "https://github.com/Muhayyuddin/VLAs"
      ],
      "source_pdf": "Din et al. - 2025 - Vision Language Action Models in Robotic Manipulation A Systematic Review.pdf",
      "alias": "SYSTEMATIC",
      "short_name": "VLA 系统综述",
      "source_kind": "项目论文附件",
      "paper_id": "P01",
      "han_characters": 2881,
      "summary_sha256": "9aad2302a711d5946c8917925ea843378720dd70bb75e7161cb36d11c8f62e8c",
      "source_sha256": "881935edbd0c6bbc66f6dd9145f5e8cee2b29abc62ef7fca66d07f8d54770f5b",
      "source_pdf_pages": 28,
      "filename": "Vision_Language_Action_Models_in_Robotic_Manipulation-A_Systematic_Review.md"
    },
    {
      "title": "A Survey on Vision-Language-Action Models: An Action Tokenization Perspective",
      "title_zh": "视觉—语言—动作模型综述：动作令牌化视角",
      "year": "2025",
      "arxiv": "2507.01925v1",
      "category": "综述与研究地图",
      "one_sentence_contribution": "用八类动作相关中间表示统一分析模型模块、数据来源与执行接口。",
      "main_limit": "分类可重叠且 token 定义宽泛；组合架构是研究判断，没有统一对照实验。",
      "official_urls": [
        "https://arxiv.org/abs/2507.01925v1",
        "https://arxiv.org/html/2507.01925v1"
      ],
      "source_pdf": "2507.01925v1.pdf",
      "alias": "TOKEN_SURVEY",
      "short_name": "动作令牌化综述",
      "source_kind": "项目论文附件",
      "paper_id": "P02",
      "han_characters": 4967,
      "summary_sha256": "544667a8bc7ad4bf4aad3431a317064d6010cf38ee1bbe2cbef8301bfac5ae19",
      "source_sha256": "f8a9e5e8902986b7f7673120a33822e2b5b0005f6647f14cb4665947bb7caec0",
      "source_pdf_pages": 70,
      "filename": "A_Survey_on_Vision-Language-Action_Models-An_Action_Tokenization_Perspective.md"
    },
    {
      "title": "面向具身操作的视觉-语言-动作模型综述",
      "title_zh": "面向具身操作的视觉-语言-动作模型综述",
      "title_en": "Survey of Vision-Language-Action Models for Embodied Manipulation",
      "year": "2025",
      "arxiv": "2508.15201v2",
      "category": "综述与研究地图",
      "one_sentence_contribution": "从架构、数据、预训练、后训练与评估五个环节梳理具身操作 VLA。",
      "main_limit": "部分方法和数据表存在转述不一致，成绩不是统一训练或统一硬件比较。",
      "official_urls": [
        "https://arxiv.org/abs/2508.15201v2"
      ],
      "source_pdf": "2508.15201v2.pdf",
      "alias": "CN_SURVEY",
      "short_name": "具身操作 VLA 中文综述",
      "source_kind": "项目论文附件",
      "paper_id": "P03",
      "han_characters": 3959,
      "summary_sha256": "f9c5bac351d3e8e53917debe787f22aae35da6ee9bab101117e7b5f3e1ac9f1a",
      "source_sha256": "dcd9c575ceb9729d19cb1d6dcf34115d30159674c58fb58b5e906cd3ca6fd210",
      "source_pdf_pages": 40,
      "filename": "面向具身操作的视觉-语言-动作模型综述.md"
    },
    {
      "title": "A Survey on Reinforcement Learning of Vision-Language-Action Models for Robotic Manipulation",
      "title_zh": "面向机器人操作的视觉—语言—动作模型强化学习综述",
      "year": 2025,
      "arxiv": null,
      "doi": "10.36227/techrxiv.176531955.54563920/v1",
      "source_pdf": "RL_VLA_Survey.pdf",
      "category": "VLA强化学习综述",
      "one_sentence_contribution": "以动作、奖励、转移建模、在线/离线/推理时范式和真实部署构成 RL-VLA 系统分类与研究路线。",
      "main_limit": "分类和结果来自异构原论文，未统一重跑；以官方同题名 PDF 补齐，原项目附件版本不可逐字节核验。",
      "official_urls": [
        "https://doi.org/10.36227/techrxiv.176531955.54563920/v1",
        "https://www.techrxiv.org/doi/full/10.36227/techrxiv.176531955.54563920/v1",
        "https://github.com/Denghaoyuan123/Awesome-RL-VLA",
        "https://github.com/Denghaoyuan123/Awesome-RL-VLA/blob/main/A_Survey_on_Reinforcement_Learning_of_Vision-Language-Action_Models_for_Robotic_Manipulation.pdf"
      ],
      "source_status": "official_fulltext_supplement_original_attachment_version_unverified",
      "alias": "RL_SURVEY",
      "short_name": "RL-VLA 综述",
      "source_kind": "项目早期记录对应的官方补充原文",
      "paper_id": "P04",
      "han_characters": 4654,
      "summary_sha256": "a2528c24d46477448ee0e062d4b8b07b280ead7b06bfd093f93b571a75e11280",
      "source_sha256": "c98dd130ab275f3050ef8ef89dc31dc6745152950bd43b8d1b4b52d1591aab81",
      "source_pdf_pages": 20,
      "filename": "A_Survey_on_Reinforcement_Learning_of_Vision-Language-Action_Models_for_Robotic_Manipulation.md"
    },
    {
      "title": "RL²: Fast Reinforcement Learning via Slow Reinforcement Learning",
      "title_zh": "RL²：通过慢速强化学习学会快速强化学习",
      "year": 2016,
      "arxiv": "1611.02779v2",
      "source_pdf": "RL2.pdf",
      "category": "元强化学习与循环状态适应",
      "one_sentence_contribution": "用慢速外层强化学习训练GRU，使其在新任务中仅依靠跨回合隐藏状态更新学习探索与利用。",
      "main_limit": "外层训练成本高且长时域优化困难，快速适应依赖任务分布与trial边界，实验不支持任意任务的通用最优学习。",
      "official_urls": [
        "https://arxiv.org/abs/1611.02779v2",
        "https://arxiv.org/abs/1611.02779",
        "https://openai.com/index/rl2/"
      ],
      "alias": "RL2",
      "short_name": "RL²",
      "source_kind": "项目早期记录对应的官方补充原文",
      "paper_id": "P05",
      "han_characters": 4045,
      "summary_sha256": "8b60de492e3975d4d312c48eed5b2adae239a5d79d670ebdee20db6d71395b99",
      "source_sha256": "39c4537b080187e8ab02127381efd3c198ab6bbbaeeef72d08c62f4a884926a8",
      "source_pdf_pages": 14,
      "filename": "RL2-Fast_Reinforcement_Learning_via_Slow_Reinforcement_Learning.md"
    },
    {
      "title": "SimBa: Simplicity Bias for Scaling Up Parameters in Deep Reinforcement Learning",
      "title_zh": "SimBa：利用简洁性偏置扩大深度强化学习网络的参数规模",
      "year": 2025,
      "arxiv": "2410.09754v2",
      "source_pdf": "SimBa.pdf",
      "category": "强化学习网络架构与规模扩展",
      "one_sentence_contribution": "以运行统计归一化、残差前馈块和输出前 LayerNorm 改造网络骨干，使更大的 critic 在多个强化学习算法中提高样本与计算效率。",
      "main_limit": "证据主要来自状态和符号输入的仿真任务；简洁性频谱是二维初始化代理且存在反例，尚未验证视觉 VLA、真实机器人或社会导航安全。",
      "official_urls": [
        "https://arxiv.org/abs/2410.09754",
        "https://sonyresearch.github.io/simba/",
        "https://github.com/SonyResearch/simba",
        "https://openreview.net/forum?id=jXLiDKsuDo"
      ],
      "alias": "SIMBA",
      "short_name": "SimBa",
      "source_kind": "项目早期记录对应的官方补充原文",
      "paper_id": "P06",
      "han_characters": 4398,
      "summary_sha256": "fb83abf2df306a1287e9b506a4729d750968f4aeabca23484b865c2e5bfc7900",
      "source_sha256": "11aed779711adaeef85ef44a9c61cc1c1f68b0161fa6cec3f84014f764126d94",
      "source_pdf_pages": 33,
      "filename": "SimBa-Simplicity_Bias_for_Scaling_Up_Parameters_in_Deep_Reinforcement_Learning.md"
    },
    {
      "title": "Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware",
      "title_zh": "利用低成本硬件学习精细双臂操作",
      "year": "2023",
      "arxiv": "2304.13705v1",
      "category": "动作生成基础",
      "one_sentence_contribution": "把 ALOHA 遥操作数据、CVAE 动作分块和时间集成结合，学习精细双臂操作。",
      "main_limit": "真实任务表现不均衡；时间集成和潜变量的作用受数据与任务条件限制。",
      "official_urls": [
        "https://arxiv.org/abs/2304.13705v1",
        "https://tonyzhaozh.github.io/aloha/"
      ],
      "source_pdf": "Zhao et al. - 2023 - Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware.pdf",
      "alias": "ACT",
      "short_name": "ACT / ALOHA",
      "source_kind": "项目论文附件",
      "paper_id": "P07",
      "han_characters": 2727,
      "summary_sha256": "6e9bdc4847a060ba3ec5866a938c0cc8f9a8bfaeb54a8cc7548b4266673be783",
      "source_sha256": "2e2fe25860f5f9cee9e655a2714e9f2264a7a5078bcd267c16d3f39af461345a",
      "source_pdf_pages": 18,
      "filename": "Learning_Fine-Grained_Bimanual_Manipulation_with_Low-Cost_Hardware.md"
    },
    {
      "title": "Diffusion Policy: Visuomotor Policy Learning via Action Diffusion",
      "title_zh": "扩散策略：通过动作扩散学习视觉运动策略",
      "year": "2023 / 2024 扩展版",
      "arxiv": "2303.04137v5",
      "category": "动作生成基础",
      "one_sentence_contribution": "在视觉条件下联合去噪生成连续动作序列，并用滚动执行形成闭环。",
      "main_limit": "多步采样、视觉条件与动作长度影响实时性；复杂双臂任务仍有较多失败。",
      "official_urls": [
        "https://arxiv.org/abs/2303.04137v5",
        "https://diffusion-policy.cs.columbia.edu/",
        "https://github.com/real-stanford/diffusion_policy"
      ],
      "source_pdf": "Chi et al. - 2024 - Diffusion Policy Visuomotor Policy Learning via Action Diffusion.pdf",
      "alias": "DP",
      "short_name": "Diffusion Policy",
      "source_kind": "项目论文附件",
      "paper_id": "P08",
      "han_characters": 2661,
      "summary_sha256": "c152ef2d38493de411b95ccb542088b5c0c4245434fff26743c8b4ef414b628a",
      "source_sha256": "b65c474b696a4802d8f1457d86b637ce2c5521412570d3aa928cd54563babc8f",
      "source_pdf_pages": 19,
      "filename": "Diffusion_Policy-Visuomotor_Policy_Learning_via_Action_Diffusion.md"
    },
    {
      "title": "π₀: A Vision-Language-Action Flow Model for General Robot Control",
      "title_zh": "π₀：用于通用机器人控制的视觉—语言—动作流模型",
      "year": "2024（所存 v1）",
      "arxiv": "2410.24164v1",
      "category": "动作生成基础",
      "one_sentence_contribution": "用预训练视觉语言主干与流匹配动作专家，学习跨机器人和多任务的连续动作块。",
      "main_limit": "广泛预训练与任务后训练的贡献需分开，控制频率不等于新视觉闭环频率。",
      "official_urls": [
        "https://arxiv.org/abs/2410.24164v1",
        "https://www.physicalintelligence.company/blog/pi0"
      ],
      "source_pdf": "Black et al. - 2026 - $π_0$ A Vision-Language-Action Flow Model for General Robot Control.pdf",
      "alias": "PI0",
      "short_name": "π₀",
      "source_kind": "项目论文附件",
      "paper_id": "P09",
      "han_characters": 2843,
      "summary_sha256": "081f2a62337431a70158fcab65abd0cb03c88c516be151f44c5ffc073c87c4f1",
      "source_sha256": "e507e0360e8be44655a789ca7a38c2457aa463fc52d320c6a39efd6bdd542bc8",
      "source_pdf_pages": 17,
      "filename": "pi0-A_Vision-Language-Action_Flow_Model_for_General_Robot_Control.md"
    },
    {
      "title": "MetaVLA: Unified Meta Co-training For Efficient Embodied Adaption",
      "title_zh": "MetaVLA：面向高效具身适应的统一元协同训练",
      "year": 2025,
      "arxiv": "2510.05580v3",
      "source_pdf": "Li et al. - 2026 - MetaVLA Unified Meta Co-training For Efficient Embodied Adaption.pdf",
      "category": "VLA多任务适配与上下文学习",
      "one_sentence_contribution": "用上下文库和 Meta-Action-Reasoner 支持统一 VLA 多任务协同训练并利用异构辅助示范。",
      "main_limit": "主证据限于模拟任务适配，未证明新技能零样本适应或真实在线学习；默认最佳配置不使用随机潜变量。",
      "official_urls": [
        "https://arxiv.org/abs/2510.05580v3",
        "https://stellar-neuron.github.io/metavla/"
      ],
      "alias": "META",
      "short_name": "MetaVLA",
      "source_kind": "项目论文附件",
      "paper_id": "P10",
      "han_characters": 2783,
      "summary_sha256": "3b1e95b96d84f50a0c6918d5696b3431d80b131a5199dbb1616e7d1898b82da9",
      "source_sha256": "d49247b60b5c8f691858db078ed3cc5f3b24c0016749404a3674296178e1d6c1",
      "source_pdf_pages": 25,
      "filename": "MetaVLA-Unified_Meta_Co-training_For_Efficient_Embodied_Adaption.md"
    },
    {
      "title": "π_RL: Online RL Fine-tuning for Flow-based Vision-Language-Action Models",
      "title_zh": "π_RL：面向流式视觉—语言—动作模型的在线强化学习微调",
      "year": 2025,
      "arxiv": "2510.25889v3",
      "source_pdf": "Chen et al. - 2026 - $π_texttt{RL}$ Online RL Fine-tuning for Flow-based Vision-Language-Action Models.pdf",
      "category": "VLA在线强化学习",
      "one_sentence_contribution": "用 Flow-Noise 与 Flow-SDE 构造可优化的随机生成过程，对流式 VLA 动作专家进行在线 RL 微调。",
      "main_limit": "仿真与计算成本较高，训练随机过程与推理有差距；原文时间方向及 score 符号需结合实现核对。",
      "official_urls": [
        "https://arxiv.org/abs/2510.25889v3",
        "https://github.com/RLinf/RLinf",
        "https://huggingface.co/RLinf",
        "https://rlinf.readthedocs.io/en/latest/rst_source/examples/embodied/pi0.html"
      ],
      "alias": "PI_RL",
      "short_name": "π_RL",
      "source_kind": "项目论文附件",
      "paper_id": "P11",
      "han_characters": 3002,
      "summary_sha256": "cd21557c935a3d02fbbf7fc551fe40fc18e9d97c6f076a89ce61f99eaf2bd94e",
      "source_sha256": "484e9f8be118be4f7adfd18470abc330e29b1b4cf457d316dc85a63e65299d36",
      "source_pdf_pages": 24,
      "filename": "pi_RL-Online_RL_Fine-tuning_for_Flow-based_Vision-Language-Action_Models.md"
    },
    {
      "title": "RL Token: Bootstrapping Online RL with Vision-Language-Action Models",
      "title_zh": "RL Token：利用视觉—语言—动作模型引导在线强化学习",
      "year": 2026,
      "arxiv": "2604.23073",
      "source_pdf": "Xu et al. - RL Token Bootstrapping Online RL with Vision-Language-Action Models.pdf",
      "category": "真实机器人在线强化学习",
      "one_sentence_contribution": "冻结 VLA 与压缩表征，利用 RL token、参考动作和小型动作块 actor-critic 高效学习关键操作阶段。",
      "main_limit": "仍依赖任务示范、人类接管与阶段管理，主要在少数精细操作任务验证，未解决整体 VLA 推理延迟。",
      "official_urls": [
        "https://arxiv.org/abs/2604.23073",
        "https://www.pi.website/research/rlt",
        "https://www.pi.website/download/rlt.pdf"
      ],
      "alias": "RLT",
      "short_name": "RL Token",
      "source_kind": "项目论文附件",
      "paper_id": "P12",
      "han_characters": 3246,
      "summary_sha256": "3d075613bc5400d55eb8426359543b9783f1ab207b85d7d188d73a1d7effa49d",
      "source_sha256": "feecf0f5799bcb3775f42d5de5d5832e61eb379bd47b627ad9a0f6c86f3ebf31",
      "source_pdf_pages": 13,
      "filename": "RL_Token-Bootstrapping_Online_RL_with_Vision-Language-Action_Models.md"
    },
    {
      "title": "What Can RL Bring to VLA Generalization? An Empirical Study",
      "title_zh": "强化学习能给视觉—语言—动作模型的泛化带来什么？一项实证研究",
      "year": 2025,
      "arxiv": "2505.19789v4",
      "source_pdf": "Liu et al. - 2026 - What Can RL Bring to VLA Generalization An Empirical Study.pdf",
      "category": "VLA强化学习与泛化评估",
      "one_sentence_contribution": "用视觉、语义和执行三维 OOD 测试说明 PPO 的主要泛化收益在执行恢复，并给出共享价值头训练配方。",
      "main_limit": "基于有限模型与规划器示范，优势并非覆盖所有语义和视觉测试；真实迁移仅有小规模验证。",
      "official_urls": [
        "https://arxiv.org/abs/2505.19789v4",
        "https://rlvla.github.io/",
        "https://github.com/gen-robot/RL4VLA",
        "https://huggingface.co/collections/gen-robot/rlvla"
      ],
      "alias": "GEN",
      "short_name": "RL 与 VLA 泛化实证",
      "source_kind": "项目论文附件",
      "paper_id": "P13",
      "han_characters": 4159,
      "summary_sha256": "fa4c44309859523e20dddcf6e2a418bc1259395b7105929f09ffeb2f99db7294",
      "source_sha256": "2dc2c73ab51cf2a97f6e3223cf5dd229632afd599a38a9b515f39d46c2e411a9",
      "source_pdf_pages": 24,
      "filename": "What_Can_RL_Bring_to_VLA_Generalization-An_Empirical_Study.md"
    },
    {
      "title": "Action-aware Dynamic Pruning for Efficient Vision-Language-Action Manipulation",
      "title_zh": "面向高效视觉—语言—动作操控的动作感知动态剪枝",
      "year": 2026,
      "arxiv": "2509.22093",
      "source_pdf": "13478_Action_aware_Dynamic_Pru.pdf",
      "category": "视觉token剪枝与动作感知计算分配",
      "one_sentence_contribution": "按近期机器人动作幅度决定是否剪枝，并利用文本—视觉相关性保留输入token。",
      "main_limit": "动作幅度只是视觉需求的代理；长任务精度下降，FLOPs加速不能等同端到端时延。",
      "official_urls": [
        "https://openreview.net/pdf?id=ea6j8k8Rnw",
        "https://arxiv.org/abs/2509.22093",
        "https://github.com/chen7086/VLA-ADP",
        "https://chen7086.github.io/VLA-ADP/"
      ],
      "alias": "ADP",
      "short_name": "ADP",
      "source_kind": "项目论文附件",
      "paper_id": "P14",
      "han_characters": 3575,
      "summary_sha256": "709463c0a78166232ebedc9cc78aa20557be243a7f27c614e9b6380c5e6756fb",
      "source_sha256": "9751dc84d05125ccbb7104b9d26a9011a11aadb4d118d1111219be00e0bbdfe6",
      "source_pdf_pages": 20,
      "filename": "Action-aware_Dynamic_Pruning_for_Efficient_Vision-Language-Action_Manipulation.md"
    },
    {
      "title": "Realtime-VLA FLASH: Speculative Inference Framework for Diffusion-based VLAs",
      "title_zh": "Realtime-VLA FLASH：面向扩散式视觉—语言—动作模型的推测推理框架",
      "year": 2026,
      "arxiv": "2605.13778v1",
      "source_pdf": "Niu et al. - 2026 - Realtime-VLA FLASH Speculative Inference Framework for Diffusion-based VLAs.pdf",
      "category": "草稿验证与推测推理加速",
      "one_sentence_contribution": "轻量草稿利用新图像生成动作，由复用旧视觉KV的动作专家并行验证连续前缀。",
      "main_limit": "验证为启发式局部一致性检查，旧缓存与分布漂移可能造成共同错误，且需要额外训练。",
      "official_urls": [
        "https://arxiv.org/abs/2605.13778v1",
        "https://github.com/dexmal/realtime-vla-flash",
        "https://dexmal.github.io/realtime-vla-flash/"
      ],
      "alias": "FLASH",
      "short_name": "Realtime-VLA FLASH",
      "source_kind": "项目论文附件",
      "paper_id": "P15",
      "han_characters": 3579,
      "summary_sha256": "a325f73cd61aa3fd51ce0e5d94465bde72826b6891d990bd5b595efe0649a8b9",
      "source_sha256": "7f69c62457d39d7e4308d0446db179a578c751248e31d886266f6c792a486b50",
      "source_pdf_pages": 15,
      "filename": "Realtime-VLA_FLASH-Speculative_Inference_Framework_for_Diffusion-based_VLAs.md"
    },
    {
      "title": "Real-Time Execution of Action Chunking Flow Policies",
      "title_zh": "动作分块流策略的实时执行",
      "year": 2025,
      "arxiv": "2506.07339v2",
      "source_pdf": "Black et al. - 2025 - Real-Time Execution of Action Chunking Flow Policies.pdf",
      "category": "异步动作执行与流模型补全",
      "one_sentence_contribution": "用旧动作前缀和软掩码引导流式动作补全，在并行推理与执行时保持动作块连续。",
      "main_limit": "单次推理因梯度引导变慢，冻结前缀期间仍无法立即利用突发环境变化，也无硬实时保证。",
      "official_urls": [
        "https://arxiv.org/abs/2506.07339v2",
        "https://papers.nips.cc/paper_files/paper/2025/hash/300ccb2187dedd4edcc07f7e76d8e553-Abstract-Conference.html",
        "https://github.com/Physical-Intelligence/real-time-chunking-kinetix",
        "https://pi.website/research/real_time_chunking"
      ],
      "alias": "RTC",
      "short_name": "RTC",
      "source_kind": "项目论文附件",
      "paper_id": "P16",
      "han_characters": 3670,
      "summary_sha256": "b980a6ca43688352427da81bd0184de512ceb2ab903fe563e507242aca016d80",
      "source_sha256": "dd6d602be1f50b894d88517bf68916c4272885b2e29d96cffc426dacff41898f",
      "source_pdf_pages": 25,
      "filename": "Real-Time_Execution_of_Action_Chunking_Flow_Policies.md"
    },
    {
      "title": "Running VLAs at Real-time Speed",
      "title_zh": "使视觉—语言—动作模型以实时速度运行",
      "year": 2025,
      "arxiv": "2510.26742v1",
      "source_pdf": "Ma et al. - 2025 - Running VLAs at Real-time Speed.pdf",
      "category": "推理内核优化与流式执行设计",
      "one_sentence_contribution": "以CUDA Graph、代数折叠和Triton融合缩短π0前向，并提出视觉/动作专家重叠的流式设计。",
      "main_limit": "完整流式与progressive chunking仍为设计草图，接笔仅为一维夹爪任务，不能据此宣称通用高速操控。",
      "official_urls": [
        "https://arxiv.org/abs/2510.26742v1",
        "https://github.com/dexmal/realtime-vla"
      ],
      "alias": "RUNNING",
      "short_name": "Running VLAs",
      "source_kind": "项目论文附件",
      "paper_id": "P17",
      "han_characters": 3602,
      "summary_sha256": "cb05acd6b59dc01094b0c5afe1e54f23d8af5d3accc31eaedcf8464b955ed291",
      "source_sha256": "c9e97418bb7ede18c58a8ad1715060d8af21e0a00e6ffe158feb0a8a81f3571a",
      "source_pdf_pages": 11,
      "filename": "Running_VLAs_at_Real-time_Speed.md"
    },
    {
      "title": "VLASH: Real-Time VLAs via Future-State-Aware Asynchronous Inference",
      "title_zh": "VLASH：通过未来状态感知的异步推理实现实时视觉—语言—动作控制",
      "year": 2025,
      "arxiv": "2512.01031v1",
      "source_pdf": "Tang et al. - 2025 - VLASH Real-Time VLAs via Future-State-Aware Asynchronous Inference.pdf",
      "category": "未来本体状态条件化与异步推理",
      "one_sentence_contribution": "用当前图像配对未来机器人状态和动作进行微调，使动作生成适应异步执行起点。",
      "main_limit": "依赖未来状态估计与偏移微调；未来本体状态不等于未来环境，量化加速可能损害精细控制。",
      "official_urls": [
        "https://arxiv.org/abs/2512.01031v1",
        "https://arxiv.org/abs/2512.01031",
        "https://github.com/mit-han-lab/vlash"
      ],
      "alias": "VLASH",
      "short_name": "VLASH",
      "source_kind": "项目论文附件",
      "paper_id": "P18",
      "han_characters": 3566,
      "summary_sha256": "0eee49b195fc4775af64fd9467279b84c7f6d64ff117a0f1ca5135c62ac86b89",
      "source_sha256": "df0a64a49dd257c348f1053311248450095c187e0268dc97f8a18d2d28b27cec",
      "source_pdf_pages": 13,
      "filename": "VLASH-Real-Time_VLAs_via_Future-State-Aware_Asynchronous_Inference.md"
    },
    {
      "title": "DynamicVLA: A Vision-Language-Action Model for Dynamic Object Manipulation",
      "title_zh": "DynamicVLA：面向动态物体操控的视觉—语言—动作模型",
      "year": 2026,
      "arxiv": "2601.22153v1",
      "source_pdf": "Xie et al. - 2026 - DynamicVLA A Vision-Language-Action Model for Dynamic Object Manipulation.pdf",
      "category": "动态对象模型、数据与动作时序",
      "one_sentence_contribution": "结合430M多帧VLA、持续推理、过期动作丢弃及DOM动态操控数据，提高动态目标操作成功率。",
      "main_limit": "旧观测后缀仍受延迟影响，扰动与运动理解较弱；v1存在频率和公式记号问题，需参照v2澄清。",
      "official_urls": [
        "https://arxiv.org/abs/2601.22153v1",
        "https://arxiv.org/abs/2601.22153",
        "https://arxiv.org/html/2601.22153v2",
        "https://github.com/hzxie/DynamicVLA",
        "https://haozhexie.com/project/dynamic-vla"
      ],
      "alias": "DYNAMIC",
      "short_name": "DynamicVLA",
      "source_kind": "项目论文附件",
      "paper_id": "P19",
      "han_characters": 3707,
      "summary_sha256": "12b088f857dee847fdc24657a7af13d611e339bc35bea3df30bdf9a48a406ab4",
      "source_sha256": "4648cbf69775c48a9cdb59c444a738243635888021d939ab12edcf36349bb66b",
      "source_pdf_pages": 14,
      "filename": "DynamicVLA-A_Vision-Language-Action_Model_for_Dynamic_Object_Manipulation.md"
    },
    {
      "title": "Towards Generalizable Robotic Manipulation in Dynamic Environments",
      "title_zh": "面向动态环境中可泛化的机器人操控",
      "year": 2026,
      "arxiv": "2603.15620v3",
      "source_pdf": "Fang et al. - 2026 - Towards Generalizable Robotic Manipulation in Dynamic Environments.pdf",
      "category": "动态操控基准与运动表征学习",
      "one_sentence_contribution": "构建117,000轨迹的DOMINO，并以历史光流和未来对象特征辅助监督训练PUMA。",
      "main_limit": "主表成功率仅17.20%，采用接触停止自主运动简化，仿真尚未系统纳入真实推理延迟。",
      "official_urls": [
        "https://arxiv.org/abs/2603.15620v3",
        "https://h-embodvis.github.io/DOMINO/",
        "https://github.com/H-EmbodVis/DOMINO",
        "https://github.com/H-EmbodVis/DOMINO/blob/main/policy/PUMA/README.md"
      ],
      "alias": "DOMINO",
      "short_name": "DOMINO / PUMA",
      "source_kind": "项目论文附件",
      "paper_id": "P20",
      "han_characters": 3951,
      "summary_sha256": "1e63c8906da101786006e94f48a816189caa6519e335f4a513a71673df14d3e0",
      "source_sha256": "a565cab8b0195a166686de774611095cf0f315c46cb411d18f32a83509ddaf02",
      "source_pdf_pages": 43,
      "filename": "Towards_Generalizable_Robotic_Manipulation_in_Dynamic_Environments.md"
    },
    {
      "title": "MotionBricks: Scalable Real-Time Motions with Modular Latent Generative Model and Smart Primitives",
      "title_zh": "MotionBricks：通过模块化潜在生成模型与智能基元实现可扩展的实时动作生成",
      "year": 2026,
      "arxiv": "2604.24833v1",
      "source_pdf": "Wang et al. - 2026 - MotionBricks Scalable Real-Time Motions with Modular Latent Generative Model and Smart Primitives.pdf",
      "category": "人体动作生成与交互",
      "one_sentence_contribution": "以根轨迹、姿态离散潜变量和智能控制基元实现低延迟多风格动作生成。",
      "main_limit": "输出运动学参考动作，导航决策与物理跟踪仍需外部系统，完整训练管线未随预览全部公开。",
      "official_urls": [
        "https://arxiv.org/abs/2604.24833v1",
        "https://nvlabs.github.io/motionbricks/",
        "https://github.com/NVlabs/GR00T-WholeBodyControl/tree/main/motionbricks",
        "https://doi.org/10.1145/3811334"
      ],
      "alias": "MOTION",
      "short_name": "MotionBricks",
      "source_kind": "项目论文附件",
      "paper_id": "P21",
      "han_characters": 3456,
      "summary_sha256": "f1b7f491662378fbb4f29f81162c098a33761b610ab6de98d369fc6cf747951a",
      "source_sha256": "44c4baabb047642d8968f53205a0d585bb169f5bbb2de6e6539ea0f0f85a40fc",
      "source_pdf_pages": 22,
      "filename": "MotionBricks-Scalable_Real-Time_Motions_with_Modular_Latent_Generative_Model_and_Smart_Primitives.md"
    },
    {
      "title": "Uni-Inter: Unifying 3D Human Motion Synthesis Across Diverse Interaction Contexts",
      "title_zh": "Uni-Inter：统一不同交互情境下的三维人体动作合成",
      "year": 2025,
      "arxiv": "2511.13032v2",
      "source_pdf": "Liu et al. - 2026 - Uni-Inter Unifying 3D Human Motion Synthesis Across Diverse Interaction Contexts.pdf",
      "category": "人体动作生成与交互",
      "one_sentence_contribution": "用统一三维语义体素条件与关节热图扩散建模人—人、人—物、人—场景交互。",
      "main_limit": "依赖给定交互上下文及未来轨迹，不直接解决因果在线预测、物理可执行性与人群导航。",
      "official_urls": [
        "https://arxiv.org/abs/2511.13032v2",
        "https://github.com/Darkdawner/Uni-Inter",
        "https://doi.org/10.1145/3757377.3763954"
      ],
      "alias": "UNI",
      "short_name": "Uni-Inter",
      "source_kind": "项目论文附件",
      "paper_id": "P22",
      "han_characters": 3229,
      "summary_sha256": "19b0393ab8628ec8cd119eb36810534d5b3f16be5faaff1e38dba977f3c5327a",
      "source_sha256": "8094edf13ae5e32305ee7cb2e1f9fadda7879c6902861fc7862a262228e69877",
      "source_pdf_pages": 11,
      "filename": "Uni-Inter-Unifying_3D_Human_Motion_Synthesis_Across_Diverse_Interaction_Contexts.md"
    },
    {
      "title": "GRAIL: Generating Humanoid Loco-Manipulation from 3D Assets and Video Priors",
      "title_zh": "GRAIL：从三维资产与视频先验生成人形机器人移动操作数据",
      "year": 2026,
      "arxiv": "2606.05160v1",
      "source_pdf": "Xie et al. - GRAIL Generating Humanoid Loco-Manipulation from 3D Assets and Video Priors.pdf",
      "category": "仿真数据生成与人形机器人控制",
      "one_sentence_contribution": "从已知三维资产、视频先验和四维重建生成可经物理策略跟踪的机器人移动操作示范。",
      "main_limit": "依赖已知场景几何、视频筛选及昂贵物理训练，不是任意现实视频的自动无误重建。",
      "official_urls": [
        "https://arxiv.org/abs/2606.05160v1",
        "https://research.nvidia.com/labs/dair/grail/",
        "https://github.com/NVlabs/GRAIL/",
        "https://huggingface.co/datasets/nvidia/PhysicalAI-Robotics-Locomanipulation-GRAIL"
      ],
      "alias": "GRAIL",
      "short_name": "GRAIL",
      "source_kind": "项目论文附件",
      "paper_id": "P23",
      "han_characters": 3507,
      "summary_sha256": "624136e70b3da2d4c0bb6630dd4b41a54e152b52b8c78bc32b80ba90291ae64b",
      "source_sha256": "d13a7769e97a2f472dbba13d383a18e1acacb48c34c3a9472adfaed1ce162a23",
      "source_pdf_pages": 24,
      "filename": "GRAIL-Generating_Humanoid_Loco-Manipulation_from_3D_Assets_and_Video_Priors.md"
    },
    {
      "title": "NavDP: Learning Sim-to-Real Navigation Diffusion Policy with Privileged Information Guidance",
      "title_zh": "NavDP：利用特权信息引导学习从仿真迁移到现实的导航扩散策略",
      "year": 2025,
      "arxiv": "2505.08712v3",
      "source_pdf": "Cai et al. - 2025 - NavDP Learning Sim-to-Real Navigation Diffusion Policy with Privileged Information Guidance.pdf",
      "category": "视觉导航策略与仿真到现实迁移",
      "one_sentence_contribution": "通过规模化仿真路径、RGB-D 扩散候选和特权几何监督的轨迹评估器，实现跨环境与跨机器人局部导航。",
      "main_limit": "主要以静态场景规划数据训练，缺少完整社会规范与长期记忆；论文的部分 critic 与训练公式还存在符号不清或排版缺项。",
      "official_urls": [
        "https://arxiv.org/abs/2505.08712",
        "https://github.com/InternRobotics/NavDP",
        "https://wzcai99.github.io/navigation-diffusion-policy.github.io/"
      ],
      "alias": "NAVDP",
      "short_name": "NavDP",
      "source_kind": "项目论文附件",
      "paper_id": "P24",
      "han_characters": 3486,
      "summary_sha256": "0d6de4fef2063b3a8da24c32e3c59e37f3e48b181740b4d6c04e21ae4c8f086f",
      "source_sha256": "ffc54542c81f7c369803a2a307927bed7f8a9581c4e45aeff6b57dee728ae22d",
      "source_pdf_pages": 8,
      "filename": "NavDP-Learning_Sim-to-Real_Navigation_Diffusion_Policy_with_Privileged_Information_Guidance.md"
    },
    {
      "title": "FLUX: Accelerating Cross-Embodiment Generative Navigation Policies via Rectified Flow and Static-to-Dynamic Learning",
      "title_zh": "FLUX：通过 Rectified Flow 与从静态到动态的学习加速跨具身生成式导航策略",
      "year": 2026,
      "arxiv": "2603.12806v1",
      "source_pdf": "2603.12806v1.pdf",
      "category": "生成式导航与动态人群强化学习",
      "one_sentence_contribution": "以 NavDP 初始化的 Rectified Flow 和动态环境强化学习改进跨具身生成式局部导航，并提供 DynBench 六类动态场景任务。",
      "main_limit": "部分奖励、概率比和社会指标定义不充分，距离阈值代理不等同物理碰撞，个别表项出现 SPL 高于成功率的内部不一致。",
      "official_urls": [
        "https://arxiv.org/abs/2603.12806",
        "https://zeying-gong.github.io/projects/flux/",
        "https://github.com/Zeying-Gong/FLUX",
        "https://github.com/Zeying-Gong/IsaacLab"
      ],
      "alias": "FLUX",
      "short_name": "FLUX",
      "source_kind": "项目论文附件",
      "paper_id": "P25",
      "han_characters": 3345,
      "summary_sha256": "f05e78804802b27d78b6cee0af52eacf630a59f8f26395e4a8f1cfd645e18bd1",
      "source_sha256": "49498848061149caa149ce34309981bfe1b82ada076bc8f17d52b7dea88e9b23",
      "source_pdf_pages": 9,
      "filename": "FLUX-Accelerating_Cross-Embodiment_Generative_Navigation_Policies_via_Rectified_Flow_and_Static-to-Dynamic_Learning.md"
    },
    {
      "title": "NavIsaacLab: Generating Realistic Crowd via Parallel Robot Learning for Benchmarking Human-aware Navigation",
      "title_zh": "NavIsaacLab：通过并行机器人学习生成真实感人群，用于人类感知导航基准评估",
      "year": 2026,
      "arxiv": "2606.26265v1",
      "source_pdf": "Xia et al. - 2026 - NavIsaacLab Generating Realistic Crowd via Parallel Robot Learning for Benchmarking Human-aware Nav.pdf",
      "category": "Isaac Lab 社交导航仿真与物理人群",
      "one_sentence_contribution": "结合场景处理、人群根轨迹生成、AMP 全身控制和视觉 PPO，使物理驱动行人能用于并行导航数据与仿真到现实评测。",
      "main_limit": "人群控制仍有较多跟踪失败；论文详细实验主要检验人体轨迹与吞吐量，公开 2.0 管线与论文 AMP 方案也需做版本区分。",
      "official_urls": [
        "https://arxiv.org/abs/2606.26265",
        "https://github.com/yh83305/NavIsaaclab2.0",
        "https://broln7.github.io/NavIsaacLab-web/",
        "https://github.com/yh83305/NavIsaaclab2.0/blob/main/FULL_PIPELINE.md"
      ],
      "alias": "NAVISAAC",
      "short_name": "NavIsaacLab",
      "source_kind": "项目论文附件",
      "paper_id": "P26",
      "han_characters": 3859,
      "summary_sha256": "219dbabb068221b09347ec2e66e391e772383b771d92732905278c7be91c8df7",
      "source_sha256": "6ec21b6608c81db75e646f1729485a5677c64fa491a1e59e3b650f62bae8b6d7",
      "source_pdf_pages": 12,
      "filename": "NavIsaacLab-Generating_Realistic_Crowd_via_Parallel_Robot_Learning_for_Benchmarking_Human-aware_Navigation.md"
    },
    {
      "title": "Arena-Bench 2.0: A Comprehensive Benchmark of Social Navigation Approaches in Collaborative Environments",
      "title_zh": "Arena-Bench 2.0：面向协作环境中社交导航方法的综合基准",
      "year": 2025,
      "arxiv": null,
      "source_pdf": "Shcherbyna et al. - 2025 - Arena-Bench 2.0 A Comprehensive Benchmark of Social Navigation Approaches in Collaborative Environm.pdf",
      "category": "ROS 2 社交导航平台与评测协议",
      "one_sentence_contribution": "以 nav2py 连接 Python 学习策略和 ROS 2 Nav2，结合语义场景、交互行为和多维指标开展可配置社交导航比较。",
      "main_limit": "其成功率允许少于两次碰撞，不等同零碰撞到达；算法分类与部分指标单位有误，统计复现细节有限。",
      "official_urls": [
        "https://doi.org/10.1109/IROS60139.2025.11246895",
        "https://ieeexplore.ieee.org/document/11246895/",
        "https://github.com/Arena-Rosnav",
        "https://github.com/Arena-Rosnav/nav2py_drlvo"
      ],
      "alias": "ARENA",
      "short_name": "Arena-Bench 2.0",
      "source_kind": "项目论文附件",
      "paper_id": "P27",
      "han_characters": 3755,
      "summary_sha256": "7e2df710a7bded19396939374b55b7a1191646e6648944b728f768d3447c396f",
      "source_sha256": "e668450523a07d596be4a08a6b5cb80e86bedee0e3cfa71a328d2689ae5b78e9",
      "source_pdf_pages": 8,
      "filename": "Arena-Bench_2.0-A_Comprehensive_Benchmark_of_Social_Navigation_Approaches_in_Collaborative_Environments.md"
    },
    {
      "title": "Social robot navigation: a review and benchmarking of learning-based methods",
      "title_zh": "社交机器人导航：基于学习的方法综述与基准评测",
      "year": 2025,
      "arxiv": null,
      "source_pdf": "Alyassi et al. - 2025 - Social robot navigation a review and benchmarking of learning-based methods.pdf",
      "category": "社交导航方法综述与统一基准",
      "one_sentence_contribution": "用五类架构梳理社交导航、训练与感知系统，并在六类场景中比较九种经典及学习式规划器实现。",
      "main_limit": "统一实验对原方法有多项结构改造，缺乏充分等预算消融与精确数表；官方 benchmark 仓库本次核查仍为 Coming Soon 预告。",
      "official_urls": [
        "https://www.frontiersin.org/journals/robotics-and-ai/articles/10.3389/frobt.2025.1658643/full",
        "https://socialnavigation.github.io/",
        "https://github.com/ralyassi/SocialNav",
        "https://github.com/ralyassi/awesome-social-navigation"
      ],
      "alias": "SOCIAL",
      "short_name": "社交导航综述",
      "source_kind": "项目论文附件",
      "paper_id": "P28",
      "han_characters": 5772,
      "summary_sha256": "874000b89ec844950732768fd8900be88a0f3c47985911332afd0d75ae800ccc",
      "source_sha256": "6cece8928d86567c0d3ccc1bf6405a433866dd8091685ecaf973ed6cbcb68566",
      "source_pdf_pages": 34,
      "filename": "Social_robot_navigation-a_review_and_benchmarking_of_learning-based_methods.md"
    },
    {
      "title": "Towards Physically Executable 3D Gaussian for Embodied Navigation",
      "title_zh": "面向具身导航的物理可执行三维高斯场景",
      "year": 2025,
      "arxiv": "2510.21307v2",
      "source_pdf": "2510.21307v2.pdf",
      "category": "三维高斯场景与导航数据基准",
      "one_sentence_contribution": "将按对象分解的语义 3DGS 与原网格的凸分解碰撞体组合，提供 InteriorGS-1k 和 SAGE-Bench 的连续导航数据与评估。",
      "main_limit": "物理层依赖已有高质量三维网格，不是任意图像生成可靠碰撞几何；部分指标和表格存在定义或数值不一致。",
      "official_urls": [
        "https://arxiv.org/abs/2510.21307",
        "https://sage-3d.github.io/",
        "https://github.com/Galery23/SAGE-3D_Official",
        "https://huggingface.co/datasets/spatialverse/SAGE-3D_Collision_Mesh"
      ],
      "alias": "SAGE",
      "short_name": "SAGE-3D",
      "source_kind": "项目论文附件",
      "paper_id": "P29",
      "han_characters": 3349,
      "summary_sha256": "cfaa934f817608f36e81e9c429dbd90561e986e8b2673395dec5441e92969ed3",
      "source_sha256": "90166752922b5c64bc7accebc7d772109e2400d494e5589a4e3c508a8808f43d",
      "source_pdf_pages": 20,
      "filename": "Towards_Physically_Executable_3D_Gaussian_for_Embodied_Navigation.md"
    },
    {
      "title": "Lyra 2.0: Explorable Generative 3D Worlds",
      "title_zh": "Lyra 2.0：可探索的生成式三维世界",
      "year": 2026,
      "arxiv": "2604.13036v1",
      "source_pdf": "Shen et al. - 2026 - Lyra 2.0 Explorable Generative 3D Worlds.pdf",
      "category": "长期空间记忆与生成式三维世界",
      "one_sentence_contribution": "通过逐帧几何检索、规范坐标对应和自增强历史条件延长视频世界探索，再将生成视图重建为 Gaussian 与网格。",
      "main_limit": "当前主要生成静态场景，未验证严谨的机器人接触碰撞几何；未知区域为生成内容且训练、推理算力要求较高。",
      "official_urls": [
        "https://arxiv.org/abs/2604.13036",
        "https://research.nvidia.com/labs/sil/lyra2/",
        "https://nv-tlabs.github.io/Project-Lyra/",
        "https://github.com/nv-tlabs/lyra",
        "https://huggingface.co/nvidia/Lyra-2.0"
      ],
      "alias": "LYRA",
      "short_name": "Lyra 2.0",
      "source_kind": "项目论文附件",
      "paper_id": "P30",
      "han_characters": 3762,
      "summary_sha256": "76121bfa96d5666cbbff77ed01de29cb0b21ebc0724815b4e4d569bbe9f3a572",
      "source_sha256": "d9de7483af288bf492d9b3ce0b7c5276db978518a22d720b9c686e68861db587",
      "source_pdf_pages": 25,
      "filename": "Lyra_2.0-Explorable_Generative_3D_Worlds.md"
    }
  ]
}