{
  "schema_version": "paper_public_manifest_v1",
  "paper_id": "roboharn-evo_2026_10",
  "slug": "roboharn-evo",
  "title": "RoboHarn-Evo：知识层RSI",
  "authors": [],
  "source": {
    "arxiv_id": "2609.37583",
    "pdf_url": "https://arxiv.org/pdf/2609.37583",
    "project_url": "",
    "github_url": "",
    "huggingface_url": "",
    "original_source": "https://arxiv.org/pdf/2609.37583"
  },
  "site": {
    "post_url": "/posts/roboharn-evo",
    "canonical_url": "https://haiguangboy.com/posts/roboharn-evo",
    "cover_image": "https://static.haiguangboy.com/papers/roboharn-evo/cover.webp"
  },
  "taxonomy": {
    "domain": "embodied_ai",
    "track": "action_generation",
    "tasks": [
      "embodied_ai",
      "action_generation",
      "robotics",
      "state_prediction",
      "具身智能",
      "机器人",
      "rsi",
      "自我改进",
      "roboharnevo"
    ],
    "related_topics": [
      {
        "paper_id": "transformer-transformer_blog_transformer_transformer_a_unified_20260807_2026_08",
        "title": "Transformer Transformer: A Unified Model for Motion-Conditioned Robot Co-design",
        "url": "https://haiguangboy.com/posts/transformer-transformer-codesign",
        "relation": "contrast",
        "summary": "固定模型与执行器，更新任务—动作两层知识 contradicts 路线主张：任务失败不应默认归因于策略，本体—控制器—策略应进入同一个优化问题",
        "strength": "strong"
      },
      {
        "paper_id": "latepost_xuhuazhe_202603_2026_03",
        "title": "latepost_xuhuazhe_202603",
        "url": "https://haiguangboy.com/posts/latepost_xuhuazhe_202603",
        "relation": "contrast",
        "summary": "固定模型与执行器，更新任务—动作两层知识 contradicts 路线赌注：行为/动作部分必须是统一模型，反对模块化拼接",
        "strength": "strong"
      },
      {
        "paper_id": "generalist_blog_gen_1_5_20260818_2026_08",
        "title": "机器人GPT-3时刻：故事很足，证据不够",
        "url": "https://haiguangboy.com/posts/gen-1-5-one-shot",
        "relation": "contrast",
        "summary": "自我改进限定在知识内容与检索资格，不是模型或控制器自我重写 contradicts 核心叙事(公司自述,未经第三方验证):这些能力直接从规模化预训练涌现,不需要任何专门机制设计",
        "strength": "strong"
      },
      {
        "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": "真机仅三任务各8场景，π0.5基线未作真机微调 validates 小样本及选优评估不能外推为稳定通用能力",
        "strength": "strong"
      },
      {
        "paper_id": "an_open_foundation_model_towards_2026_07",
        "title": "An_Open_Foundation_Model_Towards",
        "url": "https://haiguangboy.com/posts/an_open_foundation_model_towards",
        "relation": "same_track",
        "summary": "任务与动作知识组合优于各自单用 validates MM-DiT动作专家：联合全局注意力融合优于朴素DiT",
        "strength": "strong"
      },
      {
        "paper_id": "latent_action_pretraining_through_world_modeling_2026_07",
        "title": "LAWM：为什么动作标签会成为负担",
        "url": "https://haiguangboy.com/posts/latent_action_pretraining_through_world_modeling",
        "relation": "same_track",
        "summary": "LAWM：为什么动作标签会成为负担",
        "strength": "strong"
      }
    ]
  },
  "analyst_take": {
    "type": "author_opinion",
    "text": "💡 与其他改进路线对照\nPAVE按动作块计算优势，避免把失败任务里的补救动作一起判坏，但每轮仍需重训策略。本篇则把经验收益写入外部知识。\n\nSmoothRL要求优化对齐真正执行的动作，本篇要求证据对应真正执行的策略。共同问题是：失败之后，系统究竟应该改哪里，才能让下一次执行更可靠。"
  },
  "ruling": {
    "importance_score": 3.0,
    "one_sentence": "RoboHarn-Evo：知识层RSI"
  },
  "asset_base_url": "https://static.haiguangboy.com/papers/roboharn-evo",
  "assets": [
    {
      "type": "pdf_screenshot",
      "object_key": "papers/roboharn-evo/page_01.webp",
      "content_type": "image/webp",
      "upload_status": "uploaded",
      "bucket": "paper-assets",
      "url": "https://static.haiguangboy.com/papers/roboharn-evo/page_01.webp",
      "role": "paper_first_page",
      "size_bytes": 137840
    },
    {
      "type": "pdf_screenshot",
      "object_key": "papers/roboharn-evo/key_figure.webp",
      "content_type": "image/webp",
      "upload_status": "uploaded",
      "bucket": "paper-assets",
      "url": "https://static.haiguangboy.com/papers/roboharn-evo/key_figure.webp",
      "role": "method_figure",
      "size_bytes": 147656
    },
    {
      "type": "cover_image",
      "object_key": "papers/roboharn-evo/cover.webp",
      "content_type": "image/webp",
      "upload_status": "uploaded",
      "bucket": "paper-assets",
      "url": "https://static.haiguangboy.com/papers/roboharn-evo/cover.webp",
      "role": "post_cover",
      "size_bytes": 55416
    },
    {
      "type": "public_brief",
      "object_key": "papers/roboharn-evo/public_brief.md",
      "content_type": "text/markdown; charset=utf-8",
      "upload_status": "uploaded",
      "bucket": "paper-assets",
      "url": "https://static.haiguangboy.com/papers/roboharn-evo/public_brief.md",
      "role": "public_brief",
      "size_bytes": 3845
    },
    {
      "type": "public_brief",
      "object_key": "papers/roboharn-evo/public_brief.en.md",
      "content_type": "text/markdown; charset=utf-8",
      "upload_status": "uploaded",
      "bucket": "paper-assets",
      "url": "https://static.haiguangboy.com/papers/roboharn-evo/public_brief.en.md",
      "role": "public_brief_en",
      "size_bytes": 4902
    },
    {
      "type": "public_manifest",
      "object_key": "papers/roboharn-evo/public_manifest.en.json",
      "content_type": "application/json; charset=utf-8",
      "upload_status": "uploaded",
      "bucket": "paper-assets",
      "url": "https://static.haiguangboy.com/papers/roboharn-evo/public_manifest.en.json",
      "role": "public_manifest_en",
      "size_bytes": 6476
    },
    {
      "type": "public_manifest",
      "object_key": "papers/roboharn-evo/public_manifest.json",
      "content_type": "application/json; charset=utf-8",
      "upload_status": "uploaded",
      "bucket": "paper-assets",
      "url": "https://static.haiguangboy.com/papers/roboharn-evo/public_manifest.json",
      "role": "public_manifest",
      "size_bytes": 7041
    }
  ],
  "published_at": "2026-10-13T16:19:52+08:00",
  "created_at": "2026-10-13T16:19:52+08:00",
  "updated_at": "2026-10-13T16:20:04+08:00"
}