知识库分析报告（analytics）
自动生成：python3 scripts/analyze_kb.py。基于 papers.jsonl（975 篇）+ papers_pdf/（746 篇本地全文）+ papers/terse/P-*（43 篇深笔记）。
本报告是知识库「分析层」的体检，用于找：关键影响力工作、元数据缺口、引用图断链、交叉学习富矿、待补遗漏。
1. 概览
- 论文总数：975
- 开源：375（38%）
- 时间跨度：1988–2026
学习仪表盘
面向学习者的入口：每条线索的浅入口密度、推荐起点、领域图谱摘要、进阶学习路径。数据来自 site/content/*/.md（浅入口）+ clusters.jsonl / bridges.jsonl / influence.jsonl（图谱摘要）。
浅入口覆盖（按线索）
已发布入门教学页合计 294 篇，分布在 14 条线索下。入门页=一句话直觉 + 五分钟读懂，是最易切入的浅入口；该数高 = 这条线索的学习素材更密。
| 线索 | 浅入口 | 入口 |
|---|---|---|
| T01 | 18 | T01 |
| T02 | 22 | T02 |
| T03 | 6 | T03 |
| T04 | 34 | T04 |
| T05 | 23 | T05 |
| T06 | 65 | T06 |
| T07 | 16 | T07 |
| T08 | 20 | T08 |
| T09 | 16 | T09 |
| T10 | 10 | T10 |
| T11 | 36 | T11 |
| T12 | 15 | T12 |
| T13 | 6 | T13 |
| T14 | 7 | T14 |
推荐起点（浅入口覆盖最高的 3 条线索）
- T06 — 65 篇浅入口
- T11 — 36 篇浅入口
- T04 — 34 篇浅入口
知识图谱摘要
5 个研究群落（label-propagation + 模块度合并；覆盖库内 754 篇）：
| 群落 | 规模 | 主导方向 | 主题 |
|---|---|---|---|
| C0 | 265 | Foundation | 视觉-语言-动作模型（VLA） |
| C1 | 202 | Motion Prior | 人形机器人 RL / 扩散模型 / 动作跟踪 |
| C2 | 134 | Foundation | 世界模型 / 极限地形/parkour / locomotion |
| C3 | 87 | Action Gen | 动作重定向/映射 / 扩散模型 |
| C4 | 66 | Dexterous | 灵巧手内操作 |
桥接工作 Top-5（邻居跨 ≥2 群落；最值得深挖的交叉节点）：
| 工作 | 桥接度 | span | 方向 |
|---|---|---|---|
| Proximal Policy Optimization Algorithms (PPO) | 4.9 | 4 | Locomotion |
| RMA: Rapid Motor Adaptation for Legged Robots | 4.5 | 4 | Locomotion |
| Isaac Gym: High Performance GPU-Based Physics Simulation For | 4.3 | 4 | Locomotion |
| AMASS: Archive of Motion Capture as Surface Shapes | 4.0 | 3 | Motion Prior |
| GR00T N1: An Open Foundation Model for Generalist Humanoid R | 3.9 | 3 | Foundation |
影响力论文 Top-5（按 PageRank）：
| 工作 | 方向 | in-deg | PageRank |
|---|---|---|---|
| AMP: Adversarial Motion Priors for Stylized Physics-Based Ch | Action Gen | 17 | 0.0099 |
| Extreme Parkour with Legged Robots | Locomotion | 13 | 0.0065 |
| BeyondMimic: From Motion Tracking to Versatile Humanoid Cont | Action Gen | 14 | 0.0064 |
| ASE: Large-Scale Reusable Adversarial Skill Embeddings for P | Action Gen | 15 | 0.0059 |
| ThriftyDAgger: Budget-aware novelty and risk gating for inte | Imitation | 3 | 0.0055 |
领域分布速览
论文按方向 / 按年份的分布（与 §6 覆盖度同源数据，这里以视觉形式给出学习者一个直觉）。
学习路径入口
- 学习指南 — 贯穿全站的阅读顺序（线索 + 概念 + 论文）
- SOTA 与开放问题 — 各方向当前最好结果 + 未解问题
- 术语表 — 跨线索术语对照（BC / MPC / DR / VLA …）
- 全景 roadmap — 领域总图与方向关系
- 引用图 — 论文级可视化（含桥接工作高亮）
2. 元数据完整度
| 字段 | key | 填充 | 占比 |
|---|---|---|---|
| 标题 | title | 975/975 | 100% |
| 作者 | authors | 689/975 | 71% |
| 机构 | org | 655/975 | 67% |
| 年份 | year | 975/975 | 100% |
| venue | venue | 967/975 | 99% |
| arxiv id | arxiv | 787/975 | 81% |
| 代码 | code | 225/975 | 23% |
| 项目页 | project | 204/975 | 21% |
| 开源 | open_source | 975/975 | 100% |
| 方向 | category | 975/975 | 100% |
| 子方向 | subfield | 670/975 | 69% |
| 中文贡献 | contribution_zh | 975/975 | 100% |
| ⭐为何重要 | why_important_zh | 967/975 | 99% |
| builds_on | builds_on | 948/975 | 97% |
| leads_to | leads_to | 633/975 | 65% |
| 输入模态 | input_modality | 254/975 | 26% |
| 模型规模 | model_size | 42/975 | 4% |
| 上真机 | on_robot | 282/975 | 29% |
| 物理感知 | physics_aware | 282/975 | 29% |
| 实时 | real_time | 198/975 | 20% |
| 机器人 | robot | 184/975 | 19% |
| 数据集 | dataset | 181/975 | 19% |
薄弱字段（<50%，按优先级）：
- 模型规模 (
model_size)：4% — 几乎为空 - 数据集 (
dataset)：19% — 部分填充（结构性受限） - 机器人 (
robot)：19% — 部分填充（结构性受限） - 实时 (
real_time)：20% — 部分填充（结构性受限） - 项目页 (
project)：21% — 部分填充（结构性受限） - 代码 (
code)：23% — 部分填充（结构性受限） - 输入模态 (
input_modality)：26% — 部分填充（结构性受限） - 上真机 (
on_robot)：29% — 部分填充（结构性受限） - 物理感知 (
physics_aware)：29% — 部分填充（结构性受限）
3. 影响力工作榜
3.1 被「builds_on」最多（影响力 Top 20）
次数 = 多少篇后续工作显式声明建立在该工作之上。
| 被引次数 | 工作 | 方向 | 深笔记 | 本地PDF |
|---|---|---|---|---|
| 60 | OpenVLA: An Open-Source Vision-Language-Action Model | Foundation | ✓ | ✓ |
| 39 | Diffusion Policy: Visuomotor Policy Learning via Action Diffusion | Foundation | ✗ | ✓ |
| 37 | Open X-Embodiment: Robotic Learning Datasets and RT-X Models | Foundation | ✓ | ✓ |
| 36 | π0: A Vision-Language-Action Flow Model for General Robot Control | Foundation | ✓ | ✓ |
| 35 | RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control | Foundation | ✓ | ✓ |
| 30 | AMP: Adversarial Motion Priors for Stylized Physics-Based Character Control | Action Gen | ✗ | ✓ |
| 30 | OmniH2O: Universal and Dexterous Human-to-Humanoid Whole-Body Teleoperation and Learning | Action Gen | ✗ | ✓ |
| 27 | DeepMimic: Example-Guided Deep Reinforcement Learning of Physics-Based Character Skills | Locomotion | ✗ | ✓ |
| 27 | HOVER: Versatile Neural Whole-Body Controller for Humanoid Robots | Locomotion | ✗ | ✓ |
| 24 | Learning Transferable Visual Models From Natural Language Supervision (CLIP) | Foundation | ✗ | ✓ |
| 24 | Expressive Whole-Body Control for Humanoid Robots (ExBody) | Action Gen | ✗ | ✓ |
| 23 | Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware (ACT / ALOHA) | Foundation | ✗ | ✓ |
| 21 | Perpetual Humanoid Control for Real-time Simulated Avatars | Action Gen | ✗ | ✓ |
| 20 | Attention Is All You Need (Transformer) | Foundation | ✗ | ✓ |
| 19 | Human Motion Diffusion Model (MDM) | Action Gen | ✗ | ✓ |
| 17 | Auto-Encoding Variational Bayes (VAE) | Foundation | ✗ | ✓ |
| 16 | World Models | Foundation | ✓ | ✓ |
| 15 | Proximal Policy Optimization Algorithms (PPO) | Locomotion | ✗ | ✓ |
| 15 | AMASS: Archive of Motion Capture as Surface Shapes | Motion Prior | ✗ | ✓ |
| 14 | Denoising Diffusion Probabilistic Models (DDPM) | Foundation | ✗ | ✓ |
3.2 「leads_to」最多（催生后续工作最多 Top 12）
| 催生数 | 工作 | 方向 |
|---|---|---|
| 7 | AMP: Adversarial Motion Priors for Stylized Physics-Based Character Control | Action Gen |
| 6 | Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware (ACT / ALOHA) | Foundation |
| 5 | Generative Adversarial Nets (GAN) | Foundation |
| 5 | SAPIEN: A SimulAted Part-based Interactive ENvironment | Foundation |
| 5 | Denoising Diffusion Probabilistic Models (DDPM) | Foundation |
| 5 | Learning to Walk in Minutes Using Massively Parallel Deep Reinforcement Learning | Locomotion |
| 5 | Video PreTraining (VPT): Learning to Act by Watching Unlabeled Online Videos | Imitation |
| 4 | Animation of Dynamic Legged Locomotion | Motion Prior |
| 4 | Auto-Encoding Variational Bayes (VAE) | Foundation |
| 4 | SMPL: A Skinned Multi-person Linear Model | Motion Prior |
| 4 | One-Shot Imitation Learning | Imitation |
| 4 | Constrained Policy Optimization | Locomotion |
3.3 高影响力但尚无深笔记（待补 terse/P-*）
- （被引 39）Diffusion Policy: Visuomotor Policy Learning via Action Diffusion
- （被引 30）AMP: Adversarial Motion Priors for Stylized Physics-Based Character Control
- （被引 30）OmniH2O: Universal and Dexterous Human-to-Humanoid Whole-Body Teleoperation and Learning
- （被引 27）DeepMimic: Example-Guided Deep Reinforcement Learning of Physics-Based Character Skills
- （被引 27）HOVER: Versatile Neural Whole-Body Controller for Humanoid Robots
- （被引 24）Learning Transferable Visual Models From Natural Language Supervision (CLIP)
- （被引 24）Expressive Whole-Body Control for Humanoid Robots (ExBody)
- （被引 23）Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware (ACT / ALOHA)
- （被引 21）Perpetual Humanoid Control for Real-time Simulated Avatars
- （被引 20）Attention Is All You Need (Transformer)
4. 引用图健康
- 有 builds_on 的论文：948/975（97%）
- 有 leads_to 的论文：633/975（65%）
- 孤立节点（无入无出）：8（1%）
- 引用标签总数：3742，其中能解析到库内论文：2238
- 解析率（原始 = resolved/total）：59.8%
- 解析率（诚实，排除概念标签）：2238/2489 = 89.9% ← 与
validate_citations.py门同源 - 未解析标签：1504 occ / 1237 unique （其中概念 1253，真实漏收录候选 251）
未解析标签按 kb_common.classify 二分类：
- concept —— BC/MPC/DR/video diffusion 等概念/方向/平台标签，不是单篇论文，排除出解析率分母（这才是诚实口径，与门判定一致）。
- suspected_missing —— 疑似漏收录的具体论文（proper-noun 标题、版本化变体、具名数据集），属于真正的断链/漏收录，由工程 agent 据下表查补。
4.1 真实漏收录（suspected_missing，按出现次数 Top 20）
| 出现次数 | 标签 | 分类 |
|---|---|---|
| 3 | PULSE | suspected_missing |
| 2 | Allegro Hand | suspected_missing |
| 2 | CPO | suspected_missing |
| 2 | DALL-E 2 | suspected_missing |
| 2 | HM3D | suspected_missing |
| 2 | LIBERO-PRO | suspected_missing |
| 2 | Learning Complex Dexterous Manipulation with Deep Reinforcement Learning and Demonstrations | suspected_missing |
| 2 | VPP | suspected_missing |
| 2 | YOLOv5 | suspected_missing |
| 1 | 3D Diffuser Actor | suspected_missing |
| 1 | 6-DOF AUV control | suspected_missing |
| 1 | Action chunking (ACT) | suspected_missing |
| 1 | Action chunking (ACT) for reactive robot control | suspected_missing |
| 1 | Agent57 | suspected_missing |
| 1 | AlphaStar | suspected_missing |
| 1 | Ambient Diffusion (theory) | suspected_missing |
| 1 | AnyRotate | suspected_missing |
| 1 | AnySkin | suspected_missing |
| 1 | Apple Vision Pro hand tracking | suspected_missing |
| 1 | Asymmetric actor-critic PPO for dexterous manipulation | suspected_missing |
4.2 概念标签（concept，按出现次数 Top 20；非漏收录）
| 出现次数 | 标签 | 分类 |
|---|---|---|
| 26 | Behavior Cloning (BC) | concept |
| 14 | VLA | concept |
| 14 | deep RL | concept |
| 13 | reinforcement learning | concept |
| 12 | deep learning | concept |
| 7 | Diffusion models | concept |
| 7 | domain randomization | concept |
| 7 | language models | concept |
| 7 | sim2real | concept |
| 7 | tactile sensing | concept |
| 6 | BC | concept |
| 6 | ExBody2: Advanced Expressive Humanoid Whole-Body Control | concept |
| 6 | VLA models | concept |
| 6 | video diffusion | concept |
| 5 | ANYmal-blind-locomotion | concept |
| 5 | Contrastive learning | concept |
| 5 | Hierarchical IL | concept |
| 5 | MPC | concept |
| 5 | actor-critic | concept |
| 4 | Policy Gradient | concept |
5. 跨方向引用（交叉学习机会）
行方向 → 列方向：行方向的论文有多少 builds_on 指向列方向的论文。高数值 = 两个方向深度耦合 = 交叉学习富矿。
| 行 \ 列 | Action Gen | Dexterous | Foundation | Imitation | Locomotion | Motion Prior | Teleop |
|---|---|---|---|---|---|---|---|
| Action Gen | · | · | 47 | 2 | 13 | 8 | 2 |
| Dexterous | 4 | · | 28 | 4 | 6 | · | 7 |
| Foundation | 17 | 4 | · | 33 | 13 | · | 2 |
| Imitation | 1 | 2 | 88 | · | 3 | 1 | 3 |
| Locomotion | 28 | · | 9 | 3 | · | 8 | 2 |
| Motion Prior | 62 | · | 16 | 5 | 33 | · | · |
| Teleop | 20 | · | 18 | 10 | 3 | · | · |
最耦合的方向对（双向合计）：
- Foundation ⟷ Imitation：121 条引用
- Action Gen ⟷ Motion Prior：70 条引用
- Action Gen ⟷ Foundation：64 条引用
- Locomotion ⟷ Motion Prior：41 条引用
- Action Gen ⟷ Locomotion：41 条引用
- Dexterous ⟷ Foundation：32 条引用
- Foundation ⟷ Locomotion：22 条引用
- Action Gen ⟷ Teleop：22 条引用
6. 覆盖度
6.1 按方向
| 方向 | 数量 | 占比 |
|---|---|---|
| Foundation | 335 | 34% |
| Motion Prior | 132 | 14% |
| Imitation | 132 | 14% |
| Locomotion | 110 | 11% |
| Action Gen | 108 | 11% |
| Dexterous | 105 | 11% |
| Teleop | 53 | 5% |
6.2 按年份（近 8 年）
| 年份 | 数量 |
|---|---|
| 2026 | 316 |
| 2025 | 86 |
| 2024 | 99 |
| 2023 | 114 |
| 2022 | 78 |
| 2021 | 42 |
| 2020 | 41 |
| 2019 | 37 |
7. 查漏补缺（reference_expansion/ 汇总）
critical_missing.jsonl：标记 19 篇关键遗漏，其中 0 篇仍未收录。batch*_new_refs.jsonl（6 个批次）：候选 553 条，其中 30 条已收录、523 条待评估是否补入。missing_important_papers.md：列出 29 个 arxiv id，其中 5 个仍未收录。
8. 关键指标
关键指标的操作化仪表盘。前 4 项是元数据达标核心，达标状态实时计算：3/4 达标（✓=达标 / ✗=未达）。
浅入口覆盖已接入实测（扫 site/content 已发布教学页）；
后 2 项（高引收录率 / 事实 grounding 率）当前未接入实测，
标「未接入」，是否达标列记「—」——不冒充绿。
| 指标 | 操作定义 | 当前值 | 目标 | 是否达标 |
|---|---|---|---|---|
| 枢纽 why_important 完整率 | 枢纽(被引≥3)有 why_important_zh 占比 | 133/133 = 100.0% | 100% | ✓ |
| 枢纽结构化字段完整率 | 枢纽有 contribution_type+paper_type+reproducibility+primary_thread 占比 | 111/133 = 83.5% | ≥80% | ✓ |
| 引用解析率 | resolved/(resolved+真实漏收录)；概念标签不计分母（与 validate_citations 门同源） | 2238/2489 = 89.9%（原始 2238/3742 = 59.8%；概念 1253，真实漏收录 251） | ≥90% | ✗ |
| 孤立节点率 | 无 builds_on 且无 leads_to 论文占比（与 §4 同定义） | 8/975 = 0.8% | <10% | ✓ |
| 浅入口覆盖 | 枢纽(被引≥3)有已发布入门教学页占比 | 65/133 = 48.9% | 100% | ✗ |
| 高引工作收录率 | 近十年领域高引收录占比 | 未接入（需外部高引榜） | ≥95% | — |
| 事实 grounding 率 | 断言可回链 KB/PDF/权威源 或标 unconfirmed 占比 | 未接入（需 cite 提取管线） | ≥95%（余显式 unconfirmed） | — |
- 枢纽总数（被引≥3）：133 / 库内 975 篇
- 引用标签总数：3742，已解析：2238；未解析中概念 1253、真实漏收录 251（解析率分母排除概念）
- 元数据核心 4 项：3/4 达标（须 4/4 方为完整）
读法：✓/✗ 由当前值对照「目标」列实时计算（不写死）；「—」= 该指标尚未接入实测，不计入达标计数。
9. 研究群落与桥接工作
在引用图（citations_normalized.jsonl，1803 条 builds_on/leads_to 边）上跑 label-propagation 社区检测 + 模块度合并，把密集互引的论文聚成研究群落；再识别桥接工作——邻居跨越 ≥2 个群落的论文，即「交叉学习」的关键节点。由 scripts/cluster_analysis.py 产出 clusters.jsonl + bridges.jsonl，本节读其结果。
群落主表：5 个群落，覆盖库内 754 篇论文 （其余为孤立/小群落，<5 篇不计入主表）。群落编号 = 按规模降序。
| # | 规模 | 主导方向 | 主导子方向 | 主题（启发式） | Top-3 代表论文（PageRank） |
|---|---|---|---|---|---|
| C0 | 265 | Foundation | VLA | 视觉-语言-动作模型（VLA） | · RT-2: Vision-Language-Action Models Transfer Web Knowledge t（pr=0.0051） · Open X-Embodiment: Robotic Learning Datasets and RT-X Models（pr=0.0045） · OpenVLA: An Open-Source Vision-Language-Action Model（pr=0.0044） |
| C1 | 202 | Motion Prior | humanoid_rl | 人形机器人 RL / 扩散模型 / 动作跟踪 | · AMP: Adversarial Motion Priors for Stylized Physics-Based Ch（pr=0.0107） · BeyondMimic: From Motion Tracking to Versatile Humanoid Cont（pr=0.0075） · ASE: Large-Scale Reusable Adversarial Skill Embeddings for P（pr=0.0061） |
| C2 | 134 | Foundation | world_model | 世界模型 / 极限地形/parkour / locomotion | · Extreme Parkour with Legged Robots（pr=0.0070） · Parkour in the Wild: Learning a General and Extensible Agile（pr=0.0039） · World Action Models: The Next Frontier in Embodied AI（pr=0.0032） |
| C3 | 87 | Action Gen | retarget | 动作重定向/映射 / 扩散模型 | · MaskedMimic: Unified Physics-Based Character Control Through（pr=0.0054） · Human Motion Diffusion Model (MDM)（pr=0.0050） · MotionGPT: Human Motion as a Foreign Language（pr=0.0037） |
| C4 | 66 | Dexterous | in_hand_rl | 灵巧手内操作 | · Learning Dexterous In-Hand Manipulation（pr=0.0036） · Learning Robust Dexterous In-Hand Manipulation from Vision a（pr=0.0030） · DexImit: Learning Bimanual Dexterous Manipulation from Monoc（pr=0.0026） |
桥接工作 Top-15（按桥接度 = 邻居群落数 span + 0.1×跨群落邻居数 cross 排序）：
| 桥接度 | span | cross | PageRank | 工作 | 方向 | 所属群落 |
|---|---|---|---|---|---|---|
| 4.9 | 4 | 9 | 0.0009 | Proximal Policy Optimization Algorithms (PPO) | Locomotion | C1 |
| 4.5 | 4 | 5 | 0.0015 | RMA: Rapid Motor Adaptation for Legged Robots | Locomotion | C2 |
| 4.3 | 4 | 3 | 0.0006 | Isaac Gym: High Performance GPU-Based Physics Simulation For | Locomotion | C2 |
| 4.0 | 3 | 10 | 0.0014 | AMASS: Archive of Motion Capture as Surface Shapes | Motion Prior | C3 |
| 3.9 | 3 | 9 | 0.0009 | GR00T N1: An Open Foundation Model for Generalist Humanoid R | Foundation | C1 |
| 3.6 | 3 | 6 | 0.0044 | OpenVLA: An Open-Source Vision-Language-Action Model | Foundation | C0 |
| 3.6 | 3 | 6 | 0.0006 | World Models | Foundation | C2 |
| 3.5 | 3 | 5 | 0.0026 | DeepMimic: Example-Guided Deep Reinforcement Learning of Phy | Locomotion | C1 |
| 3.4 | 3 | 4 | 0.0107 | AMP: Adversarial Motion Priors for Stylized Physics-Based Ch | Locomotion | C1 |
| 3.4 | 3 | 4 | 0.0028 | Cosmos World Foundation Model Platform for Physical AI | Foundation | C2 |
| 3.4 | 3 | 4 | 0.0008 | Language Models are Few-Shot Learners (GPT-3) | Foundation | C2 |
| 3.4 | 3 | 4 | 0.0006 | Auto-Encoding Variational Bayes (VAE) | Foundation | C1 |
| 3.3 | 3 | 3 | 0.0039 | Learning Fine-Grained Bimanual Manipulation with Low-Cost Ha | Teleop | C0 |
| 3.3 | 3 | 3 | 0.0032 | GR00T N1.x (NVIDIA humanoid VLA foundation model, 2026 itera | Imitation | C1 |
| 3.3 | 3 | 3 | 0.0019 | π0: A Vision-Language-Action Flow Model for General Robot Co | Foundation | C0 |
读法：群落编号 C0–C4 与上表一致；span = 该论文邻居所属的不同群落数（含自身群落，≥2 即桥接）；cross = 跨群落的邻居数；桥接度 = span + 0.1×cross，越高越关键。桥接工作通常是 PPO / Transformer / AMASS / MuJoCo / OpenVLA / Isaac Gym 等被多个方向共同引用的基础工作——它们正是「交叉学习富矿」最值得深挖的节点。