{
  "schema_version": "paper_public_manifest_v1",
  "paper_id": "foveavla_2026_10",
  "slug": "foveavla",
  "title": "不装腕相机，VLA成功率反而更高",
  "authors": [],
  "source": {
    "arxiv_id": "2610.20064",
    "pdf_url": "https://arxiv.org/pdf/2610.20064",
    "project_url": "",
    "github_url": "",
    "huggingface_url": "",
    "original_source": "https://arxiv.org/pdf/2610.20064"
  },
  "site": {
    "post_url": "/posts/foveavla",
    "canonical_url": "https://haiguangboy.com/posts/foveavla",
    "cover_image": "https://static.haiguangboy.com/papers/foveavla/cover.webp"
  },
  "taxonomy": {
    "domain": "embodied_ai",
    "track": "vla",
    "tasks": [
      "embodied_ai",
      "vla",
      "action_generation",
      "robotics",
      "具身智能",
      "机器人",
      "foveavla",
      "主动视觉"
    ],
    "related_topics": [
      {
        "paper_id": "agile_wam_an_agile_tactile_world_action_model_for_contact_rich_robot_control_2026_09",
        "title": "Agile-WAM：轻量视触觉联合预测",
        "url": "https://haiguangboy.com/posts/agile-wam",
        "relation": "same_track",
        "summary": "48次真机完整成功率不能外推工业可靠性 validates 小样本及选优评估不能外推为稳定通用能力",
        "strength": "strong"
      },
      {
        "paper_id": "d4rt_efficiently_reconstructing_dynamic_scenes_one_2026_06",
        "title": "D4RT_Efficiently_Reconstructing_Dynamic_Scenes_One",
        "url": "https://haiguangboy.com/posts/d4rt_efficiently_reconstructing_dynamic_scenes_one",
        "relation": "same_track",
        "summary": "从原分辨率裁剪，同时保留全景 validates 可处理in-the-wild视频,连续query坐标支持任意/亚像素分辨率解码",
        "strength": "strong"
      },
      {
        "paper_id": "dextouch_wm_learning_action_conditioned_tactile_world_models_from_human_touch_fo_2026_09",
        "title": "DexTouch-WM：人类触觉预训练",
        "url": "https://haiguangboy.com/posts/dextouch-wm",
        "relation": "same_track",
        "summary": "学习人类无腕视频仍是展望 validates 人类数据不是无硬件、无机器人数据的直接迁移",
        "strength": "strong"
      },
      {
        "paper_id": "google_blog_gemini_robotics_er_2_20260730_2026_07",
        "title": "Introducing Gemini Robotics ER 2",
        "url": "https://haiguangboy.com/posts/google_blog_gemini_robotics_er_2_20260730",
        "relation": "same_track",
        "summary": "零样本排行榜对照不是受控比较 validates 数字缺对照:未点名竞品、未给基线数值、对比图表不可核",
        "strength": "strong"
      },
      {
        "paper_id": "mt_wam_reorienting_the_one_pass_predictive_representation_toward_action_generati_2026_09",
        "title": "MT-WAM：未来监督如何服务动作",
        "url": "https://haiguangboy.com/posts/mt-wam",
        "relation": "same_track",
        "summary": "主结果不是仅删相机或仅加裁剪的单变量实验 validates 尚未完全分离监督目标、容量和分支结构的因果贡献",
        "strength": "strong"
      },
      {
        "paper_id": "smoothrl_online_reinforcement_learning_during_asynchronous_execution_2026_09",
        "title": "SmoothRL:只对真正执行的动作算账",
        "url": "https://haiguangboy.com/posts/smoothrl_online_reinforcement_learning_during_asynchronous_execution",
        "relation": "same_track",
        "summary": "SmoothRL:只对真正执行的动作算账",
        "strength": "strong"
      }
    ]
  },
  "analyst_take": {
    "type": "author_opinion",
    "text": "💡 放回预测与控制的技术脉络\nMT-WAM提醒：预测任务有用，不意味着预测token都该直接成为动作条件。FoveaVLA提供另一种接法：预测结果改变下一次收集的视觉细节。\n\n值得检验的是，在遮挡、相机位置变化和更大工作空间中，这种“预测—观测—动作”循环能否持续保住收益，而不只是把目标框预测得更准。"
  },
  "ruling": {
    "importance_score": 3.0,
    "one_sentence": "不装腕相机，VLA成功率反而更高"
  },
  "asset_base_url": "https://static.haiguangboy.com/papers/foveavla",
  "assets": [
    {
      "type": "pdf_screenshot",
      "object_key": "papers/foveavla/page_01.webp",
      "content_type": "image/webp",
      "upload_status": "uploaded",
      "bucket": "paper-assets",
      "url": "https://static.haiguangboy.com/papers/foveavla/page_01.webp",
      "role": "paper_first_page",
      "size_bytes": 206026
    },
    {
      "type": "pdf_screenshot",
      "object_key": "papers/foveavla/key_figure.webp",
      "content_type": "image/webp",
      "upload_status": "uploaded",
      "bucket": "paper-assets",
      "url": "https://static.haiguangboy.com/papers/foveavla/key_figure.webp",
      "role": "method_figure",
      "size_bytes": 166234
    },
    {
      "type": "cover_image",
      "object_key": "papers/foveavla/cover.webp",
      "content_type": "image/webp",
      "upload_status": "uploaded",
      "bucket": "paper-assets",
      "url": "https://static.haiguangboy.com/papers/foveavla/cover.webp",
      "role": "post_cover",
      "size_bytes": 65152
    },
    {
      "type": "public_brief",
      "object_key": "papers/foveavla/public_brief.md",
      "content_type": "text/markdown; charset=utf-8",
      "upload_status": "uploaded",
      "bucket": "paper-assets",
      "url": "https://static.haiguangboy.com/papers/foveavla/public_brief.md",
      "role": "public_brief",
      "size_bytes": 3810
    },
    {
      "type": "public_brief",
      "object_key": "papers/foveavla/public_brief.en.md",
      "content_type": "text/markdown; charset=utf-8",
      "upload_status": "uploaded",
      "bucket": "paper-assets",
      "url": "https://static.haiguangboy.com/papers/foveavla/public_brief.en.md",
      "role": "public_brief_en",
      "size_bytes": 4798
    },
    {
      "type": "public_manifest",
      "object_key": "papers/foveavla/public_manifest.en.json",
      "content_type": "application/json; charset=utf-8",
      "upload_status": "uploaded",
      "bucket": "paper-assets",
      "url": "https://static.haiguangboy.com/papers/foveavla/public_manifest.en.json",
      "role": "public_manifest_en",
      "size_bytes": 6399
    },
    {
      "type": "public_manifest",
      "object_key": "papers/foveavla/public_manifest.json",
      "content_type": "application/json; charset=utf-8",
      "upload_status": "uploaded",
      "bucket": "paper-assets",
      "url": "https://static.haiguangboy.com/papers/foveavla/public_manifest.json",
      "role": "public_manifest",
      "size_bytes": 6895
    }
  ],
  "published_at": "2026-10-17T19:17:13+08:00",
  "created_at": "2026-10-17T19:17:13+08:00",
  "updated_at": "2026-10-17T19:17:23+08:00"
}