Upload AWSM article, metrics, manifests, and scripts

.gitignore

@@ -0,0 +1,4 @@
+__pycache__/
+qa/
+.venv/
+.publication.json

app.js

@@ -0,0 +1,101 @@
+const C={zh:{t:["输入与建模路线","位姿、原生深度与模型深度","冻结模型几何","固定 5 视角外观","同场景 100 视角外观","固定 5 视角模型深度","同场景 100 视角模型深度"],view:"视角",l:{method:"方法",allowed_inputs:"允许输入",geometry_source:"几何来源",modeling_role:"建模角色",ate_m:"ATE (m) ↓",rotation_deg:"旋转误差 (°) ↓",native_depth_absrel:"原生深度 AbsRel ↓",model_depth_absrel:"模型深度 AbsRel ↓",alignment:"配准 / 限制",model_to_gt_mean_m:"模型→GT 均值 (m) ↓",observed_gt_to_model_mean_m:"观测 GT→模型均值 (m) ↓",bidirectional_mean_m:"双向平均 (m) ↓",model_sha256:"模型 SHA256",psnr_db:"PSNR (dB) ↑",ssim:"SSIM ↑",lpips_alex_v01:"LPIPS ↓",absrel:"AbsRel ↓",rmse_m:"RMSE (m) ↓",coverage:"覆盖率 ↑",penalized_mae_m:"缺失惩罚 MAE (m) ↓"}},en:{t:["Inputs and modelling routes","Pose, native depth, and model depth","Frozen-model geometry","Fixed-five appearance","Same-scene 100-view appearance","Fixed-five model depth","Same-scene 100-view model depth"],view:"VIEW",l:{method:"Method",allowed_inputs:"Allowed inputs",geometry_source:"Geometry source",modeling_role:"Modelling role",ate_m:"ATE (m) ↓",rotation_deg:"Rotation (°) ↓",native_depth_absrel:"Native depth AbsRel ↓",model_depth_absrel:"Model depth AbsRel ↓",alignment:"Alignment / limitation",model_to_gt_mean_m:"Model→GT mean (m) ↓",observed_gt_to_model_mean_m:"Observed GT→model mean (m) ↓",bidirectional_mean_m:"Bidirectional mean (m) ↓",model_sha256:"Model SHA256",psnr_db:"PSNR (dB) ↑",ssim:"SSIM ↑",lpips_alex_v01:"LPIPS ↓",absrel:"AbsRel ↓",rmse_m:"RMSE (m) ↓",coverage:"Coverage ↑",penalized_mae_m:"Penalized MAE (m) ↓"}}};
+const lang = document.body.dataset.lang === "zh" ? "zh" : "en";
+const c = C[lang];
+const methodNames = {
+  en: {M1: "RGB-only", M2: "ViPE + DA3", M3: "ORB-SLAM3 + DA3", M4: "GT pose + DA3"},
+  zh: {M1: "纯 RGB", M2: "ViPE + DA3", M3: "ORB-SLAM3 + DA3", M4: "GT 位姿 + DA3"}
+};
+const methodLabel = id => methodNames[lang][id] ? `${id} (${methodNames[lang][id]})` : id;
+// Display translations are separate from the frozen experimental data.
+const routeText = {
+  en: {
+    M1: {allowed_inputs: "180 sampled RGB frames", geometry_source: "Visual inference without measured scale", modeling_role: "Infer the layout and author objects and materials"},
+    M2: {allowed_inputs: "Full RGB video; 180 frames selected for modeling", geometry_source: "RGB-only ViPE poses → pose-conditioned DA3 depth", modeling_role: "Organize pose and depth measurements into an object-centric scene"},
+    M3: {allowed_inputs: "Video, IMU, and camera–IMU calibration", geometry_source: "Monocular-inertial ORB-SLAM3 poses → pose-conditioned DA3 depth", modeling_role: "Model using images, poses, and depth"},
+    M4: {allowed_inputs: "Video and ground-truth camera poses", geometry_source: "GT poses → pose-conditioned DA3 depth", modeling_role: "The same type of modeling workflow as M3"}
+  },
+  zh: {
+    M1: {allowed_inputs: "180 张采样 RGB 图像", geometry_source: "视觉推断，无测量尺度", modeling_role: "推断布局，编写对象与材质"},
+    M2: {allowed_inputs: "完整 RGB 视频；选取 180 帧建模", geometry_source: "仅基于 RGB 的 ViPE 位姿 → 位姿条件下的 DA3 深度", modeling_role: "将位姿与深度测量组织为对象化场景"},
+    M3: {allowed_inputs: "视频、IMU、相机–IMU 标定", geometry_source: "ORB-SLAM3 单目惯性位姿 → 位姿条件下的 DA3 深度", modeling_role: "结合图像、位姿、深度建模"},
+    M4: {allowed_inputs: "视频及真值相机位姿", geometry_source: "真值位姿 → 位姿条件下的 DA3 深度", modeling_role: "与 M3 相同类型的建模流程"}
+  }
+};
+const directions = {
+  ate_m: "min", rotation_deg: "min", native_depth_absrel: "min", model_depth_absrel: "min",
+  model_to_gt_mean_m: "min", observed_gt_to_model_mean_m: "min", bidirectional_mean_m: "min",
+  psnr_db: "max", ssim: "max", lpips_alex_v01: "min", absrel: "min", rmse_m: "min",
+  coverage: "max", penalized_mae_m: "min"
+};
+const hidden = {2: new Set(["alignment"]), 3: new Set(["model_sha256", "alignment"])};
+function formatCell(key, value) {
+  if (value === null) return "—";
+  if (key === "method") return methodLabel(value);
+  if (typeof value !== "number") return String(value).replace(/\bM[1-4]\b/g, methodLabel);
+  return ["native_depth_absrel", "model_depth_absrel", "absrel", "coverage"].includes(key)
+    ? (100 * value).toFixed(2) + "%" : value.toFixed(4);
+}
+fetch("data/tables_1_7.json").then(response => {
+  if (!response.ok) throw Error(response.status);
+  return response.json();
+}).then(data => {
+  for (let number = 1; number <= 7; number++) {
+    const figure = document.querySelector("#table-" + number);
+    if (!figure) continue;
+    const rows = data.tables["table" + number];
+    const keys = Object.keys(rows[0]).filter(key => !hidden[number]?.has(key));
+    const best = {};
+    for (const key of keys) {
+      if (!directions[key]) continue;
+      const values = rows.map(row => row[key]).filter(value => typeof value === "number" && Number.isFinite(value));
+      if (values.length) best[key] = (directions[key] === "max" ? Math.max : Math.min)(...values);
+    }
+    figure.querySelector(".table-title").textContent = c.t[number - 1];
+    const table = document.createElement("table");
+    const header = table.createTHead().insertRow();
+    for (const key of keys) {
+      const cell = document.createElement("th");
+      cell.scope = "col";
+      cell.textContent = c.l[key] || key;
+      header.append(cell);
+    }
+    const body = table.createTBody();
+    for (const row of rows) {
+      const tr = body.insertRow();
+      keys.forEach((key, index) => {
+        const cell = document.createElement(index ? "td" : "th");
+        if (!index) cell.scope = "row";
+        if (["allowed_inputs", "geometry_source", "modeling_role", "alignment"].includes(key)) cell.className = "wrap";
+        const value = number === 1 ? routeText[lang][row.method]?.[key] ?? row[key] : row[key];
+        const text = formatCell(key, value);
+        if (typeof row[key] === "number" && row[key] === best[key]) {
+          const strong = document.createElement("strong");
+          strong.textContent = text;
+          cell.append(strong);
+        } else cell.textContent = text;
+        tr.append(cell);
+      });
+    }
+    figure.querySelector(".table-mount").replaceChildren(table);
+  }
+}).catch(error => {
+  document.querySelectorAll(".table-mount").forEach(mount => {
+    mount.textContent = (lang === "zh" ? "表格加载失败：" : "Table load failed: ") + error.message;
+  });
+  console.error(error);
+});
+const method = document.querySelector("#compare-method");
+const frame = document.querySelector("#compare-frame");
+method.disabled = false;
+frame.disabled = false;
+function refresh() {
+  const id = String(frame.value).padStart(3, "0"), m = method.value;
+  document.querySelector("#compare-pred").src = `assets/fixed_views/${m}/${id}.png`;
+  document.querySelector("#compare-gt").src = `assets/fixed_views/GT/${id}.png`;
+  document.querySelector("#compare-label").textContent = `${m} / ${c.view} ${frame.value}`;
+  document.querySelector("#figure-4").dataset.method = m;
+  document.querySelector("#figure-4").dataset.frame = frame.value;
+}
+method?.addEventListener("change", refresh);
+frame?.addEventListener("change", refresh);
+refresh();

assets/icons/LICENSE.octicons.txt

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 GitHub Inc.
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

assets/icons/README.md

@@ -0,0 +1,6 @@
+# Brand icons
+
+- `github.svg`: GitHub mark from [Primer Octicons](https://github.com/primer/octicons/blob/main/icons/mark-github-16.svg), used under the included MIT license (`LICENSE.octicons.txt`).
+- `huggingface.svg`: [Hugging Face official logo](https://huggingface.co/front/assets/huggingface_logo-noborder.svg), with a viewBox added for responsive sizing.
+
+The GitHub icon links to the AWSM repository. The Hugging Face icon is displayed without a link.

data/awsm.bib

@@ -0,0 +1,8 @@
+@misc{awsm_2026,
+  title        = {{AWSM}: Agentic World Simulation and Mapping},
+  author       = {{Phygital AI}},
+  year         = {2026},
+  howpublished = {Interactive research article},
+  url          = {https://phygital-ai.github.io/agentic-world-simulation-and-mapping/},
+  note         = {Reconstruction benchmarks, editable scene assets, and map-based embodied demonstration}
+}

data/embodied_demo.json

@@ -0,0 +1,77 @@
+{
+  "schema_version": 2,
+  "title": "Map-based embodied execution: reception-desk task",
+  "run_id": "sceneweft-walljourney-20260930-214924",
+  "delivery": "20260930-four-robot-demo-final-156s",
+  "recording_date": "2026-09-30",
+  "task_intent": {
+    "zh": "让无人机去前台，并让机器人在那里排好队。",
+    "en": "Send the drone to the reception desk and have the robots line up there.",
+    "language_to_plan_parser_demonstrated": false
+  },
+  "scene": {
+    "method": "M4",
+    "blend_sha256": "cc6cb605246a255d94e36b9dc3f8b91d5a292b047470bc0ff75b8a5c404b0cfe",
+    "manifest": "evidence/M4/manifest.json"
+  },
+  "navigation": {
+    "localization": "simulator_pose",
+    "route": "predefined",
+    "route_screening": "reconstructed_and_original_geometry",
+    "environment": "Isaac Sim / PhysX dynamics simulation",
+    "embodiments": ["drone", "humanoid", "quadruped", "wheeled_robot"],
+    "demonstration": "target marked on reconstructed map, routes specified, controller execution"
+  },
+  "media": {
+    "video": "assets/embodied-demo/four-robots-214924-156s-phone.mp4",
+    "poster": "assets/embodied-demo/navigation-214924-poster.png",
+    "poster_sha256": "fe1dfcc8d014fc47f93a6e4e343598c146373c2d3dc7b41613123bb6caafec8f",
+    "sha256": "b9008c7b754f994d30a740a632b70ea0f693f3917fcf11820e64902ddbd13596",
+    "bytes": 11648571,
+    "duration_s": 156.0,
+    "width": 1280,
+    "height": 720,
+    "playback_rate": "1x",
+    "source_fps": "100/7",
+    "output_fps": 30,
+    "frame_count": 4680,
+    "route_trimmed": true,
+    "trimmed_opening_s": 12.32,
+    "source_frame_ids": [177, 2148],
+    "later_terminal_hold_included": false,
+    "frame_rate_conversion": "Timestamp-preserving frame duplication/drop; no optical-flow interpolation or acceleration",
+    "hd": {
+      "video": "assets/embodied-demo/four-robots-214924-156s-hd.mp4",
+      "sha256": "f83dbb49e3aa97fd3f5f9c54ed79ed8624bac75dbd7e4cdddc4dc414b76ed222",
+      "bytes": 49687378
+    },
+    "map": {
+      "image": "assets/embodied-demo/planned-route-156s.png",
+      "sha256": "e3c8e766f40a975a0aceca185fa60867375153fc8ddb6db451788eb935cb9b8b",
+      "bytes": 1125683,
+      "width": 1920,
+      "height": 1080,
+      "role": "static planned-route design, not a measured execution trajectory"
+    },
+    "segments": [
+      {"start_s": 0, "end_s": 6, "role": "scene presentation"},
+      {"start_s": 6, "end_s": 12, "role": "reconstruction presentation"},
+      {"start_s": 12, "end_s": 18, "role": "static planned-route presentation; not an execution trace"},
+      {"start_s": 18, "end_s": 156, "role": "recorded native RGB execution from run 20260930-214924"}
+    ]
+  },
+  "audit": {
+    "status": "FAILED",
+    "acceptance": false,
+    "film_accepted": false,
+    "summary": "evidence/embodied-demo-audit.json",
+    "ground_final_formation_reached": true,
+    "full_flight_mission_completed": true,
+    "required_terminal_hold_s": 8.0,
+    "measured_terminal_hold_s": 7.994999821297824
+  },
+  "scope": {
+    "demonstrated": ["map-based route execution in simulation", "shared map used by multiple embodiments"],
+    "not_evaluated": ["online visual localization", "learned instruction parsing", "memory retrieval", "real-robot transfer", "navigation comparison across M1-M4", "policy sim-to-real gap"]
+  }
+}

data/fixed_views.json

@@ -0,0 +1,221 @@
+{
+  "schema_version": 1,
+  "source_manifest": "astra_blender2/report/space/results/blog_tables_1_5/reference_style_figures.json",
+  "source_manifest_sha256": "be3fb4153dbba73c6eeb2f4ba0070cf68539f848e6829c5da87c2b3398907b07",
+  "frames": [
+    0,
+    36,
+    72,
+    108,
+    144
+  ],
+  "methods": [
+    "M1",
+    "M2",
+    "M3",
+    "M4",
+    "GT"
+  ],
+  "images": [
+    {
+      "method": "M1",
+      "frame": 0,
+      "source": "world_lobby_four_trajectory_20260929/astra_blender2/evaluation/reference_style_figures/renders/M1/000.png",
+      "published": "assets/fixed_views/M1/000.png",
+      "sha256": "33b9f0d09f34871564d86bde23d61d82d48133b3f9ac2f6c251b8435f4bce5c4",
+      "bytes": 460982
+    },
+    {
+      "method": "M2",
+      "frame": 0,
+      "source": "world_lobby_four_trajectory_20260929/astra_blender2/evaluation/reference_style_figures/renders/M2/000.png",
+      "published": "assets/fixed_views/M2/000.png",
+      "sha256": "dff45b5cd9bc4c5d757f36c8b8eaff015684c6681b40ad0d6b936940e7c9f316",
+      "bytes": 465709
+    },
+    {
+      "method": "M3",
+      "frame": 0,
+      "source": "world_lobby_four_trajectory_20260929/astra_blender2/evaluation/reference_style_figures/renders/M3/000.png",
+      "published": "assets/fixed_views/M3/000.png",
+      "sha256": "a2ca215a65c82b7361c072694c6a1201464bc2ad80c18ba7c59a963927d89b44",
+      "bytes": 418136
+    },
+    {
+      "method": "M4",
+      "frame": 0,
+      "source": "world_lobby_four_trajectory_20260929/astra_blender2/evaluation/blog_tables_1_6_current_m4/renders/M4/000.png",
+      "published": "assets/fixed_views/M4/000.png",
+      "sha256": "a4e82854e7ce3e3870959f008a6d590f9c65925e4b33ca0e96f6cf7ffdd6ab43",
+      "bytes": 536558
+    },
+    {
+      "method": "GT",
+      "frame": 0,
+      "source": "vio-reconstruction/sessions/lobby_orb_success_20260929T123153/inputs/cam0/data/249080011831.png",
+      "published": "assets/fixed_views/GT/000.png",
+      "sha256": "0c95e2288b7d864133952d0ce6f552abea743a7d4ea4622103047bd794a07d04",
+      "bytes": 3688630
+    },
+    {
+      "method": "M1",
+      "frame": 36,
+      "source": "world_lobby_four_trajectory_20260929/astra_blender2/evaluation/reference_style_figures/renders/M1/036.png",
+      "published": "assets/fixed_views/M1/036.png",
+      "sha256": "d1ccc4bf1da6c2b9e90bfc38bd37d8598a0a0068b374d05a4138bac2b37a904c",
+      "bytes": 490712
+    },
+    {
+      "method": "M2",
+      "frame": 36,
+      "source": "world_lobby_four_trajectory_20260929/astra_blender2/evaluation/reference_style_figures/renders/M2/036.png",
+      "published": "assets/fixed_views/M2/036.png",
+      "sha256": "2f8cb5397a7800f0229867ba94b0ee8e8c453600176531da65f27f96e7f4f147",
+      "bytes": 478265
+    },
+    {
+      "method": "M3",
+      "frame": 36,
+      "source": "world_lobby_four_trajectory_20260929/astra_blender2/evaluation/reference_style_figures/renders/M3/036.png",
+      "published": "assets/fixed_views/M3/036.png",
+      "sha256": "6c117fef7d146ae81baab7453cb88a2df7f75b080d8e9c672e2163a9fbf00f79",
+      "bytes": 433089
+    },
+    {
+      "method": "M4",
+      "frame": 36,
+      "source": "world_lobby_four_trajectory_20260929/astra_blender2/evaluation/blog_tables_1_6_current_m4/renders/M4/036.png",
+      "published": "assets/fixed_views/M4/036.png",
+      "sha256": "a0a7a8afcd3ebc4e56c79a45aa41740f28ca08186b00a67af19161f23f773c2d",
+      "bytes": 547041
+    },
+    {
+      "method": "GT",
+      "frame": 36,
+      "source": "vio-reconstruction/sessions/lobby_orb_success_20260929T123153/inputs/cam0/data/283280013455.png",
+      "published": "assets/fixed_views/GT/036.png",
+      "sha256": "449c88399a68cd88397306fbe42c5293e6684d9a9cfb749ea9e3f5ff78eef04d",
+      "bytes": 3688630
+    },
+    {
+      "method": "M1",
+      "frame": 72,
+      "source": "world_lobby_four_trajectory_20260929/astra_blender2/evaluation/reference_style_figures/renders/M1/072.png",
+      "published": "assets/fixed_views/M1/072.png",
+      "sha256": "30b78883e701ef411d4cc04a44dffaba6df13631f6ab4c643edc42e2aa4b2415",
+      "bytes": 431098
+    },
+    {
+      "method": "M2",
+      "frame": 72,
+      "source": "world_lobby_four_trajectory_20260929/astra_blender2/evaluation/reference_style_figures/renders/M2/072.png",
+      "published": "assets/fixed_views/M2/072.png",
+      "sha256": "899b65b8c0487f9987f38d09b246c9f97b0b94217d11ebbe29b990f48f19b003",
+      "bytes": 424182
+    },
+    {
+      "method": "M3",
+      "frame": 72,
+      "source": "world_lobby_four_trajectory_20260929/astra_blender2/evaluation/reference_style_figures/renders/M3/072.png",
+      "published": "assets/fixed_views/M3/072.png",
+      "sha256": "7d40497cf6c2c3c20403a4c3e1f9b0abe478fbb8602e1f259e51c48dd535d4f3",
+      "bytes": 390554
+    },
+    {
+      "method": "M4",
+      "frame": 72,
+      "source": "world_lobby_four_trajectory_20260929/astra_blender2/evaluation/blog_tables_1_6_current_m4/renders/M4/072.png",
+      "published": "assets/fixed_views/M4/072.png",
+      "sha256": "a934d91a6f6b9e130322f526e212dc9f279e1c787398e0181da3e9dd4f60127c",
+      "bytes": 491159
+    },
+    {
+      "method": "GT",
+      "frame": 72,
+      "source": "vio-reconstruction/sessions/lobby_orb_success_20260929T123153/inputs/cam0/data/317520015081.png",
+      "published": "assets/fixed_views/GT/072.png",
+      "sha256": "222246adc897e2c42232a2b10231c614601dcd71580186647522ff1881a9b71f",
+      "bytes": 3688630
+    },
+    {
+      "method": "M1",
+      "frame": 108,
+      "source": "world_lobby_four_trajectory_20260929/astra_blender2/evaluation/reference_style_figures/renders/M1/108.png",
+      "published": "assets/fixed_views/M1/108.png",
+      "sha256": "9c98270c0e013225d168ee7714afbd2a80f5dfa764c0b158437d1985c26f4745",
+      "bytes": 309546
+    },
+    {
+      "method": "M2",
+      "frame": 108,
+      "source": "world_lobby_four_trajectory_20260929/astra_blender2/evaluation/reference_style_figures/renders/M2/108.png",
+      "published": "assets/fixed_views/M2/108.png",
+      "sha256": "291733823717146fa1e474187aafdca49d6e06b8900985dde6c48fef55684926",
+      "bytes": 464925
+    },
+    {
+      "method": "M3",
+      "frame": 108,
+      "source": "world_lobby_four_trajectory_20260929/astra_blender2/evaluation/reference_style_figures/renders/M3/108.png",
+      "published": "assets/fixed_views/M3/108.png",
+      "sha256": "12b289c7af68a32b1a00817685ed79ede039b3624af9896a57e90b1d5b99d864",
+      "bytes": 433108
+    },
+    {
+      "method": "M4",
+      "frame": 108,
+      "source": "world_lobby_four_trajectory_20260929/astra_blender2/evaluation/blog_tables_1_6_current_m4/renders/M4/108.png",
+      "published": "assets/fixed_views/M4/108.png",
+      "sha256": "4046e5e362ab8a81404aabd59da159a192f206bf7a80b694aca130c698fe2c7e",
+      "bytes": 531759
+    },
+    {
+      "method": "GT",
+      "frame": 108,
+      "source": "vio-reconstruction/sessions/lobby_orb_success_20260929T123153/inputs/cam0/data/351720016706.png",
+      "published": "assets/fixed_views/GT/108.png",
+      "sha256": "5e5d07ef9ee0c166c24160b228dfc802ab244ac6002e8cfa76e450ee4a385565",
+      "bytes": 3688630
+    },
+    {
+      "method": "M1",
+      "frame": 144,
+      "source": "world_lobby_four_trajectory_20260929/astra_blender2/evaluation/reference_style_figures/renders/M1/144.png",
+      "published": "assets/fixed_views/M1/144.png",
+      "sha256": "cf7e5355a33a83fd148daf6a648936757e8be6d09639ace537cdb2ed37a8ed98",
+      "bytes": 486611
+    },
+    {
+      "method": "M2",
+      "frame": 144,
+      "source": "world_lobby_four_trajectory_20260929/astra_blender2/evaluation/reference_style_figures/renders/M2/144.png",
+      "published": "assets/fixed_views/M2/144.png",
+      "sha256": "6c4f51435f109f71a35a09c0757e8995dd4de3ebec16c43ee32c25c94303e165",
+      "bytes": 481190
+    },
+    {
+      "method": "M3",
+      "frame": 144,
+      "source": "world_lobby_four_trajectory_20260929/astra_blender2/evaluation/reference_style_figures/renders/M3/144.png",
+      "published": "assets/fixed_views/M3/144.png",
+      "sha256": "cc546f9d355fd3d8ac4f48a08b4e4a1158c112585922bb34e8715229482a3542",
+      "bytes": 474768
+    },
+    {
+      "method": "M4",
+      "frame": 144,
+      "source": "world_lobby_four_trajectory_20260929/astra_blender2/evaluation/blog_tables_1_6_current_m4/renders/M4/144.png",
+      "published": "assets/fixed_views/M4/144.png",
+      "sha256": "1d6080b2ce9f2ae95c6a6d9fc47a620365a4439df89267ab166122287d8b61be",
+      "bytes": 529379
+    },
+    {
+      "method": "GT",
+      "frame": 144,
+      "source": "vio-reconstruction/sessions/lobby_orb_success_20260929T123153/inputs/cam0/data/385920018330.png",
+      "published": "assets/fixed_views/GT/144.png",
+      "sha256": "cb8c1a4ade400f935ef4e1e3bfe93245fe77d8beb239040bf43baea840996444",
+      "bytes": 3688630
+    }
+  ]
+}

data/office_video_media.json

@@ -0,0 +1,109 @@
+{
+  "release_id": 400742542,
+  "tag": "office-media-20261001",
+  "assets": [
+    {
+      "file": "office-input.mp4",
+      "bytes": 11309649,
+      "sha256": "b99fbe48c99410972782b0a10754d3da2e461dda2dd95aa94c4213bb8d0baeb8",
+      "url": "https://github.com/Phygital-AI/agentic-world-simulation-and-mapping/releases/download/office-media-20261001/office-input.mp4",
+      "asset_id": 602717050
+    },
+    {
+      "file": "office-pmodel.mp4",
+      "bytes": 19214748,
+      "sha256": "f2caba7641bb50ff65a231ea1d245724e884d449a41b89e829854924b53a514b",
+      "url": "https://github.com/Phygital-AI/agentic-world-simulation-and-mapping/releases/download/office-media-20261001/office-pmodel.mp4",
+      "asset_id": 602717164
+    },
+    {
+      "file": "office-scan.mp4",
+      "bytes": 153594492,
+      "sha256": "4643631ba15e6a33c43ea3dd3f259f2b68fe1d67c69a44262ad5bf1519b5daf3",
+      "url": "https://github.com/Phygital-AI/agentic-world-simulation-and-mapping/releases/download/office-media-20261001/office-scan.mp4",
+      "asset_id": 602717518
+    }
+  ],
+  "schema_version": 1,
+  "preservation": "Original source files remain byte-for-byte in the release assets. Same-origin web playback files are listed separately below.",
+  "order": [
+    "office-input.mp4",
+    "office-pmodel.mp4",
+    "office-scan.mp4"
+  ],
+  "video_spec": {
+    "width": 960,
+    "height": 540,
+    "frame_rate": "30/1",
+    "frames": 2103,
+    "duration_seconds": 70.1
+  },
+  "posters": [
+    {
+      "path": "assets/office-cafe/office-input-poster.jpg",
+      "source": "office-input.mp4",
+      "frame_index": 0,
+      "sha256": "09ee086ae0eca08afecbafb5fe5849600d0837349ae1b14e72dbdbaae9c2f237",
+      "bytes": 94250
+    },
+    {
+      "path": "assets/office-cafe/office-pmodel-poster.jpg",
+      "source": "office-pmodel.mp4",
+      "frame_index": 0,
+      "sha256": "61ea36de0e40a955879b3b0195961ad137c738373726efda760b50232b7041b8",
+      "bytes": 80038
+    },
+    {
+      "path": "assets/office-cafe/office-scan-poster.jpg",
+      "source": "office-scan.mp4",
+      "frame_index": 0,
+      "sha256": "59f64e13f254ef3c9dbf511277cd9782ac0267f25b54e99cba938269be30d4e3",
+      "bytes": 218866
+    },
+    {
+      "path": "assets/office-cafe/office-shake-heavy-poster.jpg",
+      "source": "office-shake-heavy.mp4",
+      "frame_index": 0,
+      "sha256": "e4574f22938288bf76f5d123cbba7f871643a4889c519fab176484738438be6b",
+      "bytes": 104096
+    }
+  ],
+  "web_playback": [
+    {
+      "source": "office-input.mp4",
+      "path": "assets/office-cafe/office-input-web.mp4",
+      "bytes": 27902821,
+      "sha256": "4915b934b5f9971c77d60f88d91ed110c5e164bb0932f89ac7cf8915964e9d2c",
+      "processing": "H.264 Constrained Baseline, CRF 18, maxrate 4 Mbps, fastdecode, faststart, 2-second keyframes",
+      "width": 960,
+      "height": 540,
+      "frames": 2103,
+      "frame_rate": "30/1",
+      "duration_seconds": 70.1
+    },
+    {
+      "source": "office-pmodel.mp4",
+      "path": "assets/office-cafe/office-pmodel-web.mp4",
+      "bytes": 21650736,
+      "sha256": "3b03fe58a1922cdd1ca935f6ad52c69e672f1e391cdec1543e60db99e93a4d71",
+      "processing": "H.264 Constrained Baseline, CRF 18, maxrate 4 Mbps, fastdecode, faststart, 2-second keyframes",
+      "width": 960,
+      "height": 540,
+      "frames": 2103,
+      "frame_rate": "30/1",
+      "duration_seconds": 70.1
+    },
+    {
+      "source": "office-scan.mp4",
+      "path": "assets/office-cafe/office-scan-web.mp4",
+      "bytes": 35449707,
+      "sha256": "861a074637b4b756f1c6ca115b50f20c9d03b46303caa6eb0503e812c4194b92",
+      "processing": "H.264 Constrained Baseline, CRF 18, maxrate 4 Mbps, fastdecode, faststart, 2-second keyframes",
+      "width": 960,
+      "height": 540,
+      "frames": 2103,
+      "frame_rate": "30/1",
+      "duration_seconds": 70.1
+    }
+  ]
+}

data/scene_comparison.json

@@ -0,0 +1,487 @@
+{
+  "schema_version": 2,
+  "models": {
+    "GT": {
+      "url": "assets/comparison/GT.glb",
+      "sha256": "6c28ab067c63c5b1fce19c6a972f3d66c7ac4d9bd9ce3b5a57e2e18f5641fe8f",
+      "bytes": 24816108,
+      "registration": "identity"
+    },
+    "M1": {
+      "url": "models/M1/scene.glb",
+      "sha256": "21a605be65d58dbe01606b8326bfe6f07fa418c67e133c9b2181729b68bc2832",
+      "bytes": 15104860,
+      "runtime_matrix_gltf_y_up": [
+        [
+          -0.034082710190599566,
+          -0.017474675835450443,
+          -1.003087981240217,
+          6.886067611483274
+        ],
+        [
+          -0.07852778991034609,
+          1.0006337553304165,
+          -0.014763720507024342,
+          -0.11837912846914733
+        ],
+        [
+          1.000162111167947,
+          0.07796933330392732,
+          -0.03534159003611681,
+          -25.800695086947528
+        ],
+        [
+          0.0,
+          0.0,
+          0.0,
+          1.0
+        ]
+      ],
+      "source_registration": "astra_blender2/evaluation/reference_style_figures/registration_M1.json"
+    },
+    "M2": {
+      "url": "models/M2/scene.glb",
+      "sha256": "0ee9123f38099d1df63ab593dd7082d5f1cf1b136e63833ec5afa0e072cf4c2e",
+      "bytes": 14523124,
+      "runtime_matrix_gltf_y_up": [
+        [
+          -0.002921335722089408,
+          -0.00012677219434212295,
+          -0.9999957248540663,
+          10.642407320031518
+        ],
+        [
+          0.0030202640740619557,
+          0.9999954297989613,
+          -0.00013559539998786452,
+          0.08437265493567603
+        ],
+        [
+          0.9999911718622922,
+          -0.0030206472816780125,
+          -0.002920939485494111,
+          -21.687447573786944
+        ],
+        [
+          0.0,
+          0.0,
+          0.0,
+          1.0
+        ]
+      ],
+      "source_registration": "astra_blender2/evaluation/reference_style_figures/registration_M2.json"
+    },
+    "M3": {
+      "url": "models/M3/scene.glb",
+      "sha256": "22699f8225a5f90166cdea2e6e2bdad7977fda4a1fe1086da209f0c9afe97a6c",
+      "bytes": 7701488,
+      "runtime_matrix_gltf_y_up": [
+        [
+          0.9999929680848632,
+          0.0025089340052771025,
+          0.002787298151085771,
+          7.830309933606095
+        ],
+        [
+          -0.002507975856466338,
+          0.9999967947503668,
+          -0.0003471974908570291,
+          -0.005554550842903794
+        ],
+        [
+          -0.0027881603126906785,
+          0.0003402045729258979,
+          0.9999960552036794,
+          -17.966462178842214
+        ],
+        [
+          0.0,
+          0.0,
+          0.0,
+          1.0
+        ]
+      ],
+      "source_registration": "astra_blender2/evaluation/reference_style_figures/registration_M3.json"
+    },
+    "M4": {
+      "url": "models/M4/scene.glb",
+      "sha256": "716e4b89b7b11ad8a0cb0e238b366b4f5dc80e3a2c3cddb5eedc469ef4e63774",
+      "bytes": 17202608,
+      "runtime_matrix_gltf_y_up": [
+        [
+          1,
+          0,
+          0,
+          0
+        ],
+        [
+          0,
+          0,
+          1,
+          0
+        ],
+        [
+          0,
+          -1,
+          0,
+          0
+        ],
+        [
+          0,
+          0,
+          0,
+          1
+        ]
+      ],
+      "source_registration": null
+    }
+  },
+  "camera": {
+    "target": [
+      18,
+      1.2,
+      -21
+    ],
+    "position": [
+      24,
+      23,
+      -5
+    ],
+    "fov": 42,
+    "near": 0.05,
+    "far": 500
+  },
+  "coordinate_conversion": {
+    "reason": "glTF is Y-up while registration is Blender Z-up",
+    "C_mapping": "[x,y,z] -> [x,z,-y]",
+    "C": [
+      [
+        1,
+        0,
+        0,
+        0
+      ],
+      [
+        0,
+        0,
+        1,
+        0
+      ],
+      [
+        0,
+        -1,
+        0,
+        0
+      ],
+      [
+        0,
+        0,
+        0,
+        1
+      ]
+    ],
+    "C_inverse": [
+      [
+        1,
+        0,
+        0,
+        0
+      ],
+      [
+        0,
+        0,
+        -1,
+        0
+      ],
+      [
+        0,
+        1,
+        0,
+        0
+      ],
+      [
+        0,
+        0,
+        0,
+        1
+      ]
+    ],
+    "runtime_formula": "C * T * C^-1"
+  },
+  "registrations": {
+    "M1": {
+      "source": {
+        "path": "astra_blender2/evaluation/reference_style_figures/registration_M1.json",
+        "sha256": "db6b936e5d04b1be27048790ab354f481f89d20f4d46739340ea6f6fc1048d1e",
+        "transform_direction": "GT_from_model",
+        "status": "COMPLETE_DIAGNOSTIC",
+        "scope": "GT-assisted Sim3 from frozen manual-view camera associations; shape/display diagnostic only; not metric recovery; no mesh ICP",
+        "limitations": "All M1 authored cameras remain invalid for native trajectory evaluation; this GT-assisted registration is diagnostic only."
+      },
+      "GT_from_model_blender_z_up": [
+        [
+          -0.034082710190599566,
+          1.003087981240217,
+          -0.017474675835450443,
+          6.886067611483274
+        ],
+        [
+          -1.000162111167947,
+          -0.03534159003611681,
+          -0.07796933330392732,
+          25.800695086947528
+        ],
+        [
+          -0.07852778991034609,
+          0.014763720507024342,
+          1.0006337553304165,
+          -0.11837912846914733
+        ],
+        [
+          0.0,
+          0.0,
+          0.0,
+          1.0
+        ]
+      ],
+      "runtime_matrix_gltf_y_up": [
+        [
+          -0.034082710190599566,
+          -0.017474675835450443,
+          -1.003087981240217,
+          6.886067611483274
+        ],
+        [
+          -0.07852778991034609,
+          1.0006337553304165,
+          -0.014763720507024342,
+          -0.11837912846914733
+        ],
+        [
+          1.000162111167947,
+          0.07796933330392732,
+          -0.03534159003611681,
+          -25.800695086947528
+        ],
+        [
+          0.0,
+          0.0,
+          0.0,
+          1.0
+        ]
+      ],
+      "formula": "C * GT_from_model * C^-1",
+      "diagnostic_caveat": "M1 only: GT-assisted Sim(3) from five manual camera associations; diagnostic only; not native metric recovery."
+    },
+    "M2": {
+      "source": {
+        "path": "astra_blender2/evaluation/reference_style_figures/registration_M2.json",
+        "sha256": "456f1bed62a2ece0601e05d3f6044545f6835999d2bd98394d9ae278f44022b6",
+        "transform_direction": "GT_from_model",
+        "status": "COMPLETE",
+        "scope": "one SE3 from 180 input camera correspondences; no scale fit",
+        "limitations": null
+      },
+      "GT_from_model_blender_z_up": [
+        [
+          -0.002921335722089408,
+          0.9999957248540663,
+          -0.00012677219434212295,
+          10.642407320031518
+        ],
+        [
+          -0.9999911718622922,
+          -0.002920939485494111,
+          0.0030206472816780125,
+          21.687447573786944
+        ],
+        [
+          0.0030202640740619557,
+          0.00013559539998786452,
+          0.9999954297989613,
+          0.08437265493567603
+        ],
+        [
+          0.0,
+          0.0,
+          0.0,
+          1.0
+        ]
+      ],
+      "runtime_matrix_gltf_y_up": [
+        [
+          -0.002921335722089408,
+          -0.00012677219434212295,
+          -0.9999957248540663,
+          10.642407320031518
+        ],
+        [
+          0.0030202640740619557,
+          0.9999954297989613,
+          -0.00013559539998786452,
+          0.08437265493567603
+        ],
+        [
+          0.9999911718622922,
+          -0.0030206472816780125,
+          -0.002920939485494111,
+          -21.687447573786944
+        ],
+        [
+          0.0,
+          0.0,
+          0.0,
+          1.0
+        ]
+      ],
+      "formula": "C * GT_from_model * C^-1",
+      "diagnostic_caveat": null
+    },
+    "M3": {
+      "source": {
+        "path": "astra_blender2/evaluation/reference_style_figures/registration_M3.json",
+        "sha256": "30454936b3a9a2a8163ce8ae4851979ce4df71826028ed15967f6a7e882dce49",
+        "transform_direction": "GT_from_model",
+        "status": "COMPLETE",
+        "scope": "one SE3 from 180 input camera correspondences; no scale fit",
+        "limitations": null
+      },
+      "GT_from_model_blender_z_up": [
+        [
+          0.9999929680848632,
+          -0.002787298151085771,
+          0.0025089340052771025,
+          7.830309933606095
+        ],
+        [
+          0.0027881603126906785,
+          0.9999960552036794,
+          -0.0003402045729258979,
+          17.966462178842214
+        ],
+        [
+          -0.002507975856466338,
+          0.0003471974908570291,
+          0.9999967947503668,
+          -0.005554550842903794
+        ],
+        [
+          0.0,
+          0.0,
+          0.0,
+          1.0
+        ]
+      ],
+      "runtime_matrix_gltf_y_up": [
+        [
+          0.9999929680848632,
+          0.0025089340052771025,
+          0.002787298151085771,
+          7.830309933606095
+        ],
+        [
+          -0.002507975856466338,
+          0.9999967947503668,
+          -0.0003471974908570291,
+          -0.005554550842903794
+        ],
+        [
+          -0.0027881603126906785,
+          0.0003402045729258979,
+          0.9999960552036794,
+          -17.966462178842214
+        ],
+        [
+          0.0,
+          0.0,
+          0.0,
+          1.0
+        ]
+      ],
+      "formula": "C * GT_from_model * C^-1",
+      "diagnostic_caveat": null
+    },
+    "M4": {
+      "source": {
+        "path": null,
+        "sha256": null,
+        "transform_direction": "GT_from_model",
+        "rule": "identity_world; source GLB retains Blender Z-up axes"
+      },
+      "GT_from_model_blender_z_up": [
+        [
+          1,
+          0,
+          0,
+          0
+        ],
+        [
+          0,
+          1,
+          0,
+          0
+        ],
+        [
+          0,
+          0,
+          1,
+          0
+        ],
+        [
+          0,
+          0,
+          0,
+          1
+        ]
+      ],
+      "runtime_matrix_gltf_y_up": [
+        [
+          1,
+          0,
+          0,
+          0
+        ],
+        [
+          0,
+          0,
+          1,
+          0
+        ],
+        [
+          0,
+          -1,
+          0,
+          0
+        ],
+        [
+          0,
+          0,
+          0,
+          1
+        ]
+      ],
+      "formula": "C (axis conversion only; GT_from_model is identity)",
+      "diagnostic_caveat": null
+    }
+  },
+  "display_cutaway": {
+    "scope": "runtime display only; frozen GLB/Blend bytes remain unchanged",
+    "hide_name_pattern": "ceiling|roof|wall|facade",
+    "keep_name_pattern": "emblem|art|grass|metal",
+    "applies_to": [
+      "M1",
+      "M2",
+      "M3",
+      "M4"
+    ],
+    "gt_asset": "already exported as a cutaway display model"
+  },
+  "immutability": "Display copies and runtime registration do not alter frozen source GLB bytes.",
+  "m1_diagnostic_caveat": "M1 registration is diagnostic only and is not native metric recovery.",
+  "gt_asset": {
+    "source_url": "https://wentingw.github.io/astra-world-model-blog/assets/comparison/GT.glb",
+    "published": "assets/comparison/GT.glb",
+    "sha256": "6c28ab067c63c5b1fce19c6a972f3d66c7ac4d9bd9ce3b5a57e2e18f5641fe8f",
+    "bytes": 24816108
+  }
+}

data/source_contract.json

@@ -0,0 +1,576 @@
+{
+  "schema_version": 2,
+  "old_site_baseline": "ae69dbe",
+  "source_root_name": "world_lobby_four_trajectory_20260929",
+  "table_source": {
+    "path": "astra_blender2/evaluation/blog_extended_100_20260930/tables_1_7.json",
+    "sha256": "01faed4be8948d16fbdc081bda859554f3bfc3f20c55746fba0ca70369179287",
+    "published_url": "data/tables_1_7.json"
+  },
+  "figures": {
+    "assets/figure14_abc.png": {
+      "source": "astra_blender2/report/space/assets/figure14_abc.png",
+      "published": "assets/figure14_abc.png",
+      "sha256": "6ce74542f6ce9d45a985e1916a42ef2e1205b29bbcb03959e4b1ea08380dfa39",
+      "bytes": 344653
+    },
+    "assets/fixed_five_view_comparison_m1_m4.jpg": {
+      "source": "astra_blender2/report/space/assets/fixed_five_view_comparison_m1_m4.jpg",
+      "published": "assets/fixed_five_view_comparison_m1_m4.jpg",
+      "sha256": "501a674252a492deb209a9cfd847e8f0c83103c49d5f250964ee451d233481a9",
+      "bytes": 1755470
+    },
+    "assets/model_depth_error_five_views_m1_m4.png": {
+      "source": "astra_blender2/report/space/assets/model_depth_error_five_views_m1_m4.png",
+      "published": "assets/model_depth_error_five_views_m1_m4.png",
+      "sha256": "448bc111da467461ae84941bc6ea5f5601ca5f9746d292dc1b8b279bbfa153ce",
+      "bytes": 469515
+    },
+    "assets/teaser_originals.png": {
+      "source": "teaser_originals.png",
+      "published": "assets/teaser_originals.png",
+      "sha256": "6099f12bdaea83fb2ce4f8e1ed282ac864d6f4ffb96911efb5a9101845656dae",
+      "bytes": 4000665
+    }
+  },
+  "models": {
+    "M1": {
+      "glb": {
+        "source": "astra_blender/models/M1/scene.glb",
+        "published": "models/M1/scene.glb",
+        "sha256": "21a605be65d58dbe01606b8326bfe6f07fa418c67e133c9b2181729b68bc2832",
+        "bytes": 15104860
+      },
+      "blend": {
+        "source": "astra_blender/models/M1/scene.blend",
+        "published": "models/M1/scene.blend",
+        "sha256": "80f985ad09591bdb227e99e71f6bacaa0b9e8c8b565c27da64ad4ac4ceb8910e",
+        "bytes": 2691023
+      }
+    },
+    "M2": {
+      "glb": {
+        "source": "astra_blender/models/M2/scene.glb",
+        "published": "models/M2/scene.glb",
+        "sha256": "0ee9123f38099d1df63ab593dd7082d5f1cf1b136e63833ec5afa0e072cf4c2e",
+        "bytes": 14523124
+      },
+      "blend": {
+        "source": "astra_blender/models/M2/scene.blend",
+        "published": "models/M2/scene.blend",
+        "sha256": "c19f33ec434bbcf89137e67d81c283b9adfd48f42e993d8b29d63e08318f79a6",
+        "bytes": 3886116
+      }
+    },
+    "M3": {
+      "glb": {
+        "source": "astra_blender/models/M3/scene.glb",
+        "published": "models/M3/scene.glb",
+        "sha256": "22699f8225a5f90166cdea2e6e2bdad7977fda4a1fe1086da209f0c9afe97a6c",
+        "bytes": 7701488
+      },
+      "blend": {
+        "source": "astra_blender/models/M3/scene.blend",
+        "published": "models/M3/scene.blend",
+        "sha256": "00cd43eef3b4a6b2b2b169e8a7eda787dc57222dc501a3a9be09f39af5f46a9b",
+        "bytes": 2319452
+      }
+    },
+    "M4": {
+      "glb": {
+        "source": "astra_blender/models/M4/scene.glb",
+        "published": "models/M4/scene.glb",
+        "sha256": "716e4b89b7b11ad8a0cb0e238b366b4f5dc80e3a2c3cddb5eedc469ef4e63774",
+        "bytes": 17202608
+      },
+      "blend": {
+        "source": "astra_blender/models/M4/scene.blend",
+        "published": "models/M4/scene.blend",
+        "sha256": "cc6cb605246a255d94e36b9dc3f8b91d5a292b047470bc0ff75b8a5c404b0cfe",
+        "bytes": 2268620
+      }
+    }
+  },
+  "gt_asset": {
+    "source_url": "https://wentingw.github.io/astra-world-model-blog/assets/comparison/GT.glb",
+    "published": "assets/comparison/GT.glb",
+    "sha256": "6c28ab067c63c5b1fce19c6a972f3d66c7ac4d9bd9ce3b5a57e2e18f5641fe8f",
+    "bytes": 24816108
+  },
+  "fixed_views_manifest": "data/fixed_views.json",
+  "scene_comparison_manifest": "data/scene_comparison.json",
+  "office_cafe_case_study": {
+    "office_cafe": {
+      "source_url": "https://office-cafe-vipe.hiwtishere.chatgpt.site/",
+      "presentation": "local_models_and_synchronized_original_videos",
+      "included_assets": [
+        "original_model",
+        "legacy_model",
+        "input_video",
+        "procedural_model_video",
+        "scan_video",
+        "strong_shake"
+      ],
+      "video_manifest": "data/office_video_media.json"
+    }
+  },
+  "vendor": {
+    "three.module.js": {
+      "source": "world_model_blog/private_repo/sceneweft_sync_20260925/blog/vendor/three/three.module.js",
+      "published": "vendor/three/three.module.js",
+      "sha256": "c8211c69345d2e9949dc7a8ac969380497aa0600a5a8ac6a459c8cd02dd9cb8a",
+      "bytes": 603113
+    },
+    "three.core.js": {
+      "source": "world_model_blog/private_repo/sceneweft_sync_20260925/blog/vendor/three/three.core.js",
+      "published": "vendor/three/three.core.js",
+      "sha256": "eb077d2417f61d3e6d9264c317cabc4ea35769ed6b0ab533067292a550784c20",
+      "bytes": 1403455
+    },
+    "loaders/GLTFLoader.js": {
+      "source": "world_model_blog/private_repo/sceneweft_sync_20260925/blog/vendor/three/loaders/GLTFLoader.js",
+      "published": "vendor/three/loaders/GLTFLoader.js",
+      "sha256": "67ac5551fdafa6e349bd80c8f8e5e39c136d6b2fb1ad647db9abb21dc86f9e4a",
+      "bytes": 114739
+    },
+    "controls/OrbitControls.js": {
+      "source": "world_model_blog/private_repo/sceneweft_sync_20260925/blog/vendor/three/controls/OrbitControls.js",
+      "published": "vendor/three/controls/OrbitControls.js",
+      "sha256": "b97879c748170baadeb3fb84cea1ffdf4674e283dc06042f34e2acb95a76042c",
+      "bytes": 38703
+    },
+    "environments/RoomEnvironment.js": {
+      "source": "world_model_blog/private_repo/sceneweft_sync_20260925/blog/vendor/three/environments/RoomEnvironment.js",
+      "published": "vendor/three/environments/RoomEnvironment.js",
+      "sha256": "c20f0b4677f6128a138d7152b85cbef9091f4f45cdfa05adca04d40c1697c7ae",
+      "bytes": 4899
+    },
+    "utils/BufferGeometryUtils.js": {
+      "source": "world_model_blog/private_repo/sceneweft_sync_20260925/blog/vendor/three/utils/BufferGeometryUtils.js",
+      "published": "vendor/three/utils/BufferGeometryUtils.js",
+      "sha256": "fda7e946b8e0b5ab39b779206589e7a1079a22eb24efb89d7223e03fdfb1f751",
+      "bytes": 35539
+    }
+  },
+  "tables": [
+    {
+      "number": 1,
+      "key": "table1",
+      "methods": [
+        "M1",
+        "M2",
+        "M3",
+        "M4"
+      ],
+      "columns": [
+        "method",
+        "allowed_inputs",
+        "geometry_source",
+        "modeling_role"
+      ],
+      "url": "index.html#table-1",
+      "english_url": "en.html#table-1"
+    },
+    {
+      "number": 2,
+      "key": "table2",
+      "methods": [
+        "M1",
+        "M2",
+        "M3",
+        "M4"
+      ],
+      "columns": [
+        "method",
+        "ate_m",
+        "rotation_deg",
+        "native_depth_absrel",
+        "model_depth_absrel",
+        "alignment"
+      ],
+      "url": "index.html#table-2",
+      "english_url": "en.html#table-2"
+    },
+    {
+      "number": 3,
+      "key": "table3",
+      "methods": [
+        "M1",
+        "M2",
+        "M3",
+        "M4"
+      ],
+      "columns": [
+        "method",
+        "model_to_gt_mean_m",
+        "observed_gt_to_model_mean_m",
+        "model_sha256",
+        "alignment",
+        "bidirectional_mean_m"
+      ],
+      "url": "index.html#table-3",
+      "english_url": "en.html#table-3"
+    },
+    {
+      "number": 4,
+      "key": "table4",
+      "methods": [
+        "M1",
+        "M2",
+        "M3",
+        "M4"
+      ],
+      "columns": [
+        "method",
+        "psnr_db",
+        "ssim",
+        "lpips_alex_v01"
+      ],
+      "url": "index.html#table-4",
+      "english_url": "en.html#table-4"
+    },
+    {
+      "number": 5,
+      "key": "table5",
+      "methods": [
+        "M1",
+        "M2",
+        "M3",
+        "M4"
+      ],
+      "columns": [
+        "method",
+        "psnr_db",
+        "ssim",
+        "lpips_alex_v01"
+      ],
+      "url": "index.html#table-5",
+      "english_url": "en.html#table-5"
+    },
+    {
+      "number": 6,
+      "key": "table6",
+      "methods": [
+        "M1",
+        "M2",
+        "M3",
+        "M4"
+      ],
+      "columns": [
+        "method",
+        "absrel",
+        "rmse_m",
+        "coverage",
+        "penalized_mae_m"
+      ],
+      "url": "index.html#table-6",
+      "english_url": "en.html#table-6"
+    },
+    {
+      "number": 7,
+      "key": "table7",
+      "methods": [
+        "M1",
+        "M2",
+        "M3",
+        "M4"
+      ],
+      "columns": [
+        "method",
+        "absrel",
+        "rmse_m",
+        "coverage",
+        "penalized_mae_m"
+      ],
+      "url": "index.html#table-7",
+      "english_url": "en.html#table-7"
+    }
+  ],
+  "presentation": {
+    "table2_hidden_columns": [
+      "alignment"
+    ],
+    "table3_hidden_columns": [
+      "model_sha256",
+      "alignment"
+    ],
+    "figure1_layout": "single supplied teaser image",
+    "metric_best_values": "bold, raw-value comparison"
+  },
+  "protocols": {
+    "table2": "180 fixed views × 19,200 pixels; optical-Z; GT-valid domain; no per-view fitting",
+    "table3": {
+      "model_surface_samples": 100000,
+      "model_surface_seed": 20260923,
+      "observed_gt_samples": 100000,
+      "observed_gt_seed": 20260925,
+      "distance": "exact nearest triangle",
+      "bidirectional_mean": "arithmetic mean of model→GT and observed GT→model mean distances"
+    },
+    "table4": {
+      "views": [
+        0,
+        36,
+        72,
+        108,
+        144
+      ],
+      "resolution": [
+        640,
+        480
+      ],
+      "gt_resize": "Lanczos 1280x960 to 640x480",
+      "metrics": "full-image PSNR, SSIM, LPIPS AlexNet v0.1; no crop/mask/fitting"
+    },
+    "table5": {
+      "scope": "100 seeded views drawn without replacement from the 175 modeling frames remaining after excluding fixed views; these are not unseen modeling inputs",
+      "seed": 20260930,
+      "excluded_fixed_views": [
+        0,
+        36,
+        72,
+        108,
+        144
+      ],
+      "selected_indices": [
+        1,
+        2,
+        3,
+        5,
+        9,
+        12,
+        13,
+        14,
+        16,
+        18,
+        19,
+        21,
+        22,
+        25,
+        26,
+        27,
+        28,
+        29,
+        30,
+        31,
+        32,
+        33,
+        34,
+        38,
+        40,
+        42,
+        43,
+        44,
+        45,
+        46,
+        50,
+        52,
+        55,
+        56,
+        58,
+        59,
+        60,
+        61,
+        62,
+        63,
+        64,
+        67,
+        70,
+        71,
+        73,
+        75,
+        76,
+        78,
+        79,
+        81,
+        83,
+        84,
+        85,
+        87,
+        89,
+        93,
+        94,
+        96,
+        98,
+        99,
+        101,
+        102,
+        103,
+        104,
+        106,
+        107,
+        109,
+        112,
+        113,
+        117,
+        118,
+        119,
+        120,
+        122,
+        123,
+        124,
+        127,
+        133,
+        135,
+        136,
+        137,
+        138,
+        142,
+        146,
+        148,
+        149,
+        151,
+        152,
+        155,
+        157,
+        160,
+        164,
+        166,
+        167,
+        168,
+        169,
+        170,
+        172,
+        177,
+        179
+      ],
+      "resolution": [
+        640,
+        480
+      ],
+      "preprocessing": "GT RGB Lanczos 1280x960 to 640x480; RGB float32 [0,1]; no crop, mask, or color/exposure fitting",
+      "metrics": "full-image PSNR, SSIM, LPIPS AlexNet v0.1",
+      "aggregation": "arithmetic mean over the same frozen 100 views"
+    },
+    "table6": {
+      "views": [
+        0,
+        36,
+        72,
+        108,
+        144
+      ],
+      "depth": {
+        "depth": "optical Z in metres",
+        "domain": "GT valid_domain and finite GT in [0.1, 30] m",
+        "prediction_valid": "finite prediction in [0.1, 30] m",
+        "aggregation": "pixel-weighted over all selected frames",
+        "missing_penalty": "30 m absolute error for each missing prediction"
+      }
+    },
+    "table7": {
+      "scope": "the exact same 100 modeling-frame indices as Table 5; held out only from the fixed-five check",
+      "seed": 20260930,
+      "selected_indices": [
+        1,
+        2,
+        3,
+        5,
+        9,
+        12,
+        13,
+        14,
+        16,
+        18,
+        19,
+        21,
+        22,
+        25,
+        26,
+        27,
+        28,
+        29,
+        30,
+        31,
+        32,
+        33,
+        34,
+        38,
+        40,
+        42,
+        43,
+        44,
+        45,
+        46,
+        50,
+        52,
+        55,
+        56,
+        58,
+        59,
+        60,
+        61,
+        62,
+        63,
+        64,
+        67,
+        70,
+        71,
+        73,
+        75,
+        76,
+        78,
+        79,
+        81,
+        83,
+        84,
+        85,
+        87,
+        89,
+        93,
+        94,
+        96,
+        98,
+        99,
+        101,
+        102,
+        103,
+        104,
+        106,
+        107,
+        109,
+        112,
+        113,
+        117,
+        118,
+        119,
+        120,
+        122,
+        123,
+        124,
+        127,
+        133,
+        135,
+        136,
+        137,
+        138,
+        142,
+        146,
+        148,
+        149,
+        151,
+        152,
+        155,
+        157,
+        160,
+        164,
+        166,
+        167,
+        168,
+        169,
+        170,
+        172,
+        177,
+        179
+      ],
+      "depth": {
+        "depth": "optical Z in metres",
+        "domain": "GT valid_domain and finite GT in [0.1, 30] m",
+        "prediction_valid": "finite prediction in [0.1, 30] m",
+        "aggregation": "pixel-weighted over all selected frames",
+        "missing_penalty": "30 m absolute error for each missing prediction"
+      }
+    }
+  },
+  "m1_notice": "M1 model-space metrics use one frozen GT-assisted Sim(3) from five manual camera associations. They are diagnostic and do not represent native metric recovery."
+}

data/tables_1_7.json

@@ -0,0 +1,508 @@
+{
+  "schema_version": 2,
+  "status": "COMPLETE",
+  "scope": "Only frozen models under astra_blender/models/M1..M4",
+  "reference": "https://wentingw.github.io/astra-world-model-blog/#_5",
+  "m1_notice": "M1 model-space metrics use one frozen GT-assisted Sim(3) from five manual camera associations. They are diagnostic and do not represent native metric recovery.",
+  "protocols": {
+    "table2": "180 fixed views × 19,200 pixels; optical-Z; GT-valid domain; no per-view fitting",
+    "table3": {
+      "model_surface_samples": 100000,
+      "model_surface_seed": 20260923,
+      "observed_gt_samples": 100000,
+      "observed_gt_seed": 20260925,
+      "distance": "exact nearest triangle",
+      "bidirectional_mean": "arithmetic mean of model→GT and observed GT→model mean distances"
+    },
+    "table4": {
+      "views": [
+        0,
+        36,
+        72,
+        108,
+        144
+      ],
+      "resolution": [
+        640,
+        480
+      ],
+      "gt_resize": "Lanczos 1280x960 to 640x480",
+      "metrics": "full-image PSNR, SSIM, LPIPS AlexNet v0.1; no crop/mask/fitting"
+    },
+    "table5": {
+      "scope": "100 seeded views drawn without replacement from the 175 modeling frames remaining after excluding fixed views; these are not unseen modeling inputs",
+      "seed": 20260930,
+      "excluded_fixed_views": [
+        0,
+        36,
+        72,
+        108,
+        144
+      ],
+      "selected_indices": [
+        1,
+        2,
+        3,
+        5,
+        9,
+        12,
+        13,
+        14,
+        16,
+        18,
+        19,
+        21,
+        22,
+        25,
+        26,
+        27,
+        28,
+        29,
+        30,
+        31,
+        32,
+        33,
+        34,
+        38,
+        40,
+        42,
+        43,
+        44,
+        45,
+        46,
+        50,
+        52,
+        55,
+        56,
+        58,
+        59,
+        60,
+        61,
+        62,
+        63,
+        64,
+        67,
+        70,
+        71,
+        73,
+        75,
+        76,
+        78,
+        79,
+        81,
+        83,
+        84,
+        85,
+        87,
+        89,
+        93,
+        94,
+        96,
+        98,
+        99,
+        101,
+        102,
+        103,
+        104,
+        106,
+        107,
+        109,
+        112,
+        113,
+        117,
+        118,
+        119,
+        120,
+        122,
+        123,
+        124,
+        127,
+        133,
+        135,
+        136,
+        137,
+        138,
+        142,
+        146,
+        148,
+        149,
+        151,
+        152,
+        155,
+        157,
+        160,
+        164,
+        166,
+        167,
+        168,
+        169,
+        170,
+        172,
+        177,
+        179
+      ],
+      "resolution": [
+        640,
+        480
+      ],
+      "preprocessing": "GT RGB Lanczos 1280x960 to 640x480; RGB float32 [0,1]; no crop, mask, or color/exposure fitting",
+      "metrics": "full-image PSNR, SSIM, LPIPS AlexNet v0.1",
+      "aggregation": "arithmetic mean over the same frozen 100 views"
+    },
+    "table6": {
+      "views": [
+        0,
+        36,
+        72,
+        108,
+        144
+      ],
+      "depth": {
+        "depth": "optical Z in metres",
+        "domain": "GT valid_domain and finite GT in [0.1, 30] m",
+        "prediction_valid": "finite prediction in [0.1, 30] m",
+        "aggregation": "pixel-weighted over all selected frames",
+        "missing_penalty": "30 m absolute error for each missing prediction"
+      }
+    },
+    "table7": {
+      "scope": "the exact same 100 modeling-frame indices as Table 5; held out only from the fixed-five check",
+      "seed": 20260930,
+      "selected_indices": [
+        1,
+        2,
+        3,
+        5,
+        9,
+        12,
+        13,
+        14,
+        16,
+        18,
+        19,
+        21,
+        22,
+        25,
+        26,
+        27,
+        28,
+        29,
+        30,
+        31,
+        32,
+        33,
+        34,
+        38,
+        40,
+        42,
+        43,
+        44,
+        45,
+        46,
+        50,
+        52,
+        55,
+        56,
+        58,
+        59,
+        60,
+        61,
+        62,
+        63,
+        64,
+        67,
+        70,
+        71,
+        73,
+        75,
+        76,
+        78,
+        79,
+        81,
+        83,
+        84,
+        85,
+        87,
+        89,
+        93,
+        94,
+        96,
+        98,
+        99,
+        101,
+        102,
+        103,
+        104,
+        106,
+        107,
+        109,
+        112,
+        113,
+        117,
+        118,
+        119,
+        120,
+        122,
+        123,
+        124,
+        127,
+        133,
+        135,
+        136,
+        137,
+        138,
+        142,
+        146,
+        148,
+        149,
+        151,
+        152,
+        155,
+        157,
+        160,
+        164,
+        166,
+        167,
+        168,
+        169,
+        170,
+        172,
+        177,
+        179
+      ],
+      "depth": {
+        "depth": "optical Z in metres",
+        "domain": "GT valid_domain and finite GT in [0.1, 30] m",
+        "prediction_valid": "finite prediction in [0.1, 30] m",
+        "aggregation": "pixel-weighted over all selected frames",
+        "missing_penalty": "30 m absolute error for each missing prediction"
+      }
+    }
+  },
+  "tables": {
+    "table1": [
+      {
+        "method": "M1",
+        "allowed_inputs": "180 张采样 RGB",
+        "geometry_source": "视觉推断，无测量尺度",
+        "modeling_role": "推断布局，编写对象与材质"
+      },
+      {
+        "method": "M2",
+        "allowed_inputs": "完整 RGB 视频；选取 180 帧建模",
+        "geometry_source": "ViPE RGB-only 位姿 → pose-conditioned DA3 深度",
+        "modeling_role": "将位姿与深度测量组织为对象化场景"
+      },
+      {
+        "method": "M3",
+        "allowed_inputs": "视频、IMU、相机–IMU 标定",
+        "geometry_source": "ORB-SLAM3 单目惯性位姿 → pose-conditioned DA3 深度",
+        "modeling_role": "结合图像、位姿、深度建模"
+      },
+      {
+        "method": "M4",
+        "allowed_inputs": "视频及 GT camera pose",
+        "geometry_source": "GT 位姿 → pose-conditioned DA3 深度",
+        "modeling_role": "与 M3 相同类型的建模流程"
+      }
+    ],
+    "table2": [
+      {
+        "method": "M1",
+        "ate_m": null,
+        "rotation_deg": null,
+        "native_depth_absrel": null,
+        "model_depth_absrel": 0.16439823806285858,
+        "alignment": "GT-assisted Sim(3), diagnostic only"
+      },
+      {
+        "method": "M2",
+        "ate_m": 0.1668,
+        "rotation_deg": 0.3065,
+        "native_depth_absrel": 0.19056472003176061,
+        "model_depth_absrel": 0.09286404401063919,
+        "alignment": "one global SE(3), scale=1"
+      },
+      {
+        "method": "M3",
+        "ate_m": 0.1205,
+        "rotation_deg": 0.1943,
+        "native_depth_absrel": 0.20379653842953233,
+        "model_depth_absrel": 0.08371810615062714,
+        "alignment": "one global SE(3), scale=1"
+      },
+      {
+        "method": "M4",
+        "ate_m": "-",
+        "rotation_deg": "-",
+        "native_depth_absrel": 0.1886706026379072,
+        "model_depth_absrel": 0.07600006461143494,
+        "alignment": "GT coordinates"
+      }
+    ],
+    "table3": [
+      {
+        "method": "M1",
+        "model_to_gt_mean_m": 0.34486243741226075,
+        "observed_gt_to_model_mean_m": 0.40664686199496186,
+        "model_sha256": "80f985ad09591bdb227e99e71f6bacaa0b9e8c8b565c27da64ad4ac4ceb8910e",
+        "alignment": "GT-assisted Sim3 from frozen manual-view camera associations; shape/display diagnostic only; not metric recovery; no mesh ICP",
+        "bidirectional_mean_m": 0.3757546497036113
+      },
+      {
+        "method": "M2",
+        "model_to_gt_mean_m": 0.13236121125603983,
+        "observed_gt_to_model_mean_m": 0.10298419313248852,
+        "model_sha256": "c19f33ec434bbcf89137e67d81c283b9adfd48f42e993d8b29d63e08318f79a6",
+        "alignment": "one SE3 from 180 input camera correspondences; no scale fit",
+        "bidirectional_mean_m": 0.11767270219426418
+      },
+      {
+        "method": "M3",
+        "model_to_gt_mean_m": 0.06935983382743138,
+        "observed_gt_to_model_mean_m": 0.09502824761255278,
+        "model_sha256": "00cd43eef3b4a6b2b2b169e8a7eda787dc57222dc501a3a9be09f39af5f46a9b",
+        "alignment": "one SE3 from 180 input camera correspondences; no scale fit",
+        "bidirectional_mean_m": 0.08219404071999209
+      },
+      {
+        "method": "M4",
+        "model_to_gt_mean_m": 0.06866006729939131,
+        "observed_gt_to_model_mean_m": 0.07449522976921306,
+        "model_sha256": "cc6cb605246a255d94e36b9dc3f8b91d5a292b047470bc0ff75b8a5c404b0cfe",
+        "alignment": "permitted GT input coordinate frame; astra_blender/models/M4 T_input_model is identity",
+        "bidirectional_mean_m": 0.07157764853430218
+      }
+    ],
+    "table4": [
+      {
+        "method": "M1",
+        "psnr_db": 10.260909193969148,
+        "ssim": 0.3091735512018204,
+        "lpips_alex_v01": 0.6575803160667419
+      },
+      {
+        "method": "M2",
+        "psnr_db": 12.95293150317823,
+        "ssim": 0.3896766543388367,
+        "lpips_alex_v01": 0.5310081720352173
+      },
+      {
+        "method": "M3",
+        "psnr_db": 12.052547333787004,
+        "ssim": 0.3958937466144562,
+        "lpips_alex_v01": 0.5791207671165466
+      },
+      {
+        "method": "M4",
+        "psnr_db": 12.294263798590851,
+        "ssim": 0.34136943221092225,
+        "lpips_alex_v01": 0.531274390220642
+      }
+    ],
+    "table5": [
+      {
+        "method": "M1",
+        "psnr_db": 11.691789500467046,
+        "ssim": 0.31951158225536347,
+        "lpips_alex_v01": 0.6073989069461823
+      },
+      {
+        "method": "M2",
+        "psnr_db": 13.207560508419194,
+        "ssim": 0.3776748192310333,
+        "lpips_alex_v01": 0.5329300162196159
+      },
+      {
+        "method": "M3",
+        "psnr_db": 12.252687193116532,
+        "ssim": 0.3972503730654717,
+        "lpips_alex_v01": 0.5683315402269363
+      },
+      {
+        "method": "M4",
+        "psnr_db": 12.513978931521383,
+        "ssim": 0.33674563601613045,
+        "lpips_alex_v01": 0.5237806642055511
+      }
+    ],
+    "table6": [
+      {
+        "method": "M1",
+        "absrel": 0.2901068925857544,
+        "rmse_m": 3.049535036087036,
+        "coverage": 0.9996145833333333,
+        "penalized_mae_m": 1.88661527633667
+      },
+      {
+        "method": "M2",
+        "absrel": 0.09218113869428635,
+        "rmse_m": 1.116154432296753,
+        "coverage": 0.9978125,
+        "penalized_mae_m": 0.577746570110321
+      },
+      {
+        "method": "M3",
+        "absrel": 0.09866034984588623,
+        "rmse_m": 1.1816775798797607,
+        "coverage": 0.9984166666666666,
+        "penalized_mae_m": 0.5551674365997314
+      },
+      {
+        "method": "M4",
+        "absrel": 0.08832013607025146,
+        "rmse_m": 1.1883933544158936,
+        "coverage": 0.9985833333333334,
+        "penalized_mae_m": 0.5010335445404053
+      }
+    ],
+    "table7": [
+      {
+        "method": "M1",
+        "absrel": 0.164479598402977,
+        "rmse_m": 1.7184345722198486,
+        "coverage": 0.9924354166666667,
+        "penalized_mae_m": 1.2859067916870117
+      },
+      {
+        "method": "M2",
+        "absrel": 0.09284797310829163,
+        "rmse_m": 1.1448776721954346,
+        "coverage": 0.9977067708333334,
+        "penalized_mae_m": 0.6155765056610107
+      },
+      {
+        "method": "M3",
+        "absrel": 0.08247262239456177,
+        "rmse_m": 1.0896246433258057,
+        "coverage": 0.9991479166666667,
+        "penalized_mae_m": 0.4945947229862213
+      },
+      {
+        "method": "M4",
+        "absrel": 0.07466284930706024,
+        "rmse_m": 1.1474982500076294,
+        "coverage": 0.998678125,
+        "penalized_mae_m": 0.47199496626853943
+      }
+    ]
+  },
+  "source_files": {
+    "base_tables_1_5": {
+      "path": "astra_blender2/report/space/results/blog_tables_1_5/tables_1_5.json",
+      "sha256": "35379c83f758cacad2b9a07c1d1159f332bf109e94cf3cddd7e6ccde0d0b8270"
+    },
+    "view_subset_manifest": {
+      "path": "astra_blender2/evaluation/blog_extended_100_20260930/view_subset_manifest.json",
+      "sha256": "f03f7561b01fe1132dcbc16890e9abb10af73b5c4589f75b38aa698a3819feca"
+    },
+    "appearance_metrics": {
+      "path": "astra_blender2/evaluation/blog_extended_100_20260930/appearance_metrics.json",
+      "sha256": "bd8a3e4cd06ef42030a1e49008b321b4ee51f134059c681d2a709bf5e39874fd"
+    },
+    "depth_metrics": {
+      "path": "astra_blender2/evaluation/blog_extended_100_20260930/depth_metrics.json",
+      "sha256": "709809701db36c49ab848d665da330a7cb213cf26675212dbcf214d7e4cd5b24"
+    }
+  }
+}

editorial.css

@@ -0,0 +1,66 @@
+:root{--masthead-width:1080px;--masthead-gutter:40px}
+.insight-strip{grid-template-columns:repeat(2,minmax(0,1fr))}
+.insight-strip .metric-reduction{display:block;margin-top:12px;color:var(--green);font-size:13px;font-weight:600;line-height:1.5}
+@media(max-width:700px){.insight-strip{grid-template-columns:1fr}}
+.office-links{display:flex;flex-wrap:wrap;gap:12px 24px;font-size:14px}
+body>nav,body>header,body>.hero{max-width:var(--masthead-width);padding-left:var(--masthead-gutter);padding-right:var(--masthead-gutter);margin-left:auto;margin-right:auto}
+body>nav{padding-top:24px;padding-bottom:24px;border-bottom:0;position:relative}
+body>nav::after{content:"";position:absolute;bottom:0;left:var(--masthead-gutter);right:var(--masthead-gutter);height:1px;background:var(--line)}
+body>header{margin-top:48px;margin-bottom:34px}
+header .title-name{margin:0 0 14px;color:var(--green);font-size:clamp(14px,1.35vw,17px);font-weight:700;letter-spacing:.18em;line-height:1.6;text-transform:uppercase}
+header .title-lockup{display:flex;align-items:baseline;flex-wrap:wrap;gap:0 34px}
+header h1{margin:0;color:#182f28;font-family:Georgia,"Times New Roman",serif;font-size:clamp(104px,16vw,190px);font-weight:500;letter-spacing:-.045em;line-height:1.04}
+header .pronunciation{margin:0;color:#7c8477;font:italic 400 clamp(22px,2.5vw,32px)/1.4 Georgia,"Noto Serif CJK SC",serif;white-space:nowrap}
+header .lead{max-width:none;margin:24px 0 0;color:#182f28;font-family:Georgia,"Noto Serif CJK SC",serif;font-size:clamp(26px,3.15vw,40px);font-weight:500;line-height:1.35;letter-spacing:-.02em;text-wrap:pretty}
+header .lead .tagline-emphasis{color:var(--green);font-weight:inherit;text-decoration:none}
+header .brand-note{margin:16px 0 0;color:var(--muted);font-size:14px;line-height:1.6}
+body[data-lang="zh"] header .title-name{letter-spacing:.16em}
+body[data-lang="zh"] header .lead{line-height:1.6;letter-spacing:0}
+header .publication-date{margin:20px 0 0;color:#60736b;font-size:12px;letter-spacing:.08em;text-transform:uppercase}
+body>.hero{margin-top:0;margin-bottom:72px}
+body>.hero>video{display:block;width:100%;height:auto;aspect-ratio:16/9;background:#182c29;border:1px solid #dbe2d9;border-radius:5px;box-shadow:0 12px 36px -24px #243b3033}
+.project-links{display:flex;justify-content:center;align-items:center;flex-wrap:wrap;gap:12px;margin-top:24px}
+.project-links .project-resource{display:inline-flex;align-items:center;justify-content:center;gap:9px;min-height:44px;padding:9px 18px;border:1px solid var(--line);border-radius:999px;background:var(--white);color:var(--ink);font-size:14px;font-weight:600;text-decoration:none}
+.project-links img{display:block;width:24px;height:24px;object-fit:contain;background:transparent}
+.project-links a{position:relative;padding-right:26px}
+.project-external-arrow{position:absolute;top:6px;right:6px;width:12px;height:12px;color:var(--green);pointer-events:none}
+.project-links a:hover{border-color:#8eaa9e;background:#edf1e9}
+.project-links a:focus-visible{outline:2px solid var(--green);outline-offset:4px}
+.finding-comparison{border-bottom:0;padding-bottom:0;margin-bottom:0}
+.finding-comparison + .analysis-block{border-top:0;padding-top:0;margin-top:24px}
+header .title-name,header .title-lockup,header .lead{animation:masthead-enter .7s cubic-bezier(.22,1,.36,1) both}
+header .lead{animation-delay:.08s}
+.table-figure tbody th{white-space:normal;min-width:155px;max-width:215px;font-weight:600}
+@keyframes masthead-enter{from{opacity:0;transform:translateY(10px)}to{opacity:1;transform:translateY(0)}}
+.citation-block{border-radius:5px;border-color:#dbe2d9;background:#fbfcf8;white-space:pre-wrap;overflow-wrap:anywhere}
+.citation-download{display:inline-flex;align-items:center;gap:20px;min-height:44px;padding:8px 15px;border:1px solid var(--line);border-radius:4px;font-size:13px;font-weight:600;text-decoration:none;transition:background-color .18s ease,border-color .18s ease}
+.citation-download:hover{background:#edf1e9;border-color:#8eaa9e}
+.citation-download:focus-visible{outline:2px solid var(--green);outline-offset:4px}
+.embodied-demo{scroll-margin-top:30px}
+.mission-command{margin:28px 0 20px;padding:22px 26px;border-left:3px solid var(--green);background:#edf1e9;font-size:clamp(20px,2.4vw,27px);line-height:1.6}
+.mission-command>span{display:block;margin-bottom:8px;font:700 11px/1.5 system-ui,sans-serif;letter-spacing:.12em;color:var(--green)}
+.mission-flow{display:grid;grid-template-columns:repeat(4,minmax(0,1fr));gap:0;list-style:none;padding:0;margin:24px 0;border-top:1px solid var(--line);border-bottom:1px solid var(--line)}
+.mission-flow li{padding:14px 12px;font-size:13px;line-height:1.7}
+.mission-flow li+li{border-left:1px solid var(--line)}
+.mission-flow span{display:block;color:var(--green);font-size:11px;letter-spacing:.12em}
+.demo-film{margin:24px 0 14px}
+.demo-film video{display:block;width:100%;height:auto;aspect-ratio:16/9;background:#15231f;border:1px solid var(--line)}
+.demo-film figcaption{font-size:13px;line-height:1.8}
+.demo-map{margin:28px 0 20px}
+.demo-map img{width:100%;height:auto;border:1px solid var(--line);border-radius:4px}
+.demo-map a{display:block}
+.demo-map a:focus-visible{outline:2px solid var(--green);outline-offset:4px}
+.demo-links{display:flex;flex-wrap:wrap;justify-content:space-between;gap:8px 22px;font-size:13px}
+.demo-protocol{margin:20px 0 0;border-top:1px solid var(--line);padding:14px 0;font-size:14px;color:var(--muted)}
+.demo-protocol summary{cursor:pointer;color:var(--green);font-weight:650;min-height:30px}
+.demo-protocol p{font-size:14px}
+.grounding-block{margin-top:32px;padding-top:22px;border-top:1px solid var(--line)}
+.spatial-foundation{margin-top:38px;padding-top:25px;border-top:1px solid var(--line)}
+.capability-grid{display:grid;grid-template-columns:repeat(2,minmax(0,1fr));gap:0 28px;margin:18px 0}
+.capability-grid>div{padding:12px 0;border-top:1px solid var(--line)}
+.capability-grid h4{margin:0 0 8px;font-size:17px;color:var(--green)}
+.capability-grid p{margin:0;font-size:15px;line-height:1.85}
+.demo-protocol summary:focus-visible,.demo-links a:focus-visible{outline:2px solid var(--green);outline-offset:4px}
+@media(max-width:800px){:root{--masthead-gutter:28px}body>header{margin-top:40px}header .title-lockup{column-gap:24px}}
+@media(max-width:600px){:root{--masthead-gutter:22px}body>header{margin-top:32px;margin-bottom:28px}header h1{font-size:clamp(82px,24vw,140px)}header .title-name{font-size:14px;letter-spacing:.13em}header .title-lockup{gap:8px 20px}header .pronunciation{font-size:24px}header .lead{margin-top:24px;font-size:28px}body>.hero{margin-bottom:44px}nav{gap:15px;flex-wrap:wrap}.brand{font-size:11px}.mission-command{padding:18px}.mission-flow{grid-template-columns:repeat(2,minmax(0,1fr))}.mission-flow li:nth-child(3){border-left:0}.mission-flow li:nth-child(n+3){border-top:1px solid var(--line)}.capability-grid{grid-template-columns:1fr}.demo-links{flex-direction:column}}
+@media(prefers-reduced-motion:reduce){html{scroll-behavior:auto}header .title-name,header .title-lockup,header .lead{animation:none}.citation-download{transition:none}}

en.html

@@ -0,0 +1,34 @@
+<!doctype html><html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"><title>AWSM: Agentic World Simulation and Mapping — Phygital AI</title><meta name="description" content="From real spaces to worlds phygital agents can use."><meta property="og:type" content="article"><meta property="og:locale" content="en_US"><meta property="og:title" content="AWSM: Agentic World Simulation and Mapping"><meta property="og:description" content="From real spaces to worlds phygital agents can use."><meta property="og:image" content="https://phygital-ai.github.io/agentic-world-simulation-and-mapping/assets/teaser_originals.png"><meta property="og:url" content="https://phygital-ai.github.io/agentic-world-simulation-and-mapping/index.html"><meta name="twitter:card" content="summary_large_image"><meta name="twitter:title" content="AWSM: Agentic World Simulation and Mapping"><meta name="twitter:description" content="From real spaces to worlds phygital agents can use."><meta name="twitter:image" content="https://phygital-ai.github.io/agentic-world-simulation-and-mapping/assets/teaser_originals.png"><link rel="canonical" href="https://phygital-ai.github.io/agentic-world-simulation-and-mapping/"><link rel="alternate" hreflang="zh-CN" href="https://phygital-ai.github.io/agentic-world-simulation-and-mapping/zh.html"><link rel="alternate" hreflang="en" href="https://phygital-ai.github.io/agentic-world-simulation-and-mapping/"><link rel="stylesheet" href="style.css?v=cbba47743ba2"><link rel="stylesheet" href="editorial.css?v=462f0ffc2b51"><script type="importmap">{"imports":{"three":"./vendor/three/three.module.js"}}</script><script src="app.js?v=d70db3a6816b" defer></script><script type="module" src="scene-compare.js?v=bd749465e05d"></script><link rel="stylesheet" href="office.css?v=2e8c4a2f0718"><link rel="alternate" hreflang="x-default" href="https://phygital-ai.github.io/agentic-world-simulation-and-mapping/"><meta property="article:published_time" content="2026-10-01"></head>
+<body data-lang="en"><a class="skip" href="#main">Skip to article</a><nav><a class="brand" href="index.html">PHYGITAL AI / AWSM</a><div><a href="zh.html" >中文</a><a href="index.html" aria-current='page'>EN</a><a href="https://github.com/wentingw/AWSM" target="_blank" rel="noopener noreferrer">GitHub ↗</a></div></nav><header><p class="title-name">Agentic World Simulation and Mapping</p><div class="title-lockup"><h1 aria-label="AWSM: Agentic World Simulation and Mapping">AWSM</h1><p class="pronunciation">pronounced “awesome”</p></div><p class="lead">From real spaces to <strong class="tagline-emphasis">worlds phygital agents can use.</strong></p><p class="brand-note">Phygital = physical + digital.</p><p class="publication-date">Published <time datetime="2026-10-01">October 1, 2026</time></p></header>
+<div class="hero" id="overview-video"><video controls playsinline preload="metadata" poster="assets/awsm-promo-v5-poster.jpg" aria-label="AWSM introduction video"><source src="assets/awsm-promo-v5.mp4" type="video/mp4"><a href="assets/awsm-promo-v5.mp4">AWSM introduction video</a></video><div class="project-links" role="group" aria-label="AWSM project resources"><a class="project-resource" href="https://github.com/wentingw/AWSM" target="_blank" rel="noopener noreferrer" aria-label="GitHub (opens in a new window)" title="GitHub (opens in a new window)"><img src="assets/icons/github.svg" width="24" height="24" alt="">GitHub<svg class="project-external-arrow" viewBox="0 0 16 16" width="12" height="12" fill="none" stroke="currentColor" stroke-width="1.6" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><path d="M4 12 12 4M4 4h8v8"/></svg></a><span class="project-resource"><img src="assets/icons/huggingface.svg" width="24" height="24" alt="">Hugging Face (Coming soon)</span></div></div><div class="layout"><aside class="toc"><p>ON THIS PAGE</p><a href="#results">Overview</a><a href="#motivation">Motivation</a><a href="#workflow">Workflow</a><a href="#more-results">Results & analysis</a><a href="#embodied-demo">Embodied demo</a><a href="#related-work">Related work</a><a href="#limitations">Limitations</a><a href="#citation">Citation</a><a href="#references">References</a></aside><main id="main">
+<section id="results"><p class="section-tag">01 / OVERVIEW</p><aside class="tldr"><strong>TL;DR</strong><p>AWSM uses large-model agents to turn visual observations into editable 3D scenes, grounded in camera-pose and depth estimates—including IMU-informed visual–inertial constraints that tie geometry to physical scale. In one controlled Isaac Sim scene, four reconstruction routes test spatial fidelity, not just visual plausibility. The aligned comparison from RGB-only M1 to GT-pose-conditioned M4 shows ≈81% lower bidirectional surface error and ≈54% lower model-depth AbsRel across 180 evaluation views. A multi-robot demo then illustrates goal-directed navigation and task execution on the reconstructed map: follow predefined routes to a marked reception desk and line up there. Next steps are higher efficiency and accuracy, reusable simulation-ready assets, and deeper integration with phygital agents for persistent spatial memory, object search, and interaction.</p></aside><h2>A scene can look plausible and still be spatially wrong.</h2><p class="standfirst">Agentic reconstruction uses a large model to inspect observations, call modeling tools, and iteratively build and check an editable scene. The challenge is not only to produce a convincing room, but to preserve the scale, shape, and spatial relationships that make it useful beyond rendering. AWSM asks how geometric evidence can constrain this process so that its outputs become more faithful spatial references for downstream agents.</p><p>We reconstruct World Lobby, a simulated scene in NVIDIA Isaac Sim, through four complete engineering routes, from RGB-only modeling to ground-truth-pose-conditioned modeling. Each route produces a frozen Blender scene, not merely a point cloud or a novel-view renderer. The experiment asks what additional geometric evidence buys, where it fails to help, and which errors survive object-centric reconstruction.</p><div class="insight-strip"><div><strong>0.376 → 0.072 m</strong><span>bidirectional surface error</span><small class="metric-reduction">≈81% relative reduction · M1 → M4</small></div><div><strong>16.44% → 7.60%</strong><span>model-depth AbsRel · 180 evaluation views</span><small class="metric-reduction">≈54% relative reduction · M1 → M4</small></div></div><p class="metric-caveat">Single-scene M1 → M4 diagnostic comparison: M1 uses GT-assisted Sim(3) alignment; M4 uses GT camera poses. Relative reductions compare complete pipelines, not IMU alone.</p></section>
+<section id="motivation"><p class="section-tag">02 / MOTIVATION</p><h2>From plausible reconstruction to a usable spatial reference</h2><p>A reconstruction agent can assemble a plausible room while getting a corner, distance, or passage wrong. For an embodied agent, those are not merely visual defects: they change where a destination lies and which route a body can follow. Geometry grounding means constraining the modeling process with pose, depth, and metric evidence where available, rather than relying on visual plausibility alone. The question is how those constraints survive the conversion from observations into editable objects.</p><p>The reception-desk task makes this connection concrete: mark a destination in the reconstructed map, specify routes, and execute them with robot controllers. We use ‘scene representation’ for the editable objects and geometry, ‘map’ for their navigation role, and ‘simulation assets’ for their integration into a simulator. The ‘world’ in AWSM refers to this reusable spatial environment, not a learned dynamics predictor. A persistent spatial foundation connecting navigation, memory, interaction, and simulation is the longer-term goal; the experiments and demo examine its geometric basis and one downstream use.</p><aside class="claim"><strong>Research question</strong><p>How does progressively stronger geometric evidence change the fidelity—and the remaining failure modes—of an editable object-centric world?</p></aside><h3>Contributions</h3><ol class="contributions"><li>A four-route study of geometry-grounded agentic reconstruction, keeping the scene, 180 RGB identities, modeling objective, and editable output format fixed while comparing complete pipelines with different geometric evidence.</li><li>An evaluation protocol that separates trajectory error, native depth, final-scene depth, surface geometry, and rendered appearance, with frozen Blend/GLB assets, reproducible evaluation subsets, public hashes, and interactive comparisons.</li><li>A map-based multi-robot simulation example connecting a reconstructed scene to destination marking, predefined routes, and controller execution. It illustrates downstream use rather than a controlled navigation comparison across reconstruction routes.</li></ol><div id="scale-and-sensors" class="analysis-block grounding-block"><h3>Geometry grounding: scale is necessary, but not sufficient</h3>
+<p>Monocular projection geometry leaves a global scale ambiguity; learned metric depth adds a useful prior, not an independent measurement of every real-world dimension. Calibrated RGB-D, a known baseline or length, and visual-inertial estimation can supply metric constraints. IMU fusion can recover metric motion under sufficient excitation and reliable initialization, bias estimation, synchronization, and calibration. An IMU is not a direct distance sensor, and its presence alone does not guarantee accurate scale.</p>
+<p>Scale, local shape, camera pose, and navigable connectivity must be checked separately. A curved corner simplified into a square cannot be repaired by rescaling alone. Pose-conditioned depth can improve reconstruction, but neither the DA3 ablation nor our M2/M3 results imply that every depth metric improves monotonically. The four routes below are system comparisons, not an isolated IMU ablation.</p>
+<p class="inline-sources"><a href="https://arxiv.org/abs/1708.03852" target="_blank" rel="noopener">VINS-Mono</a> · <a href="https://arxiv.org/html/2511.10647v1#S7.SS2.SSS3" target="_blank" rel="noopener">DA3 pose-conditioning ablation</a></p></div></section>
+<section id="workflow"><p class="section-tag">03 / WORKFLOW</p><h2>One agentic workflow, four reconstruction routes</h2><p>The reconstruction agent follows a shared observe–build–verify loop: inspect the available evidence, write Blender Python to construct objects, render review views, and revise the scene before freezing it for evaluation. This tool-using modeling process is what ‘agentic’ describes here; the downstream robot controllers have a separate role.</p><p>All routes receive the same 180 RGB identities and the same Astra–Blender modeling objective. M1 tests how far visual priors can go without metric measurements. M2 adds RGB-only ViPE poses and pose-conditioned DA3 depth. M3 uses monocular-inertial ORB-SLAM3 poses with the same class of DA3 observations. M4 supplies ground-truth camera poses to diagnose the remaining modeling error; it does not receive the GT mesh or GT depth as modeling input.</p><figure class="table-figure" id="table-1" data-table-key="table1"><figcaption><span>Table 1.</span> <span class="table-title"></span></figcaption><div class="table-scroll table-mount" role="region" tabindex="0" aria-label="Table 1"></div></figure><h3>Agentic reconstruction loop</h3><div class="workflow-strip"><div><strong>01 · Observe</strong><span>Inspect all 180 RGB frames and, where allowed, pose-conditioned DA3 geometry.</span></div><div><strong>02 · Build</strong><span>Write Blender Python that decomposes the room into named objects, materials, cameras, and collision proxies.</span></div><div><strong>03 · Verify</strong><span>Compare paired RGB and optical-Z renders at fixed review cameras; revise within a fixed version budget.</span></div><div><strong>04 · Freeze</strong><span>Export Blend/GLB, manifests, semantic IDs, and hashes before any GT score is visible to the author.</span></div><div><strong>05 · Evaluate</strong><span>Measure pose, native depth, model depth, surface distance, and appearance under one frozen protocol.</span></div></div><p>Ground truth is introduced only after the scenes are frozen. M2 and M3 are evaluated under one global SE(3) alignment with scale fixed to one; no mesh ICP or per-view fitting is used. M1 has no native metric trajectory, so its model-space numbers use one frozen GT-assisted Sim(3) from five manual camera associations and remain diagnostic rather than evidence of metric recovery.</p><aside class="notice"><strong>M1 diagnostic limitation.</strong> Its Sim(3)-aligned geometry and depth numbers describe shape after GT-assisted display alignment, not native metric recovery.</aside></section>
+<section id="more-results"><p class="section-tag">04 / RESULTS AND ANALYSIS</p><h2>Three findings matter more than a leaderboard</h2><article class="finding"><p class="finding-kicker">FINDING 01</p><h3>1. Better pose does not mechanically imply better native depth</h3><p>M3 improves trajectory accuracy over M2—ATE falls from 0.167 m to 0.121 m and rotation error from 0.31° to 0.19°. Yet its native DA3 AbsRel is slightly worse (20.38% versus 19.06%). After modeling, the ordering reverses: M3 reaches 8.37% model-depth AbsRel versus 9.29% for M2. Upstream metrics and downstream scene fidelity are related, but they are not interchangeable.</p></article><figure class="table-figure" id="table-2" data-table-key="table2"><figcaption><span>Table 2.</span> <span class="table-title"></span></figcaption><div class="table-scroll table-mount" role="region" tabindex="0" aria-label="Table 2"></div></figure><figure id="figure-1"><img loading="lazy" src="assets/figure14_abc.png" alt="Trajectory, native depth, and model depth"><figcaption><span>Figure 1.</span> The upstream trajectory and native-depth ranking does not map one-to-one onto frozen-scene depth.</figcaption></figure><article class="finding"><p class="finding-kicker">FINDING 02</p><h3>2. Geometric evidence turns a plausible layout into a more faithful space</h3><p>Under the declared alignment and without mesh ICP, bidirectional surface error falls from 0.376 m for the RGB-only diagnostic baseline to 0.118 m with ViPE, 0.082 m with ORB-SLAM3, and 0.072 m with GT poses. M4 is strongest on geometry in this case, while M3 closes most of the gap using an estimated trajectory.</p></article><figure class="table-figure" id="table-3" data-table-key="table3"><figcaption><span>Table 3.</span> <span class="table-title"></span></figcaption><div class="table-scroll table-mount" role="region" tabindex="0" aria-label="Table 3"></div></figure><article class="finding finding-comparison"><p class="finding-kicker">FINDING 03</p><h3>3. Qualitative comparison</h3></article><div class="analysis-block"><h3>Inspect the reconstructed scenes, not just aggregate scores</h3><p>Use the shared camera to compare each reconstruction with GT. Start with M1 to see how a semantically plausible room can drift in metric layout; compare M2 and M3 to find regions where geometry improves without a matching gain in RGB similarity; then inspect M4 to isolate errors that remain even when camera pose is no longer the bottleneck.</p><figure id="figure-2"><img loading="lazy" src="assets/fixed_five_view_comparison_m1_m4.jpg" alt="M1 to M4 and input GT across five fixed views"><figcaption><span>Figure 2.</span> M1–M4 and input GT across five fixed views.</figcaption></figure><figure class="scene-comparison" id="figure-3" data-state="idle"><h3>Rotate, zoom, and inspect the frozen scenes</h3>
+<div class="scene-toolbar"><label class="method-control">Reconstruction / GT<select id="model-select"><option value="M1">M1</option><option value="M2">M2</option><option value="M3">M3</option><option value="M4" selected>M4</option><option value="GT">GT</option></select></label><label>Display<select id="scene-mode"><option value="compare">Split comparison with GT</option><option value="single">View selected scene alone</option></select></label><button id="scene-reset" type="button">Reset view</button></div>
+<div class="scene-stage"><span class="scene-label scene-label-left">M4</span><span class="scene-label scene-label-right">GT</span><button class="scene-divider" type="button" role="slider" aria-label="Comparison divider" aria-valuemin="2" aria-valuemax="98" aria-valuenow="50"></button><input id="scene-split" class="scene-split" type="range" min="2" max="98" value="50" aria-label="Comparison split"><p class="scene-status">Scroll here to load the 3D comparison.</p></div><button id="scene-retry" class="scene-retry" type="button">Retry</button>
+<figcaption><span>Figure 3.</span> One camera and one full viewport compare the selected frozen model with GT. Ceilings and surrounding walls are hidden at runtime for the same cutaway-style inspection; source files remain unchanged.</figcaption></figure><figure class="fixed-comparison" id="figure-4"><div class="controlrow"><label>Method<select id="compare-method" disabled><option value="M1">M1</option><option value="M2">M2</option><option value="M3">M3</option><option value="M4">M4</option></select></label><label>Fixed view<select id="compare-frame" disabled><option value="0">0</option><option value="36">36</option><option value="72">72</option><option value="108">108</option><option value="144">144</option></select></label></div><div class="compare-pair"><div><img id="compare-pred" src="assets/fixed_views/M1/000.png" alt="Selected frozen model rendering"><p id="compare-label" class="compare-subcaption">M1 / VIEW 0</p></div><div><img id="compare-gt" src="assets/fixed_views/GT/000.png" alt="Input GT RGB"><p class="compare-subcaption">INPUT / GT RGB</p></div></div><figcaption><span>Figure 4.</span> Frozen-model rendering and input GT RGB from exactly the same view.</figcaption></figure><p>The browser hides ceilings and surrounding walls only for cutaway inspection. Runtime registration and display choices do not modify the downloadable frozen GLB or Blend files.</p></div><div class="analysis-block"><h3>What geometric grounding changes—and what it does not</h3><h4>Modeling behaves like a structured—and lossy—regularizer</h4><p>For M2–M4, the final scene depth is substantially closer to GT than native DA3 depth on the shared 180-view protocol: AbsRel falls by roughly 51–60%. One plausible interpretation is that object-level aggregation suppresses inconsistent local predictions. This is a hypothesis, not a controlled causal result: the same abstraction can also erase real surfaces and fine geometry.</p><h4>Measurement constrains generation; it does not solve appearance</h4><p>The strongest geometric route does not dominate PSNR or SSIM. Pose removes one source of uncertainty, but appearance still depends on material estimation, lighting, object detail, and renderer mismatch. Evaluating only RGB would miss spatial progress; evaluating only geometry would miss perceptual failure.</p><h4>The 100-view check measures consistency, not unseen-view generalization</h4><p>Table 7 uses a seeded set of 100 cameras drawn from the 175 modeling frames outside the fixed-five check. These are additional evaluation views relative to that check, but their RGB images were available during modeling. On this set, M4 has the lowest depth AbsRel (7.47%), while M3 has the lowest RMSE (1.090 m).</p><figure id="figure-5"><img loading="lazy" src="assets/model_depth_error_five_views_m1_m4.png" alt="Model-depth error at five fixed evaluation views"><figcaption><span>Figure 5.</span> Model-depth error at the fixed-five check; teal marks missing or invalid predictions.</figcaption></figure><figure class="table-figure" id="table-6" data-table-key="table6"><figcaption><span>Table 6.</span> <span class="table-title"></span></figcaption><div class="table-scroll table-mount" role="region" tabindex="0" aria-label="Table 6"></div></figure><figure class="table-figure" id="table-7" data-table-key="table7"><figcaption><span>Table 7.</span> <span class="table-title"></span></figcaption><div class="table-scroll table-mount" role="region" tabindex="0" aria-label="Table 7"></div></figure></div><div id="office-cafe" class="analysis-block external-case"><h3>Beyond the lobby: what survives in a phone-captured space?</h3><p>Office Café is not a second run of the World Lobby benchmark and is not quantitative validation under the same protocol. It is an external, real-capture case study showing how an object-centric scene, a solved source-camera trajectory, video-to-model comparison, and physics replays can coexist in one inspectable artifact. An author-observed failure—a curved real-world corner simplified into a square one—motivates separating local shape from global scale. A correct scale alone cannot repair that shape error; this is a qualitative observation, not another measured benchmark.</p><p class="office-links"><a href="https://office-cafe-vipe.hiwtishere.chatgpt.site/?lang=en" target="_blank" rel="noopener">Open full-screen model ↗</a><a href="https://office-cafe-vipe.hiwtishere.chatgpt.site/shake?lang=en" target="_blank" rel="noopener">Open shake experiment ↗</a></p><div class="office-external"><h4>Explore the original interactive apps</h4><div class="office-external-grid"><figure class="office-external-card"><a href="https://office-cafe-vipe.hiwtishere.chatgpt.site/?lang=en" target="_blank" rel="noopener noreferrer"><img loading="lazy" src="assets/office-cafe/office-pmodel-poster.jpg" alt="Interactive model: Local preview from the published model-render video."></a><figcaption><strong>Interactive model</strong><p>Local preview from the published model-render video.</p><button type="button" data-office-external="https://office-cafe-vipe.hiwtishere.chatgpt.site/?lang=en" aria-controls="office-external-slot">Load external app · English requested</button></figcaption></figure><figure class="office-external-card"><a href="https://office-cafe-vipe.hiwtishere.chatgpt.site/shake?lang=en" target="_blank" rel="noopener noreferrer"><img loading="lazy" src="assets/office-cafe/office-shake-heavy-poster.jpg" alt="Shake experiment: Local preview from the published strong-shake replay."></a><figcaption><strong>Shake experiment</strong><p>Local preview from the published strong-shake replay.</p><button type="button" data-office-external="https://office-cafe-vipe.hiwtishere.chatgpt.site/shake?lang=en" aria-controls="office-external-slot">Load external app · English requested</button></figcaption></figure></div><p class="office-external-note">Links request English with lang=en; the external host controls language and availability. These local content previews are not screenshots of the external website. If embedding is unavailable, use the full-screen links or the local models and videos below.</p><div id="office-external-slot" class="office-external-slot" hidden></div></div><div class="office-model-grid"><figure class="office-model-card"><div class="office-model-stage" data-src="assets/office-cafe/scene.glb" aria-label="Interactive Office Café ViPE model"></div><figcaption><strong>Current model</strong><span>Video plus estimated depth as input; drag to orbit, right-drag to pan, and scroll to zoom.</span></figcaption></figure><figure class="office-model-card"><div class="office-model-stage" data-src="assets/office-cafe/atlas-scene.glb" aria-label="Interactive legacy Office Café model"></div><figcaption><strong>Legacy model</strong><span>Video-only input; this legacy layout was not recalibrated against the ViPE scan.</span></figcaption></figure></div><p class="office-model-note">Both models open in an elevated cutaway view: upper geometry is clipped only for display to expose the interior. Toggle Cutaway to inspect the complete model; Reset view restores the overview. Authored coordinates, materials, and model files are unchanged. Linked orbit controls compare layouts, not registered geometry.</p><div class="office-video-grid office-video-sync"><figure class="office-video-card"><video data-synced-video muted preload="metadata" playsinline poster="assets/office-cafe/office-input-poster.jpg" src="assets/office-cafe/office-input-web.mp4?v=4915b934b5f9" aria-label="Input video"></video><figcaption><strong>Input video</strong><span>Original phone walkthrough.</span></figcaption></figure><figure class="office-video-card"><video data-synced-video muted preload="metadata" playsinline poster="assets/office-cafe/office-pmodel-poster.jpg" src="assets/office-cafe/office-pmodel-web.mp4?v=3b03fe58a192" aria-label="Procedural model"></video><figcaption><strong>Procedural model</strong><span>Model rendered along the source-camera trajectory.</span></figcaption></figure><figure class="office-video-card"><video data-synced-video muted preload="metadata" playsinline poster="assets/office-cafe/office-scan-poster.jpg" src="assets/office-cafe/office-scan-web.mp4?v=861a074637b4" aria-label="Point cloud"></video><figcaption><strong>Point cloud</strong><span>Scan rendered along the same trajectory.</span></figcaption></figure></div>
+<div class="office-video-controls"><button class="office-video-play" type="button" aria-label="Play all three videos" aria-pressed="false">▶</button><input class="office-video-progress" type="range" min="0" max="1000" value="0" aria-label="Shared video progress"><output class="office-video-time">00:00 / 01:10</output><p class="office-video-status" role="status" aria-live="polite">Shared playback · 70.1 seconds · 30 fps</p></div><div class="office-video-grid"><figure class="office-video-card office-video-card-wide"><video controls preload="metadata" playsinline poster="assets/office-cafe/office-shake-heavy-poster.jpg" src="assets/office-cafe/office-shake-heavy.mp4"></video><figcaption><strong>Strong earthquake effect</strong><span>Every object in the scene is interactable, and cabinet doors can open; this replay shows the physical response under strong shaking.</span></figcaption></figure></div><script type="module" src="office-models.js?v=8752a7cc4f64"></script></div></section>
+<section id="embodied-demo" class="embodied-demo">
+<p class="section-tag">DEMO / MAP-BASED EMBODIED EXECUTION</p><h2>A reconstructed map. A destination. Robots in motion.</h2><p>The demo connects reconstruction to downstream use in NVIDIA Isaac Sim. The reception desk is marked in the navigation map, routes are specified, and a drone, humanoid, quadruped, and wheeled robot execute the task. The same spatial reference connects a destination, route constraints, and motion control: a scene to inspect becomes a map to act with.</p>
+<blockquote class="mission-command"><span>Task intent: </span>Send the robots to the reception desk and have them line up there.</blockquote>
+<ol class="mission-flow" aria-label="Task workflow"><li><span>01</span>Task instruction</li><li><span>02</span>Locate the reception desk</li><li><span>03</span>Set routes on the map</li><li><span>04</span>Execute with controllers</li></ol>
+<figure class="demo-film"><video controls playsinline preload="none" poster="assets/embodied-demo/navigation-214924-poster.png" width="1280" height="720" aria-label="Four-robot map-based navigation demonstration"><source src="assets/embodied-demo/four-robots-214924-156s-phone.mp4" type="video/mp4">Your browser cannot play this video. Use the download link below.</video><figcaption>Map-based navigation in simulation · predefined routes · simulator-pose feedback. The authors confirm that the demonstration has been manually reviewed.</figcaption></figure>
+<div class="demo-links"><a href="assets/embodied-demo/four-robots-214924-156s-phone.mp4" download>Download mobile edition · 11.65 MB · 2:36</a><a href="assets/embodied-demo/four-robots-214924-156s-hd.mp4" download>Download HD edition · 49.69 MB · 2:36</a><a href="data/embodied_demo.json">Run &amp; media record ↗</a></div>
+<figure class="demo-map" id="demo-route-map"><a href="assets/embodied-demo/planned-route-156s.png" target="_blank" rel="noopener"><img loading="lazy" src="assets/embodied-demo/planned-route-156s.png" width="1920" height="1080" alt="Planned routes for four robots, with the reception desk and waypoints marked; static design, not execution trajectories"></a><figcaption><span>Navigation map.</span> The reception desk and planned routes share the reconstructed scene. This is a static design preview, not a recorded trajectory. Open the image to inspect the full-resolution map.</figcaption></figure>
+<details class="demo-protocol"><summary>Demo protocol and scope</summary><p>The instruction above describes the intended task, not a demonstrated language-to-plan parser. Routes were manually specified and screened against both the M4 reconstructed map and original geometry. This is not a reconstruction-only planning benchmark.</p><p>The 156-second presentation edit plays at 1×: Scene (0–6s), Reconstruction (6–12s), Planned route (12–18s), and Navigation (18–156s). Navigation uses the new run <code>20260930-214924</code>; the opening 12.32 seconds of waiting are omitted, and the later terminal hold is outside this cut. The map is a static route design, not a measured execution trace.</p><p>This demo does not evaluate online visual localization, learned instruction understanding, memory retrieval, or real-robot transfer. Those are separate capabilities; the map-based execution shown here remains the intended demonstration.</p><p>Manual presentation review is separate from the historical automated checks retained below; it does not revise their recorded verdicts.</p><p><a href="evidence/embodied-demo-audit.json">Historical automated check record ↗</a></p></details>
+</section>
+<section id="related-work"><p class="section-tag">05 / RELATED WORK</p><h2>From geometric measurement to editable Real2Sim</h2><p>AWSM connects two roles: an agent that constructs an editable scene, and downstream agents that use it as a spatial reference. SLAM and visual-inertial estimation supply pose and metric constraints; ViPE and Depth Anything 3 contribute geometric observations. Concurrent work AHa-3D explores video-driven, tool-based Real2Sim, including camera and 3D-reference estimation from video. Our emphasis is geometric grounding: bringing additional physical measurements, especially IMU-informed visual-inertial pose and metric constraints, into agentic scene construction rather than relying on video-derived evidence alone. The four-route comparison tests reconstruction systems, not the isolated effect of IMU; the demo illustrates a map-based application, not a complete autonomous embodied system.</p><p>Our evaluation therefore spans both sides of the interface: upstream systems are judged as measurement pipelines, while the Blender outputs are judged as persistent spatial artifacts. This differs from evaluating pose, depth, or novel-view synthesis in isolation.</p><span class="inline-sources"><a href="https://research.nvidia.com/labs/toronto-ai/vipe/" target="_blank" rel="noopener">ViPE</a> · <a href="https://arxiv.org/abs/2007.11898" target="_blank" rel="noopener">ORB-SLAM3</a> · <a href="https://arxiv.org/abs/2511.10647" target="_blank" rel="noopener">Depth Anything 3</a> · <a href="https://kevinxu02.github.io/real2sim-indoor-site/" target="_blank" rel="noopener">AHa-3D</a></span><div id="spatial-foundation" class="analysis-block spatial-foundation"><p class="section-tag">RESEARCH DIRECTION</p><h3>Toward persistent spaces for phygital agents</h3><p>The current outputs are editable scene snapshots, and the demo shows a scene-specific simulation integration. The longer-term goal is to make geometry-grounded agentic reconstruction a reusable process for building and maintaining spaces that phygital agents can use. Saving an editable asset is a starting point, not a demonstration of lifelong map maintenance. The directions below extend the present evidence.</p><div class="capability-grid"><div><h4>Navigate</h4><p>Use destinations, free space, and body-specific clearance in a shared geometric frame. Better reconstruction can reduce map mismatch; reliable navigation also needs localization and current observations.</p></div><div><h4>Remember & retrieve</h4><p>Link object identities and locations to source views so language, images, and past observations can refer to the same place. Retrieval and safe arrival should be evaluated separately.</p></div><div><h4>Interact & maintain</h4><p>Edit object properties and relationships, revisit changed areas, and preserve evidence and versions. Persistent memory must distinguish observed geometry, inferred completion, and unknown space.</p></div><div><h4>Simulate & learn</h4><p>Move from the current scene-specific integration to reusable simulation-ready exports with checked scale, collision geometry, and physical parameters. These can support task rehearsal, scenario variation, and training; reduced policy sim-to-real gap remains a separate question to test.</p></div></div><p>A full mesh is not required for every retrieval task: Memory Over Maps uses posed RGB-D keyframes for on-demand localization, while 3D-Mem and task-oriented scene graphs explore complementary memory and representation choices. Our proposed direction is hybrid: preserve visual evidence, refine task-relevant geometry, and build editable simulation assets where they add value. The next question is when an agent has enough evidence to act—and when it should observe again and update its map.</p><p>The immediate agenda is higher reconstruction efficiency and accuracy, reusable simulation-ready scene generation, and deeper integration with phygital agents. Together, these steps connect the reconstruction agent’s modeling and revision loop to the downstream agent’s navigation, memory, and interaction needs.</p><p class="inline-sources"><a href="https://arxiv.org/abs/2603.20530" target="_blank" rel="noopener">Memory Over Maps</a> · <a href="https://arxiv.org/abs/2411.17735" target="_blank" rel="noopener">3D-Mem</a> · <a href="https://arxiv.org/abs/2404.13696" target="_blank" rel="noopener">Clio</a> · <a href="https://arxiv.org/abs/2404.06609" target="_blank" rel="noopener">GOAT-Bench</a></p></div></section>
+<section id="limitations"><p class="section-tag">06 / LIMITATIONS</p><h2>The boundaries are part of the result</h2><ul class="limits-list"><li>One synthetic scene and one engineering run per route do not establish general superiority or statistical significance.</li><li>The four routes are complete systems, not a strict single-variable ablation; M2 versus M3 cannot be attributed to IMU alone.</li><li>M4 uses GT camera poses and is a diagnostic route, not a deployable baseline or a theoretical upper bound.</li><li>M1 model-space metrics rely on GT-assisted Sim(3) and must not be ranked as native metric recovery.</li><li>The 100-view appearance and depth sets contain modeling RGB inputs; they are not a held-out novel-view benchmark.</li><li>Appearance metrics mix geometry, material, illumination, and rendering error. M4 revision 4 has not received a complete independent visual re-review.</li><li>The reception-desk demo is map-based simulation execution, not a controlled comparison of navigation across M1–M4. Routes are predefined and localization uses simulator pose; see the demo protocol for scope.</li><li>Multimodal memory, long-term map maintenance, and reduced policy sim-to-real gap are research goals, not measured outcomes of the four-robot demo. The Office Café corner observation has no matched-view quantitative GT measurement here.</li><li>Historical downstream documentation describes the same M4 Blend hash using a different depth frontend. The frozen tables and assets are retained unchanged; their upstream naming needs reconciliation before attributing historical task outcomes to DA3. See the provenance note.</li></ul><p>Every table is generated from frozen assets. The public manifests expose model hashes, registrations, fixed views, and the exact seeded evaluation indices.</p><div class="asset-downloads"><p><strong>M1</strong> <a href="models/M1/scene.blend" download>scene.blend ↓</a> · <a href="models/M1/scene.glb" download>scene.glb ↓</a> · <a href="evidence/M1/manifest.json">manifest ↗</a></p><p><strong>M2</strong> <a href="models/M2/scene.blend" download>scene.blend ↓</a> · <a href="models/M2/scene.glb" download>scene.glb ↓</a> · <a href="evidence/M2/manifest.json">manifest ↗</a></p><p><strong>M3</strong> <a href="models/M3/scene.blend" download>scene.blend ↓</a> · <a href="models/M3/scene.glb" download>scene.glb ↓</a> · <a href="evidence/M3/manifest.json">manifest ↗</a></p><p><strong>M4</strong> <a href="models/M4/scene.blend" download>scene.blend ↓</a> · <a href="models/M4/scene.glb" download>scene.glb ↓</a> · <a href="evidence/M4/manifest.json">manifest ↗</a></p></div><p class="evidence-links"><a href="data/tables_1_7.json">Tables JSON</a> · <a href="data/scene_comparison.json">Scene manifest</a> · <a href="data/fixed_views.json">Fixed views</a> · <a href="evidence/publication_manifest.json">Publication manifest</a> · <a href="evidence/SHA256SUMS">SHA256</a></p><p><a href="evidence/provenance-note.md">Version and provenance note ↗</a></p></section>
+<section id="citation"><p class="section-tag">07 / CITATION</p><h2>Cite this work</h2><p>If this controlled study, its frozen assets, or its evaluation protocol supports your work, please cite the interactive article below.</p><pre class="citation-block"><code>@misc{awsm_2026,
+  title        = {{AWSM}: Agentic World Simulation and Mapping},
+  author       = {{Phygital AI}},
+  year         = {2026},
+  howpublished = {Interactive research article},
+  url          = {https://phygital-ai.github.io/agentic-world-simulation-and-mapping/},
+  note         = {Reconstruction benchmarks, editable scene assets, and map-based embodied demonstration}
+}</code></pre><a class="citation-download" href="data/awsm.bib" download="awsm.bib">Download BibTeX <span aria-hidden="true">↓</span></a></section>
+<section id="references"><p class="section-tag">08 / REFERENCES</p><h2>References</h2><ol class="references-list"><li><a href="https://kevinxu02.github.io/real2sim-indoor-site/" target="_blank" rel="noopener">AHa-3D: Agentic Tool Use for Real2Sim with GPT-6 Astra.</a></li><li><a href="https://arxiv.org/abs/2508.10934" target="_blank" rel="noopener">ViPE: Video Pose Engine for camera motion and near-metric geometry.</a></li><li><a href="https://arxiv.org/abs/2007.11898" target="_blank" rel="noopener">Campos et al. ORB-SLAM3: An Accurate Open-Source Library for Visual, Visual–Inertial, and Multimap SLAM. IEEE Transactions on Robotics, 2021.</a></li><li><a href="https://arxiv.org/abs/2511.10647" target="_blank" rel="noopener">Depth Anything 3: Recovering the Visual Space from Any Views.</a></li><li><a href="https://doi.org/10.1109/TIP.2003.819861" target="_blank" rel="noopener">Wang et al. Image Quality Assessment: From Error Visibility to Structural Similarity. IEEE Transactions on Image Processing, 2004.</a></li><li><a href="https://arxiv.org/abs/1801.03924" target="_blank" rel="noopener">Zhang et al. The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. CVPR, 2018.</a></li><li><a href="https://www.blender.org/" target="_blank" rel="noopener">Blender Foundation. Blender: Free and Open Source 3D Creation Suite.</a></li><li><a href="https://arxiv.org/abs/1708.03852" target="_blank" rel="noopener">Qin et al. VINS-Mono: A Robust and Versatile Monocular Visual-Inertial State Estimator. TRO, 2018.</a></li><li><a href="https://arxiv.org/abs/2509.13414" target="_blank" rel="noopener">Keetha et al. MapAnything: Universal Feed-Forward Metric 3D Reconstruction. 2025.</a></li><li><a href="https://arxiv.org/abs/2603.20530" target="_blank" rel="noopener">Zhou et al. Memory Over Maps: 3D Object Localization Without Reconstruction. 2026.</a></li><li><a href="https://arxiv.org/abs/2411.17735" target="_blank" rel="noopener">Yang et al. 3D-Mem: 3D Scene Memory for Embodied Exploration and Reasoning. 2024.</a></li><li><a href="https://arxiv.org/abs/2404.06609" target="_blank" rel="noopener">Khanna et al. GOAT-Bench: A Benchmark for Multi-Modal Lifelong Navigation. CVPR, 2024.</a></li><li><a href="https://arxiv.org/abs/2309.16650" target="_blank" rel="noopener">ConceptGraphs: Open-Vocabulary 3D Scene Graphs for Perception and Planning. ICRA, 2024.</a></li><li><a href="https://arxiv.org/abs/2404.13696" target="_blank" rel="noopener">Clio: Real-time Task-Driven Open-Set 3D Scene Graphs. RA-L, 2024.</a></li></ol></section><noscript>JavaScript is required for tables and interactive figures.</noscript></main></div><footer>Phygital AI · AWSM · frozen evidence, editable outputs · <a href="zh.html">中文</a></footer></body></html>

evidence/M1/manifest.json

@@ -0,0 +1,68 @@
+{
+  "schema_version": 2,
+  "method": "M1",
+  "files": {
+    "glb": {
+      "source": "astra_blender/models/M1/scene.glb",
+      "published": "models/M1/scene.glb",
+      "sha256": "21a605be65d58dbe01606b8326bfe6f07fa418c67e133c9b2181729b68bc2832",
+      "bytes": 15104860
+    },
+    "blend": {
+      "source": "astra_blender/models/M1/scene.blend",
+      "published": "models/M1/scene.blend",
+      "sha256": "80f985ad09591bdb227e99e71f6bacaa0b9e8c8b565c27da64ad4ac4ceb8910e",
+      "bytes": 2691023
+    }
+  },
+  "table_rows": {
+    "table1": {
+      "method": "M1",
+      "allowed_inputs": "180 张采样 RGB",
+      "geometry_source": "视觉推断，无测量尺度",
+      "modeling_role": "推断布局，编写对象与材质"
+    },
+    "table2": {
+      "method": "M1",
+      "ate_m": null,
+      "rotation_deg": null,
+      "native_depth_absrel": null,
+      "model_depth_absrel": 0.16439823806285858,
+      "alignment": "GT-assisted Sim(3), diagnostic only"
+    },
+    "table3": {
+      "method": "M1",
+      "model_to_gt_mean_m": 0.34486243741226075,
+      "observed_gt_to_model_mean_m": 0.40664686199496186,
+      "model_sha256": "80f985ad09591bdb227e99e71f6bacaa0b9e8c8b565c27da64ad4ac4ceb8910e",
+      "alignment": "GT-assisted Sim3 from frozen manual-view camera associations; shape/display diagnostic only; not metric recovery; no mesh ICP",
+      "bidirectional_mean_m": 0.3757546497036113
+    },
+    "table4": {
+      "method": "M1",
+      "psnr_db": 10.260909193969148,
+      "ssim": 0.3091735512018204,
+      "lpips_alex_v01": 0.6575803160667419
+    },
+    "table5": {
+      "method": "M1",
+      "psnr_db": 11.691789500467046,
+      "ssim": 0.31951158225536347,
+      "lpips_alex_v01": 0.6073989069461823
+    },
+    "table6": {
+      "method": "M1",
+      "absrel": 0.2901068925857544,
+      "rmse_m": 3.049535036087036,
+      "coverage": 0.9996145833333333,
+      "penalized_mae_m": 1.88661527633667
+    },
+    "table7": {
+      "method": "M1",
+      "absrel": 0.164479598402977,
+      "rmse_m": 1.7184345722198486,
+      "coverage": 0.9924354166666667,
+      "penalized_mae_m": 1.2859067916870117
+    }
+  }
+}

evidence/M2/manifest.json

@@ -0,0 +1,68 @@
+{
+  "schema_version": 2,
+  "method": "M2",
+  "files": {
+    "glb": {
+      "source": "astra_blender/models/M2/scene.glb",
+      "published": "models/M2/scene.glb",
+      "sha256": "0ee9123f38099d1df63ab593dd7082d5f1cf1b136e63833ec5afa0e072cf4c2e",
+      "bytes": 14523124
+    },
+    "blend": {
+      "source": "astra_blender/models/M2/scene.blend",
+      "published": "models/M2/scene.blend",
+      "sha256": "c19f33ec434bbcf89137e67d81c283b9adfd48f42e993d8b29d63e08318f79a6",
+      "bytes": 3886116
+    }
+  },
+  "table_rows": {
+    "table1": {
+      "method": "M2",
+      "allowed_inputs": "完整 RGB 视频；选取 180 帧建模",
+      "geometry_source": "ViPE RGB-only 位姿 → pose-conditioned DA3 深度",
+      "modeling_role": "将位姿与深度测量组织为对象化场景"
+    },
+    "table2": {
+      "method": "M2",
+      "ate_m": 0.1668,
+      "rotation_deg": 0.3065,
+      "native_depth_absrel": 0.19056472003176061,
+      "model_depth_absrel": 0.09286404401063919,
+      "alignment": "one global SE(3), scale=1"
+    },
+    "table3": {
+      "method": "M2",
+      "model_to_gt_mean_m": 0.13236121125603983,
+      "observed_gt_to_model_mean_m": 0.10298419313248852,
+      "model_sha256": "c19f33ec434bbcf89137e67d81c283b9adfd48f42e993d8b29d63e08318f79a6",
+      "alignment": "one SE3 from 180 input camera correspondences; no scale fit",
+      "bidirectional_mean_m": 0.11767270219426418
+    },
+    "table4": {
+      "method": "M2",
+      "psnr_db": 12.95293150317823,
+      "ssim": 0.3896766543388367,
+      "lpips_alex_v01": 0.5310081720352173
+    },
+    "table5": {
+      "method": "M2",
+      "psnr_db": 13.207560508419194,
+      "ssim": 0.3776748192310333,
+      "lpips_alex_v01": 0.5329300162196159
+    },
+    "table6": {
+      "method": "M2",
+      "absrel": 0.09218113869428635,
+      "rmse_m": 1.116154432296753,
+      "coverage": 0.9978125,
+      "penalized_mae_m": 0.577746570110321
+    },
+    "table7": {
+      "method": "M2",
+      "absrel": 0.09284797310829163,
+      "rmse_m": 1.1448776721954346,
+      "coverage": 0.9977067708333334,
+      "penalized_mae_m": 0.6155765056610107
+    }
+  }
+}

evidence/M3/manifest.json

@@ -0,0 +1,68 @@
+{
+  "schema_version": 2,
+  "method": "M3",
+  "files": {
+    "glb": {
+      "source": "astra_blender/models/M3/scene.glb",
+      "published": "models/M3/scene.glb",
+      "sha256": "22699f8225a5f90166cdea2e6e2bdad7977fda4a1fe1086da209f0c9afe97a6c",
+      "bytes": 7701488
+    },
+    "blend": {
+      "source": "astra_blender/models/M3/scene.blend",
+      "published": "models/M3/scene.blend",
+      "sha256": "00cd43eef3b4a6b2b2b169e8a7eda787dc57222dc501a3a9be09f39af5f46a9b",
+      "bytes": 2319452
+    }
+  },
+  "table_rows": {
+    "table1": {
+      "method": "M3",
+      "allowed_inputs": "视频、IMU、相机–IMU 标定",
+      "geometry_source": "ORB-SLAM3 单目惯性位姿 → pose-conditioned DA3 深度",
+      "modeling_role": "结合图像、位姿、深度建模"
+    },
+    "table2": {
+      "method": "M3",
+      "ate_m": 0.1205,
+      "rotation_deg": 0.1943,
+      "native_depth_absrel": 0.20379653842953233,
+      "model_depth_absrel": 0.08371810615062714,
+      "alignment": "one global SE(3), scale=1"
+    },
+    "table3": {
+      "method": "M3",
+      "model_to_gt_mean_m": 0.06935983382743138,
+      "observed_gt_to_model_mean_m": 0.09502824761255278,
+      "model_sha256": "00cd43eef3b4a6b2b2b169e8a7eda787dc57222dc501a3a9be09f39af5f46a9b",
+      "alignment": "one SE3 from 180 input camera correspondences; no scale fit",
+      "bidirectional_mean_m": 0.08219404071999209
+    },
+    "table4": {
+      "method": "M3",
+      "psnr_db": 12.052547333787004,
+      "ssim": 0.3958937466144562,
+      "lpips_alex_v01": 0.5791207671165466
+    },
+    "table5": {
+      "method": "M3",
+      "psnr_db": 12.252687193116532,
+      "ssim": 0.3972503730654717,
+      "lpips_alex_v01": 0.5683315402269363
+    },
+    "table6": {
+      "method": "M3",
+      "absrel": 0.09866034984588623,
+      "rmse_m": 1.1816775798797607,
+      "coverage": 0.9984166666666666,
+      "penalized_mae_m": 0.5551674365997314
+    },
+    "table7": {
+      "method": "M3",
+      "absrel": 0.08247262239456177,
+      "rmse_m": 1.0896246433258057,
+      "coverage": 0.9991479166666667,
+      "penalized_mae_m": 0.4945947229862213
+    }
+  }
+}

evidence/M4/manifest.json

@@ -0,0 +1,68 @@
+{
+  "schema_version": 2,
+  "method": "M4",
+  "files": {
+    "glb": {
+      "source": "astra_blender/models/M4/scene.glb",
+      "published": "models/M4/scene.glb",
+      "sha256": "716e4b89b7b11ad8a0cb0e238b366b4f5dc80e3a2c3cddb5eedc469ef4e63774",
+      "bytes": 17202608
+    },
+    "blend": {
+      "source": "astra_blender/models/M4/scene.blend",
+      "published": "models/M4/scene.blend",
+      "sha256": "cc6cb605246a255d94e36b9dc3f8b91d5a292b047470bc0ff75b8a5c404b0cfe",
+      "bytes": 2268620
+    }
+  },
+  "table_rows": {
+    "table1": {
+      "method": "M4",
+      "allowed_inputs": "视频及 GT camera pose",
+      "geometry_source": "GT 位姿 → pose-conditioned DA3 深度",
+      "modeling_role": "与 M3 相同类型的建模流程"
+    },
+    "table2": {
+      "method": "M4",
+      "ate_m": "-",
+      "rotation_deg": "-",
+      "native_depth_absrel": 0.1886706026379072,
+      "model_depth_absrel": 0.07600006461143494,
+      "alignment": "GT coordinates"
+    },
+    "table3": {
+      "method": "M4",
+      "model_to_gt_mean_m": 0.06866006729939131,
+      "observed_gt_to_model_mean_m": 0.07449522976921306,
+      "model_sha256": "cc6cb605246a255d94e36b9dc3f8b91d5a292b047470bc0ff75b8a5c404b0cfe",
+      "alignment": "permitted GT input coordinate frame; astra_blender/models/M4 T_input_model is identity",
+      "bidirectional_mean_m": 0.07157764853430218
+    },
+    "table4": {
+      "method": "M4",
+      "psnr_db": 12.294263798590851,
+      "ssim": 0.34136943221092225,
+      "lpips_alex_v01": 0.531274390220642
+    },
+    "table5": {
+      "method": "M4",
+      "psnr_db": 12.513978931521383,
+      "ssim": 0.33674563601613045,
+      "lpips_alex_v01": 0.5237806642055511
+    },
+    "table6": {
+      "method": "M4",
+      "absrel": 0.08832013607025146,
+      "rmse_m": 1.1883933544158936,
+      "coverage": 0.9985833333333334,
+      "penalized_mae_m": 0.5010335445404053
+    },
+    "table7": {
+      "method": "M4",
+      "absrel": 0.07466284930706024,
+      "rmse_m": 1.1474982500076294,
+      "coverage": 0.998678125,
+      "penalized_mae_m": 0.47199496626853943
+    }
+  }
+}

evidence/embodied-demo-audit.json

@@ -0,0 +1,20 @@
+{
+  "schema_version": 1,
+  "run_id": "sceneweft-walljourney-20260930-214924",
+  "source": "wall-journey-audit.json from the recorded run; summarized without changing the verdict",
+  "status": "FAILED",
+  "passed": false,
+  "ground_final_formation_reached": true,
+  "full_flight_mission_completed": true,
+  "end_formation_hold_s": 7.994999821297824,
+  "passed_checks": ["computed_current_body_stream_map_clearance", "computed_current_body_stream_inter_robot_clearance", "measured_ground_path_length_and_waypoint_traversal", "fresh_current_run_closure_receipt"],
+  "failed_checks": [
+    {
+      "check": "measured_terminal_health_and_formation_hold",
+      "required_hold_seconds": 8.0,
+      "measured_hold_seconds": 7.994999821297824,
+      "interpretation": "Measured hold falls short of the unchanged 8.0-second criterion. The later terminal hold is outside the 156-second presentation edit."
+    }
+  ],
+  "publication_role": "execution illustration, not an accepted navigation benchmark"
+}

evidence/embodied-demo-delivery.json

@@ -0,0 +1,68 @@
+{
+  "schema_version": 1,
+  "delivery": "20260930-four-robot-demo-final-156s",
+  "run_id": "sceneweft-walljourney-20260930-214924",
+  "exported_at": "2026-09-30T23:37:26.372334+08:00",
+  "media_copied_without_reencoding": true,
+  "original_records_retained_in_local_delivery": true,
+  "independent_film_accepted": false,
+  "source_records": [
+    {
+      "path": "README.md",
+      "sha256": "3eeb1197ae05a623f5a4f0241cd118a961534acd357cca215ab0a562a475a3cb",
+      "bytes": 2070
+    },
+    {
+      "path": "SHA256SUMS",
+      "sha256": "0cfd060f13facd03d376fae3b48da26350a511283868e56f8b13b1708fc54465",
+      "bytes": 235373
+    },
+    {
+      "path": "05_records/final-export-receipt.json",
+      "sha256": "325a667714c568d7ed7c252e9b72839eb143e1487f13168fe4f932917f7c74f9",
+      "bytes": 43265
+    },
+    {
+      "path": "05_records/wall-journey-audit.json",
+      "sha256": "5b15f5af8ebf031337f5c514af3054acf03a25176f1312a07b7defbf056b2755",
+      "bytes": 22408
+    },
+    {
+      "path": "05_records/flight-report.json",
+      "sha256": "f8324dd461588b90fb9e128d58f83b6eef9e18a696debaaca6b63822ca96a39b",
+      "bytes": 11779
+    },
+    {
+      "path": "05_records/navigation-completed-status.json",
+      "sha256": "8141b0142cebdfc39c6b07fff44ea813bd3bfcb5ade4a1e7d72c0950e3ad0da8",
+      "bytes": 20662
+    },
+    {
+      "path": "05_records/segment-validation.json",
+      "sha256": "d3719eed6bb02c70385c50b2a836f17e211a7e4f9a52a5ce3fea59dbd338251b",
+      "bytes": 814
+    }
+  ],
+  "asset_bindings": [
+    {
+      "source": "00_final/sceneweft-four-robots-final-156s-phone.mp4",
+      "published": "assets/embodied-demo/four-robots-214924-156s-phone.mp4",
+      "sha256": "b9008c7b754f994d30a740a632b70ea0f693f3917fcf11820e64902ddbd13596"
+    },
+    {
+      "source": "00_final/sceneweft-four-robots-final-156s.mp4",
+      "published": "assets/embodied-demo/four-robots-214924-156s-hd.mp4",
+      "sha256": "f83dbb49e3aa97fd3f5f9c54ed79ed8624bac75dbd7e4cdddc4dc414b76ed222"
+    },
+    {
+      "source": "03_planned_route/planned-route-original.png",
+      "published": "assets/embodied-demo/planned-route-156s.png",
+      "sha256": "e3c8e766f40a975a0aceca185fa60867375153fc8ddb6db451788eb935cb9b8b"
+    },
+    {
+      "source": "04_navigation/middle.png",
+      "published": "assets/embodied-demo/navigation-214924-poster.png",
+      "sha256": "fe1dfcc8d014fc47f93a6e4e343598c146373c2d3dc7b41613123bb6caafec8f"
+    }
+  ]
+}

evidence/provenance-note.md

@@ -0,0 +1,41 @@
+# Publication scope and provenance — 30 September 2026
+
+## This revision
+
+The Phygital AI edition changes the bilingual narrative and adds the reception-desk embodied demo. It retains the seven frozen experiment tables, six original figures, model files, registration transforms, and fixed-view images from upstream commit `3c5f27ca16fffca03f9e071f70bcc234e34739ee` of `wentingw/agentic-world-blog`.
+
+The demo expresses map-based navigation: the destination is marked on the reconstructed map, routes are specified, and robot controllers execute them. It is not presented as an online visual-localization experiment. The task wording describes the intent, not a demonstrated language-to-plan parser.
+
+The demo now uses the supplied `20260930-four-robot-demo-final-156s` delivery, copied without re-encoding. The mobile edition is 11,648,571 bytes (SHA-256 `b9008c7b754f994d30a740a632b70ea0f693f3917fcf11820e64902ddbd13596`); the HD edition is 49,687,378 bytes (SHA-256 `f83dbb49e3aa97fd3f5f9c54ed79ed8624bac75dbd7e4cdddc4dc414b76ed222`). Both are 156 seconds, 1280×720, 30 fps, at 1× playback speed. The package's original planned-route PNG and navigation screenshot are copied unchanged as the standalone map and video poster.
+
+The first 18 seconds present Scene, Reconstruction, and Planned route; the remaining 138 seconds show native RGB execution from run `sceneweft-walljourney-20260930-214924`. The package omits 12.32 seconds of opening waiting and truncates later footage, including terminal formation hold. The static map is a planned-route design, not a measured execution trace. These are presentation edits, not additional experiments.
+
+The new source run reports completed flight and formation. Unlike the superseded `144825` run, its independent audit passes map and inter-robot clearance, waypoint traversal, and closure checks. The total verdict remains FAILED because measured terminal health/formation hold is 7.994999821297824 seconds against an unchanged 8.0-second requirement. Media export checks do not establish physical or film acceptance. The old run's clearance, waypoint, and flight-closure failures are not attributed to this new run. Current conditions and audit details are recorded in `data/embodied_demo.json` and `evidence/embodied-demo-audit.json`; source-receipt hashes are in `evidence/embodied-demo-delivery.json`. Full production records and raw frames remain in the local delivery, rather than uploading the 2.61 GB package to the blog.
+
+## Display and positioning update — 1 October 2026
+
+The masthead carries the publication date 1 October 2026. This is the article's publication date, not a claim about the earliest public disclosure of the method.
+
+Figure 3 now opens at a higher shared camera angle, with aspect-aware framing. Mirror-wall shells are no longer exempted from its existing runtime wall cutaway. Office Café opens in an elevated section view; its Cutaway button restores or disables the display-only horizontal clipping plane, and Reset view restores the overview. Linked Office controls are a viewing aid, not a geometric registration. Model bytes, evaluation values, registration transforms, and demo media are unchanged.
+
+AHa-3D is described as concurrent work rather than an antecedent to AWSM. Its primary project sections `sections/workflow.html` and `sections/reference.html` at `https://kevinxu02.github.io/real2sim-indoor-site/` (checked 1 October 2026) describe video input, camera/depth estimation with Pi3X, segmentation with SAM3, and a TSDF reference. Thus, “video-driven” does not mean “without estimated geometry.” AWSM's emphasized distinction is additional physical evidence, especially IMU-informed visual-inertial constraints. No dated Twitter source has been attached here, so this update makes no public-priority claim. The frozen four-route experiment is not an isolated IMU ablation.
+
+## Upstream naming discrepancy
+
+The frozen M4 Blend SHA-256 is `cc6cb605246a255d94e36b9dc3f8b91d5a292b047470bc0ff75b8a5c404b0cfe`.
+
+- The current upstream table describes M4 as GT camera poses with pose-conditioned DA3 depth.
+- Historical downstream documentation dated 24 September 2026 describes that same Blend hash as GT camera poses with pose-conditioned MapAnything depth.
+- Historical M3 documentation also names OpenVINS, whereas the current frozen table names ORB-SLAM3.
+
+This may reflect documentation or pipeline-version drift. The asset hash establishes identical output, not which upstream description is correct. This publication does not silently change the frozen data or infer a causal result from that discrepancy. Resolving it requires the original input packets and modeling-run records. Historical retrieval/navigation outcomes are not combined with this demo or attributed to DA3.
+
+## Scope
+
+The reconstruction measurements, map-based execution demonstration, and longer-term spatial-memory/simulation research agenda are separate evidence levels. More faithful geometry can reduce geometric mismatch; dependable real-world navigation and policy transfer also depend on localization, physical properties, dynamics, sensing, and change handling. No real-robot transfer or reduction in training sim-to-real gap is measured here.
+
+## 中文说明
+
+本轮仅调整叙事并加入前台导航 demo，保留上游七张实验表、六张原图、模型和配准。Demo 表达“目标在重建地图中标出—路线在地图上设定—机器人执行”，并非在线视觉定位或语言自动规划实验。
+
+新旧文档对同一 M4 模型的深度前端，以及 M3 的位姿前端，存在命名差异；需要用原始输入包和运行记录核对。本页不更改冻结结果，也不把旧检索任务收益直接归因于新版方法名称。多模态记忆、长期维护及降低 sim-to-real 差距作为研究方向呈现。

evidence/publication_manifest.json

@@ -0,0 +1,456 @@
+{
+  "schema_version": 2,
+  "source_contract": "data/source_contract.json",
+  "files": [
+    {
+      "path": ".nojekyll",
+      "sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+      "bytes": 0
+    },
+    {
+      "path": "app.js",
+      "sha256": "d70db3a6816b9e91abbd6f9711fbcf006113e913fa0800334ccb4af986c83012",
+      "bytes": 7499
+    },
+    {
+      "path": "assets/awsm-promo-v5-poster.jpg",
+      "sha256": "3c9d54318ae7cc4fd580cfd97569933645f49e7348a9052eec5c33c9856840c3",
+      "bytes": 130204
+    },
+    {
+      "path": "assets/awsm-promo-v5.mp4",
+      "sha256": "41d40c2cabcd34d3e8760b927e6f3ae05dce21237b6ba6d25205103d0e055f2f",
+      "bytes": 39621539
+    },
+    {
+      "path": "assets/comparison/GT.glb",
+      "sha256": "6c28ab067c63c5b1fce19c6a972f3d66c7ac4d9bd9ce3b5a57e2e18f5641fe8f",
+      "bytes": 24816108
+    },
+    {
+      "path": "assets/embodied-demo/four-robots-214924-156s-hd.mp4",
+      "sha256": "f83dbb49e3aa97fd3f5f9c54ed79ed8624bac75dbd7e4cdddc4dc414b76ed222",
+      "bytes": 49687378
+    },
+    {
+      "path": "assets/embodied-demo/four-robots-214924-156s-phone.mp4",
+      "sha256": "b9008c7b754f994d30a740a632b70ea0f693f3917fcf11820e64902ddbd13596",
+      "bytes": 11648571
+    },
+    {
+      "path": "assets/embodied-demo/navigation-214924-poster.png",
+      "sha256": "fe1dfcc8d014fc47f93a6e4e343598c146373c2d3dc7b41613123bb6caafec8f",
+      "bytes": 1227190
+    },
+    {
+      "path": "assets/embodied-demo/planned-route-156s.png",
+      "sha256": "e3c8e766f40a975a0aceca185fa60867375153fc8ddb6db451788eb935cb9b8b",
+      "bytes": 1125683
+    },
+    {
+      "path": "assets/figure14_abc.png",
+      "sha256": "6ce74542f6ce9d45a985e1916a42ef2e1205b29bbcb03959e4b1ea08380dfa39",
+      "bytes": 344653
+    },
+    {
+      "path": "assets/fixed_five_view_comparison_m1_m4.jpg",
+      "sha256": "501a674252a492deb209a9cfd847e8f0c83103c49d5f250964ee451d233481a9",
+      "bytes": 1755470
+    },
+    {
+      "path": "assets/fixed_views/GT/000.png",
+      "sha256": "0c95e2288b7d864133952d0ce6f552abea743a7d4ea4622103047bd794a07d04",
+      "bytes": 3688630
+    },
+    {
+      "path": "assets/fixed_views/GT/036.png",
+      "sha256": "449c88399a68cd88397306fbe42c5293e6684d9a9cfb749ea9e3f5ff78eef04d",
+      "bytes": 3688630
+    },
+    {
+      "path": "assets/fixed_views/GT/072.png",
+      "sha256": "222246adc897e2c42232a2b10231c614601dcd71580186647522ff1881a9b71f",
+      "bytes": 3688630
+    },
+    {
+      "path": "assets/fixed_views/GT/108.png",
+      "sha256": "5e5d07ef9ee0c166c24160b228dfc802ab244ac6002e8cfa76e450ee4a385565",
+      "bytes": 3688630
+    },
+    {
+      "path": "assets/fixed_views/GT/144.png",
+      "sha256": "cb8c1a4ade400f935ef4e1e3bfe93245fe77d8beb239040bf43baea840996444",
+      "bytes": 3688630
+    },
+    {
+      "path": "assets/fixed_views/M1/000.png",
+      "sha256": "33b9f0d09f34871564d86bde23d61d82d48133b3f9ac2f6c251b8435f4bce5c4",
+      "bytes": 460982
+    },
+    {
+      "path": "assets/fixed_views/M1/036.png",
+      "sha256": "d1ccc4bf1da6c2b9e90bfc38bd37d8598a0a0068b374d05a4138bac2b37a904c",
+      "bytes": 490712
+    },
+    {
+      "path": "assets/fixed_views/M1/072.png",
+      "sha256": "30b78883e701ef411d4cc04a44dffaba6df13631f6ab4c643edc42e2aa4b2415",
+      "bytes": 431098
+    },
+    {
+      "path": "assets/fixed_views/M1/108.png",
+      "sha256": "9c98270c0e013225d168ee7714afbd2a80f5dfa764c0b158437d1985c26f4745",
+      "bytes": 309546
+    },
+    {
+      "path": "assets/fixed_views/M1/144.png",
+      "sha256": "cf7e5355a33a83fd148daf6a648936757e8be6d09639ace537cdb2ed37a8ed98",
+      "bytes": 486611
+    },
+    {
+      "path": "assets/fixed_views/M2/000.png",
+      "sha256": "dff45b5cd9bc4c5d757f36c8b8eaff015684c6681b40ad0d6b936940e7c9f316",
+      "bytes": 465709
+    },
+    {
+      "path": "assets/fixed_views/M2/036.png",
+      "sha256": "2f8cb5397a7800f0229867ba94b0ee8e8c453600176531da65f27f96e7f4f147",
+      "bytes": 478265
+    },
+    {
+      "path": "assets/fixed_views/M2/072.png",
+      "sha256": "899b65b8c0487f9987f38d09b246c9f97b0b94217d11ebbe29b990f48f19b003",
+      "bytes": 424182
+    },
+    {
+      "path": "assets/fixed_views/M2/108.png",
+      "sha256": "291733823717146fa1e474187aafdca49d6e06b8900985dde6c48fef55684926",
+      "bytes": 464925
+    },
+    {
+      "path": "assets/fixed_views/M2/144.png",
+      "sha256": "6c4f51435f109f71a35a09c0757e8995dd4de3ebec16c43ee32c25c94303e165",
+      "bytes": 481190
+    },
+    {
+      "path": "assets/fixed_views/M3/000.png",
+      "sha256": "a2ca215a65c82b7361c072694c6a1201464bc2ad80c18ba7c59a963927d89b44",
+      "bytes": 418136
+    },
+    {
+      "path": "assets/fixed_views/M3/036.png",
+      "sha256": "6c117fef7d146ae81baab7453cb88a2df7f75b080d8e9c672e2163a9fbf00f79",
+      "bytes": 433089
+    },
+    {
+      "path": "assets/fixed_views/M3/072.png",
+      "sha256": "7d40497cf6c2c3c20403a4c3e1f9b0abe478fbb8602e1f259e51c48dd535d4f3",
+      "bytes": 390554
+    },
+    {
+      "path": "assets/fixed_views/M3/108.png",
+      "sha256": "12b289c7af68a32b1a00817685ed79ede039b3624af9896a57e90b1d5b99d864",
+      "bytes": 433108
+    },
+    {
+      "path": "assets/fixed_views/M3/144.png",
+      "sha256": "cc546f9d355fd3d8ac4f48a08b4e4a1158c112585922bb34e8715229482a3542",
+      "bytes": 474768
+    },
+    {
+      "path": "assets/fixed_views/M4/000.png",
+      "sha256": "a4e82854e7ce3e3870959f008a6d590f9c65925e4b33ca0e96f6cf7ffdd6ab43",
+      "bytes": 536558
+    },
+    {
+      "path": "assets/fixed_views/M4/036.png",
+      "sha256": "a0a7a8afcd3ebc4e56c79a45aa41740f28ca08186b00a67af19161f23f773c2d",
+      "bytes": 547041
+    },
+    {
+      "path": "assets/fixed_views/M4/072.png",
+      "sha256": "a934d91a6f6b9e130322f526e212dc9f279e1c787398e0181da3e9dd4f60127c",
+      "bytes": 491159
+    },
+    {
+      "path": "assets/fixed_views/M4/108.png",
+      "sha256": "4046e5e362ab8a81404aabd59da159a192f206bf7a80b694aca130c698fe2c7e",
+      "bytes": 531759
+    },
+    {
+      "path": "assets/fixed_views/M4/144.png",
+      "sha256": "1d6080b2ce9f2ae95c6a6d9fc47a620365a4439df89267ab166122287d8b61be",
+      "bytes": 529379
+    },
+    {
+      "path": "assets/icons/LICENSE.octicons.txt",
+      "sha256": "da259c8bd0de62713ccdcf88910aebca810644f98c2c912bad814fc79ea778df",
+      "bytes": 1068
+    },
+    {
+      "path": "assets/icons/README.md",
+      "sha256": "7460123d09d83ab450e8b4012f1f2af3ff87daf6f63ac84426faacbb9e28a342",
+      "bytes": 460
+    },
+    {
+      "path": "assets/icons/github.svg",
+      "sha256": "7421820090b50ac79d7c2bf7c951a433d7098a8a9c474809207bdd45f62d9d46",
+      "bytes": 822
+    },
+    {
+      "path": "assets/icons/huggingface.svg",
+      "sha256": "9cb9d71c1513fd2c57181cede84cbf552fbd8da3e3e32da514b30412a37c93a6",
+      "bytes": 4654
+    },
+    {
+      "path": "assets/model_depth_error_five_views_m1_m4.png",
+      "sha256": "448bc111da467461ae84941bc6ea5f5601ca5f9746d292dc1b8b279bbfa153ce",
+      "bytes": 469515
+    },
+    {
+      "path": "assets/office-cafe/atlas-scene.glb",
+      "sha256": "2f966df63078306062d0d68e3246d19a8f1179857a2fcc936835b33de1fcf87b",
+      "bytes": 10011388
+    },
+    {
+      "path": "assets/office-cafe/office-input-poster.jpg",
+      "sha256": "09ee086ae0eca08afecbafb5fe5849600d0837349ae1b14e72dbdbaae9c2f237",
+      "bytes": 94250
+    },
+    {
+      "path": "assets/office-cafe/office-input-web.mp4",
+      "sha256": "4915b934b5f9971c77d60f88d91ed110c5e164bb0932f89ac7cf8915964e9d2c",
+      "bytes": 27902821
+    },
+    {
+      "path": "assets/office-cafe/office-pmodel-poster.jpg",
+      "sha256": "61ea36de0e40a955879b3b0195961ad137c738373726efda760b50232b7041b8",
+      "bytes": 80038
+    },
+    {
+      "path": "assets/office-cafe/office-pmodel-web.mp4",
+      "sha256": "3b03fe58a1922cdd1ca935f6ad52c69e672f1e391cdec1543e60db99e93a4d71",
+      "bytes": 21650736
+    },
+    {
+      "path": "assets/office-cafe/office-scan-poster.jpg",
+      "sha256": "59f64e13f254ef3c9dbf511277cd9782ac0267f25b54e99cba938269be30d4e3",
+      "bytes": 218866
+    },
+    {
+      "path": "assets/office-cafe/office-scan-web.mp4",
+      "sha256": "861a074637b4b756f1c6ca115b50f20c9d03b46303caa6eb0503e812c4194b92",
+      "bytes": 35449707
+    },
+    {
+      "path": "assets/office-cafe/office-scan2model.mp4",
+      "sha256": "11a2db99879d1e9e4b681f5ce3f8e2bb68ea2d66116013c17036b1b0a0231b7a",
+      "bytes": 36167269
+    },
+    {
+      "path": "assets/office-cafe/office-shake-heavy-poster.jpg",
+      "sha256": "e4574f22938288bf76f5d123cbba7f871643a4889c519fab176484738438be6b",
+      "bytes": 104096
+    },
+    {
+      "path": "assets/office-cafe/office-shake-heavy.mp4",
+      "sha256": "7a5a3e79a0060c8658703208daf0c1a4c76c039f1165a9841193dc3ca64e517b",
+      "bytes": 3229810
+    },
+    {
+      "path": "assets/office-cafe/office-video2model.mp4",
+      "sha256": "4f0f2d4f243bf97e1ace0c874a111d7f743223109a53e84a47d036ccd0e51f2e",
+      "bytes": 10769208
+    },
+    {
+      "path": "assets/office-cafe/scene.glb",
+      "sha256": "cc2e2afcfd9426335f83cd6152f90d582f3a0e15dd4dfe5cf58d85e65c5b72ac",
+      "bytes": 9703664
+    },
+    {
+      "path": "assets/teaser_originals.png",
+      "sha256": "6099f12bdaea83fb2ce4f8e1ed282ac864d6f4ffb96911efb5a9101845656dae",
+      "bytes": 4000665
+    },
+    {
+      "path": "data/awsm.bib",
+      "sha256": "89ef6c3e8ad37bb83e9ed34be4a5348fea7d03f4f36c7019d4ab65ca2726b578",
+      "bytes": 386
+    },
+    {
+      "path": "data/embodied_demo.json",
+      "sha256": "716cf0950b01d09546799ae3caeda019a6c672ba3e979714ccd2c6dca41eea59",
+      "bytes": 3306
+    },
+    {
+      "path": "data/fixed_views.json",
+      "sha256": "67611f8ecc63b44187d494984c9fea6cfe6ab1fb2e99cbe9196ec9ba1e3fb911",
+      "bytes": 8773
+    },
+    {
+      "path": "data/office_video_media.json",
+      "sha256": "2a8329b175a06592b61153eb1f05450d6b40e899ddca7f348fb583a2eb4bac43",
+      "bytes": 3823
+    },
+    {
+      "path": "data/scene_comparison.json",
+      "sha256": "567d50e99a789db24b4e9c31bebd6d134502a748e78dc3d0ed96e757acf0a5e8",
+      "bytes": 11113
+    },
+    {
+      "path": "data/source_contract.json",
+      "sha256": "a0055850dab53d209b508a63939c21e5daf0dca73d518c4ce890542c63893af6",
+      "bytes": 14011
+    },
+    {
+      "path": "data/tables_1_7.json",
+      "sha256": "01faed4be8948d16fbdc081bda859554f3bfc3f20c55746fba0ca70369179287",
+      "bytes": 12692
+    },
+    {
+      "path": "editorial.css",
+      "sha256": "462f0ffc2b51cbfa93bd7187c16541ce3e28046440a5ce7d893775f16f727247",
+      "bytes": 7130
+    },
+    {
+      "path": "en.html",
+      "sha256": "a207a87659c1b12ac4b468f50ff230910a166b5b597fe9d22a7403a300049908",
+      "bytes": 41112
+    },
+    {
+      "path": "evidence/M1/manifest.json",
+      "sha256": "1035edb2af53edc907402472fa01b749100f01e4186b8437c4a26209fc64b952",
+      "bytes": 2132
+    },
+    {
+      "path": "evidence/M2/manifest.json",
+      "sha256": "1d702281c98bfd1995f5dc073b0d91dd548f639bb813d4d8023bd6cc236ffcbd",
+      "bytes": 2126
+    },
+    {
+      "path": "evidence/M3/manifest.json",
+      "sha256": "ab0f7a5542843ed7db656306f983051c9e16e7655a120e3bf8a9db16625b486e",
+      "bytes": 2127
+    },
+    {
+      "path": "evidence/M4/manifest.json",
+      "sha256": "9bbe33317f64a670eed12a689a0291a4cdda422719c3d62111c53a42d638347a",
+      "bytes": 2101
+    },
+    {
+      "path": "evidence/embodied-demo-audit.json",
+      "sha256": "4ea81e875634a2565aa50ac653a190a1544c154b516507345f8119e0ab2fc3f8",
+      "bytes": 994
+    },
+    {
+      "path": "evidence/embodied-demo-delivery.json",
+      "sha256": "fc2ec44f10a04f7b1c162e3071bf96e1e3a5249cdf4df25b4b5f47b84a675cca",
+      "bytes": 2434
+    },
+    {
+      "path": "evidence/provenance-note.md",
+      "sha256": "a6091e2335d01e5201b03df418e51e3cca50992945af1856fc4630bcdcb7bfed",
+      "bytes": 5842
+    },
+    {
+      "path": "index.html",
+      "sha256": "a207a87659c1b12ac4b468f50ff230910a166b5b597fe9d22a7403a300049908",
+      "bytes": 41112
+    },
+    {
+      "path": "models/M1/scene.blend",
+      "sha256": "80f985ad09591bdb227e99e71f6bacaa0b9e8c8b565c27da64ad4ac4ceb8910e",
+      "bytes": 2691023
+    },
+    {
+      "path": "models/M1/scene.glb",
+      "sha256": "21a605be65d58dbe01606b8326bfe6f07fa418c67e133c9b2181729b68bc2832",
+      "bytes": 15104860
+    },
+    {
+      "path": "models/M2/scene.blend",
+      "sha256": "c19f33ec434bbcf89137e67d81c283b9adfd48f42e993d8b29d63e08318f79a6",
+      "bytes": 3886116
+    },
+    {
+      "path": "models/M2/scene.glb",
+      "sha256": "0ee9123f38099d1df63ab593dd7082d5f1cf1b136e63833ec5afa0e072cf4c2e",
+      "bytes": 14523124
+    },
+    {
+      "path": "models/M3/scene.blend",
+      "sha256": "00cd43eef3b4a6b2b2b169e8a7eda787dc57222dc501a3a9be09f39af5f46a9b",
+      "bytes": 2319452
+    },
+    {
+      "path": "models/M3/scene.glb",
+      "sha256": "22699f8225a5f90166cdea2e6e2bdad7977fda4a1fe1086da209f0c9afe97a6c",
+      "bytes": 7701488
+    },
+    {
+      "path": "models/M4/scene.blend",
+      "sha256": "cc6cb605246a255d94e36b9dc3f8b91d5a292b047470bc0ff75b8a5c404b0cfe",
+      "bytes": 2268620
+    },
+    {
+      "path": "models/M4/scene.glb",
+      "sha256": "716e4b89b7b11ad8a0cb0e238b366b4f5dc80e3a2c3cddb5eedc469ef4e63774",
+      "bytes": 17202608
+    },
+    {
+      "path": "office-models.js",
+      "sha256": "8752a7cc4f64dfac992029d218e8ed1ee747f12e6ead0ea9fcbefc9293a8ea17",
+      "bytes": 15700
+    },
+    {
+      "path": "office.css",
+      "sha256": "2e8c4a2f0718520449a9a4534241290e387af74d2133e5db41f91c0682353f6f",
+      "bytes": 3447
+    },
+    {
+      "path": "scene-compare.js",
+      "sha256": "bd749465e05d4d19a746945f6914ff55bb06da743e5f5cd25e5630dd0cc0e732",
+      "bytes": 7483
+    },
+    {
+      "path": "style.css",
+      "sha256": "cbba47743ba27dea797754d7061d00aa7440649d472e35824537cc04cb49365f",
+      "bytes": 8962
+    },
+    {
+      "path": "vendor/three/controls/OrbitControls.js",
+      "sha256": "b97879c748170baadeb3fb84cea1ffdf4674e283dc06042f34e2acb95a76042c",
+      "bytes": 38703
+    },
+    {
+      "path": "vendor/three/environments/RoomEnvironment.js",
+      "sha256": "c20f0b4677f6128a138d7152b85cbef9091f4f45cdfa05adca04d40c1697c7ae",
+      "bytes": 4899
+    },
+    {
+      "path": "vendor/three/loaders/GLTFLoader.js",
+      "sha256": "67ac5551fdafa6e349bd80c8f8e5e39c136d6b2fb1ad647db9abb21dc86f9e4a",
+      "bytes": 114739
+    },
+    {
+      "path": "vendor/three/three.core.js",
+      "sha256": "eb077d2417f61d3e6d9264c317cabc4ea35769ed6b0ab533067292a550784c20",
+      "bytes": 1403455
+    },
+    {
+      "path": "vendor/three/three.module.js",
+      "sha256": "c8211c69345d2e9949dc7a8ac969380497aa0600a5a8ac6a459c8cd02dd9cb8a",
+      "bytes": 603113
+    },
+    {
+      "path": "vendor/three/utils/BufferGeometryUtils.js",
+      "sha256": "fda7e946b8e0b5ab39b779206589e7a1079a22eb24efb89d7223e03fdfb1f751",
+      "bytes": 35539
+    },
+    {
+      "path": "zh.html",
+      "sha256": "99c4919e62bd89da3c4ba6a8a98f948a597d005518ce234d07bbcdf06df6780a",
+      "bytes": 38832
+    }
+  ]
+}

index.html

@@ -0,0 +1,34 @@
+<!doctype html><html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"><title>AWSM: Agentic World Simulation and Mapping — Phygital AI</title><meta name="description" content="From real spaces to worlds phygital agents can use."><meta property="og:type" content="article"><meta property="og:locale" content="en_US"><meta property="og:title" content="AWSM: Agentic World Simulation and Mapping"><meta property="og:description" content="From real spaces to worlds phygital agents can use."><meta property="og:image" content="https://phygital-ai.github.io/agentic-world-simulation-and-mapping/assets/teaser_originals.png"><meta property="og:url" content="https://phygital-ai.github.io/agentic-world-simulation-and-mapping/index.html"><meta name="twitter:card" content="summary_large_image"><meta name="twitter:title" content="AWSM: Agentic World Simulation and Mapping"><meta name="twitter:description" content="From real spaces to worlds phygital agents can use."><meta name="twitter:image" content="https://phygital-ai.github.io/agentic-world-simulation-and-mapping/assets/teaser_originals.png"><link rel="canonical" href="https://phygital-ai.github.io/agentic-world-simulation-and-mapping/"><link rel="alternate" hreflang="zh-CN" href="https://phygital-ai.github.io/agentic-world-simulation-and-mapping/zh.html"><link rel="alternate" hreflang="en" href="https://phygital-ai.github.io/agentic-world-simulation-and-mapping/"><link rel="stylesheet" href="style.css?v=cbba47743ba2"><link rel="stylesheet" href="editorial.css?v=462f0ffc2b51"><script type="importmap">{"imports":{"three":"./vendor/three/three.module.js"}}</script><script src="app.js?v=d70db3a6816b" defer></script><script type="module" src="scene-compare.js?v=bd749465e05d"></script><link rel="stylesheet" href="office.css?v=2e8c4a2f0718"><link rel="alternate" hreflang="x-default" href="https://phygital-ai.github.io/agentic-world-simulation-and-mapping/"><meta property="article:published_time" content="2026-10-01"></head>
+<body data-lang="en"><a class="skip" href="#main">Skip to article</a><nav><a class="brand" href="index.html">PHYGITAL AI / AWSM</a><div><a href="zh.html" >中文</a><a href="index.html" aria-current='page'>EN</a><a href="https://github.com/wentingw/AWSM" target="_blank" rel="noopener noreferrer">GitHub ↗</a></div></nav><header><p class="title-name">Agentic World Simulation and Mapping</p><div class="title-lockup"><h1 aria-label="AWSM: Agentic World Simulation and Mapping">AWSM</h1><p class="pronunciation">pronounced “awesome”</p></div><p class="lead">From real spaces to <strong class="tagline-emphasis">worlds phygital agents can use.</strong></p><p class="brand-note">Phygital = physical + digital.</p><p class="publication-date">Published <time datetime="2026-10-01">October 1, 2026</time></p></header>
+<div class="hero" id="overview-video"><video controls playsinline preload="metadata" poster="assets/awsm-promo-v5-poster.jpg" aria-label="AWSM introduction video"><source src="assets/awsm-promo-v5.mp4" type="video/mp4"><a href="assets/awsm-promo-v5.mp4">AWSM introduction video</a></video><div class="project-links" role="group" aria-label="AWSM project resources"><a class="project-resource" href="https://github.com/wentingw/AWSM" target="_blank" rel="noopener noreferrer" aria-label="GitHub (opens in a new window)" title="GitHub (opens in a new window)"><img src="assets/icons/github.svg" width="24" height="24" alt="">GitHub<svg class="project-external-arrow" viewBox="0 0 16 16" width="12" height="12" fill="none" stroke="currentColor" stroke-width="1.6" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><path d="M4 12 12 4M4 4h8v8"/></svg></a><span class="project-resource"><img src="assets/icons/huggingface.svg" width="24" height="24" alt="">Hugging Face (Coming soon)</span></div></div><div class="layout"><aside class="toc"><p>ON THIS PAGE</p><a href="#results">Overview</a><a href="#motivation">Motivation</a><a href="#workflow">Workflow</a><a href="#more-results">Results & analysis</a><a href="#embodied-demo">Embodied demo</a><a href="#related-work">Related work</a><a href="#limitations">Limitations</a><a href="#citation">Citation</a><a href="#references">References</a></aside><main id="main">
+<section id="results"><p class="section-tag">01 / OVERVIEW</p><aside class="tldr"><strong>TL;DR</strong><p>AWSM uses large-model agents to turn visual observations into editable 3D scenes, grounded in camera-pose and depth estimates—including IMU-informed visual–inertial constraints that tie geometry to physical scale. In one controlled Isaac Sim scene, four reconstruction routes test spatial fidelity, not just visual plausibility. The aligned comparison from RGB-only M1 to GT-pose-conditioned M4 shows ≈81% lower bidirectional surface error and ≈54% lower model-depth AbsRel across 180 evaluation views. A multi-robot demo then illustrates goal-directed navigation and task execution on the reconstructed map: follow predefined routes to a marked reception desk and line up there. Next steps are higher efficiency and accuracy, reusable simulation-ready assets, and deeper integration with phygital agents for persistent spatial memory, object search, and interaction.</p></aside><h2>A scene can look plausible and still be spatially wrong.</h2><p class="standfirst">Agentic reconstruction uses a large model to inspect observations, call modeling tools, and iteratively build and check an editable scene. The challenge is not only to produce a convincing room, but to preserve the scale, shape, and spatial relationships that make it useful beyond rendering. AWSM asks how geometric evidence can constrain this process so that its outputs become more faithful spatial references for downstream agents.</p><p>We reconstruct World Lobby, a simulated scene in NVIDIA Isaac Sim, through four complete engineering routes, from RGB-only modeling to ground-truth-pose-conditioned modeling. Each route produces a frozen Blender scene, not merely a point cloud or a novel-view renderer. The experiment asks what additional geometric evidence buys, where it fails to help, and which errors survive object-centric reconstruction.</p><div class="insight-strip"><div><strong>0.376 → 0.072 m</strong><span>bidirectional surface error</span><small class="metric-reduction">≈81% relative reduction · M1 → M4</small></div><div><strong>16.44% → 7.60%</strong><span>model-depth AbsRel · 180 evaluation views</span><small class="metric-reduction">≈54% relative reduction · M1 → M4</small></div></div><p class="metric-caveat">Single-scene M1 → M4 diagnostic comparison: M1 uses GT-assisted Sim(3) alignment; M4 uses GT camera poses. Relative reductions compare complete pipelines, not IMU alone.</p></section>
+<section id="motivation"><p class="section-tag">02 / MOTIVATION</p><h2>From plausible reconstruction to a usable spatial reference</h2><p>A reconstruction agent can assemble a plausible room while getting a corner, distance, or passage wrong. For an embodied agent, those are not merely visual defects: they change where a destination lies and which route a body can follow. Geometry grounding means constraining the modeling process with pose, depth, and metric evidence where available, rather than relying on visual plausibility alone. The question is how those constraints survive the conversion from observations into editable objects.</p><p>The reception-desk task makes this connection concrete: mark a destination in the reconstructed map, specify routes, and execute them with robot controllers. We use ‘scene representation’ for the editable objects and geometry, ‘map’ for their navigation role, and ‘simulation assets’ for their integration into a simulator. The ‘world’ in AWSM refers to this reusable spatial environment, not a learned dynamics predictor. A persistent spatial foundation connecting navigation, memory, interaction, and simulation is the longer-term goal; the experiments and demo examine its geometric basis and one downstream use.</p><aside class="claim"><strong>Research question</strong><p>How does progressively stronger geometric evidence change the fidelity—and the remaining failure modes—of an editable object-centric world?</p></aside><h3>Contributions</h3><ol class="contributions"><li>A four-route study of geometry-grounded agentic reconstruction, keeping the scene, 180 RGB identities, modeling objective, and editable output format fixed while comparing complete pipelines with different geometric evidence.</li><li>An evaluation protocol that separates trajectory error, native depth, final-scene depth, surface geometry, and rendered appearance, with frozen Blend/GLB assets, reproducible evaluation subsets, public hashes, and interactive comparisons.</li><li>A map-based multi-robot simulation example connecting a reconstructed scene to destination marking, predefined routes, and controller execution. It illustrates downstream use rather than a controlled navigation comparison across reconstruction routes.</li></ol><div id="scale-and-sensors" class="analysis-block grounding-block"><h3>Geometry grounding: scale is necessary, but not sufficient</h3>
+<p>Monocular projection geometry leaves a global scale ambiguity; learned metric depth adds a useful prior, not an independent measurement of every real-world dimension. Calibrated RGB-D, a known baseline or length, and visual-inertial estimation can supply metric constraints. IMU fusion can recover metric motion under sufficient excitation and reliable initialization, bias estimation, synchronization, and calibration. An IMU is not a direct distance sensor, and its presence alone does not guarantee accurate scale.</p>
+<p>Scale, local shape, camera pose, and navigable connectivity must be checked separately. A curved corner simplified into a square cannot be repaired by rescaling alone. Pose-conditioned depth can improve reconstruction, but neither the DA3 ablation nor our M2/M3 results imply that every depth metric improves monotonically. The four routes below are system comparisons, not an isolated IMU ablation.</p>
+<p class="inline-sources"><a href="https://arxiv.org/abs/1708.03852" target="_blank" rel="noopener">VINS-Mono</a> · <a href="https://arxiv.org/html/2511.10647v1#S7.SS2.SSS3" target="_blank" rel="noopener">DA3 pose-conditioning ablation</a></p></div></section>
+<section id="workflow"><p class="section-tag">03 / WORKFLOW</p><h2>One agentic workflow, four reconstruction routes</h2><p>The reconstruction agent follows a shared observe–build–verify loop: inspect the available evidence, write Blender Python to construct objects, render review views, and revise the scene before freezing it for evaluation. This tool-using modeling process is what ‘agentic’ describes here; the downstream robot controllers have a separate role.</p><p>All routes receive the same 180 RGB identities and the same Astra–Blender modeling objective. M1 tests how far visual priors can go without metric measurements. M2 adds RGB-only ViPE poses and pose-conditioned DA3 depth. M3 uses monocular-inertial ORB-SLAM3 poses with the same class of DA3 observations. M4 supplies ground-truth camera poses to diagnose the remaining modeling error; it does not receive the GT mesh or GT depth as modeling input.</p><figure class="table-figure" id="table-1" data-table-key="table1"><figcaption><span>Table 1.</span> <span class="table-title"></span></figcaption><div class="table-scroll table-mount" role="region" tabindex="0" aria-label="Table 1"></div></figure><h3>Agentic reconstruction loop</h3><div class="workflow-strip"><div><strong>01 · Observe</strong><span>Inspect all 180 RGB frames and, where allowed, pose-conditioned DA3 geometry.</span></div><div><strong>02 · Build</strong><span>Write Blender Python that decomposes the room into named objects, materials, cameras, and collision proxies.</span></div><div><strong>03 · Verify</strong><span>Compare paired RGB and optical-Z renders at fixed review cameras; revise within a fixed version budget.</span></div><div><strong>04 · Freeze</strong><span>Export Blend/GLB, manifests, semantic IDs, and hashes before any GT score is visible to the author.</span></div><div><strong>05 · Evaluate</strong><span>Measure pose, native depth, model depth, surface distance, and appearance under one frozen protocol.</span></div></div><p>Ground truth is introduced only after the scenes are frozen. M2 and M3 are evaluated under one global SE(3) alignment with scale fixed to one; no mesh ICP or per-view fitting is used. M1 has no native metric trajectory, so its model-space numbers use one frozen GT-assisted Sim(3) from five manual camera associations and remain diagnostic rather than evidence of metric recovery.</p><aside class="notice"><strong>M1 diagnostic limitation.</strong> Its Sim(3)-aligned geometry and depth numbers describe shape after GT-assisted display alignment, not native metric recovery.</aside></section>
+<section id="more-results"><p class="section-tag">04 / RESULTS AND ANALYSIS</p><h2>Three findings matter more than a leaderboard</h2><article class="finding"><p class="finding-kicker">FINDING 01</p><h3>1. Better pose does not mechanically imply better native depth</h3><p>M3 improves trajectory accuracy over M2—ATE falls from 0.167 m to 0.121 m and rotation error from 0.31° to 0.19°. Yet its native DA3 AbsRel is slightly worse (20.38% versus 19.06%). After modeling, the ordering reverses: M3 reaches 8.37% model-depth AbsRel versus 9.29% for M2. Upstream metrics and downstream scene fidelity are related, but they are not interchangeable.</p></article><figure class="table-figure" id="table-2" data-table-key="table2"><figcaption><span>Table 2.</span> <span class="table-title"></span></figcaption><div class="table-scroll table-mount" role="region" tabindex="0" aria-label="Table 2"></div></figure><figure id="figure-1"><img loading="lazy" src="assets/figure14_abc.png" alt="Trajectory, native depth, and model depth"><figcaption><span>Figure 1.</span> The upstream trajectory and native-depth ranking does not map one-to-one onto frozen-scene depth.</figcaption></figure><article class="finding"><p class="finding-kicker">FINDING 02</p><h3>2. Geometric evidence turns a plausible layout into a more faithful space</h3><p>Under the declared alignment and without mesh ICP, bidirectional surface error falls from 0.376 m for the RGB-only diagnostic baseline to 0.118 m with ViPE, 0.082 m with ORB-SLAM3, and 0.072 m with GT poses. M4 is strongest on geometry in this case, while M3 closes most of the gap using an estimated trajectory.</p></article><figure class="table-figure" id="table-3" data-table-key="table3"><figcaption><span>Table 3.</span> <span class="table-title"></span></figcaption><div class="table-scroll table-mount" role="region" tabindex="0" aria-label="Table 3"></div></figure><article class="finding finding-comparison"><p class="finding-kicker">FINDING 03</p><h3>3. Qualitative comparison</h3></article><div class="analysis-block"><h3>Inspect the reconstructed scenes, not just aggregate scores</h3><p>Use the shared camera to compare each reconstruction with GT. Start with M1 to see how a semantically plausible room can drift in metric layout; compare M2 and M3 to find regions where geometry improves without a matching gain in RGB similarity; then inspect M4 to isolate errors that remain even when camera pose is no longer the bottleneck.</p><figure id="figure-2"><img loading="lazy" src="assets/fixed_five_view_comparison_m1_m4.jpg" alt="M1 to M4 and input GT across five fixed views"><figcaption><span>Figure 2.</span> M1–M4 and input GT across five fixed views.</figcaption></figure><figure class="scene-comparison" id="figure-3" data-state="idle"><h3>Rotate, zoom, and inspect the frozen scenes</h3>
+<div class="scene-toolbar"><label class="method-control">Reconstruction / GT<select id="model-select"><option value="M1">M1</option><option value="M2">M2</option><option value="M3">M3</option><option value="M4" selected>M4</option><option value="GT">GT</option></select></label><label>Display<select id="scene-mode"><option value="compare">Split comparison with GT</option><option value="single">View selected scene alone</option></select></label><button id="scene-reset" type="button">Reset view</button></div>
+<div class="scene-stage"><span class="scene-label scene-label-left">M4</span><span class="scene-label scene-label-right">GT</span><button class="scene-divider" type="button" role="slider" aria-label="Comparison divider" aria-valuemin="2" aria-valuemax="98" aria-valuenow="50"></button><input id="scene-split" class="scene-split" type="range" min="2" max="98" value="50" aria-label="Comparison split"><p class="scene-status">Scroll here to load the 3D comparison.</p></div><button id="scene-retry" class="scene-retry" type="button">Retry</button>
+<figcaption><span>Figure 3.</span> One camera and one full viewport compare the selected frozen model with GT. Ceilings and surrounding walls are hidden at runtime for the same cutaway-style inspection; source files remain unchanged.</figcaption></figure><figure class="fixed-comparison" id="figure-4"><div class="controlrow"><label>Method<select id="compare-method" disabled><option value="M1">M1</option><option value="M2">M2</option><option value="M3">M3</option><option value="M4">M4</option></select></label><label>Fixed view<select id="compare-frame" disabled><option value="0">0</option><option value="36">36</option><option value="72">72</option><option value="108">108</option><option value="144">144</option></select></label></div><div class="compare-pair"><div><img id="compare-pred" src="assets/fixed_views/M1/000.png" alt="Selected frozen model rendering"><p id="compare-label" class="compare-subcaption">M1 / VIEW 0</p></div><div><img id="compare-gt" src="assets/fixed_views/GT/000.png" alt="Input GT RGB"><p class="compare-subcaption">INPUT / GT RGB</p></div></div><figcaption><span>Figure 4.</span> Frozen-model rendering and input GT RGB from exactly the same view.</figcaption></figure><p>The browser hides ceilings and surrounding walls only for cutaway inspection. Runtime registration and display choices do not modify the downloadable frozen GLB or Blend files.</p></div><div class="analysis-block"><h3>What geometric grounding changes—and what it does not</h3><h4>Modeling behaves like a structured—and lossy—regularizer</h4><p>For M2–M4, the final scene depth is substantially closer to GT than native DA3 depth on the shared 180-view protocol: AbsRel falls by roughly 51–60%. One plausible interpretation is that object-level aggregation suppresses inconsistent local predictions. This is a hypothesis, not a controlled causal result: the same abstraction can also erase real surfaces and fine geometry.</p><h4>Measurement constrains generation; it does not solve appearance</h4><p>The strongest geometric route does not dominate PSNR or SSIM. Pose removes one source of uncertainty, but appearance still depends on material estimation, lighting, object detail, and renderer mismatch. Evaluating only RGB would miss spatial progress; evaluating only geometry would miss perceptual failure.</p><h4>The 100-view check measures consistency, not unseen-view generalization</h4><p>Table 7 uses a seeded set of 100 cameras drawn from the 175 modeling frames outside the fixed-five check. These are additional evaluation views relative to that check, but their RGB images were available during modeling. On this set, M4 has the lowest depth AbsRel (7.47%), while M3 has the lowest RMSE (1.090 m).</p><figure id="figure-5"><img loading="lazy" src="assets/model_depth_error_five_views_m1_m4.png" alt="Model-depth error at five fixed evaluation views"><figcaption><span>Figure 5.</span> Model-depth error at the fixed-five check; teal marks missing or invalid predictions.</figcaption></figure><figure class="table-figure" id="table-6" data-table-key="table6"><figcaption><span>Table 6.</span> <span class="table-title"></span></figcaption><div class="table-scroll table-mount" role="region" tabindex="0" aria-label="Table 6"></div></figure><figure class="table-figure" id="table-7" data-table-key="table7"><figcaption><span>Table 7.</span> <span class="table-title"></span></figcaption><div class="table-scroll table-mount" role="region" tabindex="0" aria-label="Table 7"></div></figure></div><div id="office-cafe" class="analysis-block external-case"><h3>Beyond the lobby: what survives in a phone-captured space?</h3><p>Office Café is not a second run of the World Lobby benchmark and is not quantitative validation under the same protocol. It is an external, real-capture case study showing how an object-centric scene, a solved source-camera trajectory, video-to-model comparison, and physics replays can coexist in one inspectable artifact. An author-observed failure—a curved real-world corner simplified into a square one—motivates separating local shape from global scale. A correct scale alone cannot repair that shape error; this is a qualitative observation, not another measured benchmark.</p><p class="office-links"><a href="https://office-cafe-vipe.hiwtishere.chatgpt.site/?lang=en" target="_blank" rel="noopener">Open full-screen model ↗</a><a href="https://office-cafe-vipe.hiwtishere.chatgpt.site/shake?lang=en" target="_blank" rel="noopener">Open shake experiment ↗</a></p><div class="office-external"><h4>Explore the original interactive apps</h4><div class="office-external-grid"><figure class="office-external-card"><a href="https://office-cafe-vipe.hiwtishere.chatgpt.site/?lang=en" target="_blank" rel="noopener noreferrer"><img loading="lazy" src="assets/office-cafe/office-pmodel-poster.jpg" alt="Interactive model: Local preview from the published model-render video."></a><figcaption><strong>Interactive model</strong><p>Local preview from the published model-render video.</p><button type="button" data-office-external="https://office-cafe-vipe.hiwtishere.chatgpt.site/?lang=en" aria-controls="office-external-slot">Load external app · English requested</button></figcaption></figure><figure class="office-external-card"><a href="https://office-cafe-vipe.hiwtishere.chatgpt.site/shake?lang=en" target="_blank" rel="noopener noreferrer"><img loading="lazy" src="assets/office-cafe/office-shake-heavy-poster.jpg" alt="Shake experiment: Local preview from the published strong-shake replay."></a><figcaption><strong>Shake experiment</strong><p>Local preview from the published strong-shake replay.</p><button type="button" data-office-external="https://office-cafe-vipe.hiwtishere.chatgpt.site/shake?lang=en" aria-controls="office-external-slot">Load external app · English requested</button></figcaption></figure></div><p class="office-external-note">Links request English with lang=en; the external host controls language and availability. These local content previews are not screenshots of the external website. If embedding is unavailable, use the full-screen links or the local models and videos below.</p><div id="office-external-slot" class="office-external-slot" hidden></div></div><div class="office-model-grid"><figure class="office-model-card"><div class="office-model-stage" data-src="assets/office-cafe/scene.glb" aria-label="Interactive Office Café ViPE model"></div><figcaption><strong>Current model</strong><span>Video plus estimated depth as input; drag to orbit, right-drag to pan, and scroll to zoom.</span></figcaption></figure><figure class="office-model-card"><div class="office-model-stage" data-src="assets/office-cafe/atlas-scene.glb" aria-label="Interactive legacy Office Café model"></div><figcaption><strong>Legacy model</strong><span>Video-only input; this legacy layout was not recalibrated against the ViPE scan.</span></figcaption></figure></div><p class="office-model-note">Both models open in an elevated cutaway view: upper geometry is clipped only for display to expose the interior. Toggle Cutaway to inspect the complete model; Reset view restores the overview. Authored coordinates, materials, and model files are unchanged. Linked orbit controls compare layouts, not registered geometry.</p><div class="office-video-grid office-video-sync"><figure class="office-video-card"><video data-synced-video muted preload="metadata" playsinline poster="assets/office-cafe/office-input-poster.jpg" src="assets/office-cafe/office-input-web.mp4?v=4915b934b5f9" aria-label="Input video"></video><figcaption><strong>Input video</strong><span>Original phone walkthrough.</span></figcaption></figure><figure class="office-video-card"><video data-synced-video muted preload="metadata" playsinline poster="assets/office-cafe/office-pmodel-poster.jpg" src="assets/office-cafe/office-pmodel-web.mp4?v=3b03fe58a192" aria-label="Procedural model"></video><figcaption><strong>Procedural model</strong><span>Model rendered along the source-camera trajectory.</span></figcaption></figure><figure class="office-video-card"><video data-synced-video muted preload="metadata" playsinline poster="assets/office-cafe/office-scan-poster.jpg" src="assets/office-cafe/office-scan-web.mp4?v=861a074637b4" aria-label="Point cloud"></video><figcaption><strong>Point cloud</strong><span>Scan rendered along the same trajectory.</span></figcaption></figure></div>
+<div class="office-video-controls"><button class="office-video-play" type="button" aria-label="Play all three videos" aria-pressed="false">▶</button><input class="office-video-progress" type="range" min="0" max="1000" value="0" aria-label="Shared video progress"><output class="office-video-time">00:00 / 01:10</output><p class="office-video-status" role="status" aria-live="polite">Shared playback · 70.1 seconds · 30 fps</p></div><div class="office-video-grid"><figure class="office-video-card office-video-card-wide"><video controls preload="metadata" playsinline poster="assets/office-cafe/office-shake-heavy-poster.jpg" src="assets/office-cafe/office-shake-heavy.mp4"></video><figcaption><strong>Strong earthquake effect</strong><span>Every object in the scene is interactable, and cabinet doors can open; this replay shows the physical response under strong shaking.</span></figcaption></figure></div><script type="module" src="office-models.js?v=8752a7cc4f64"></script></div></section>
+<section id="embodied-demo" class="embodied-demo">
+<p class="section-tag">DEMO / MAP-BASED EMBODIED EXECUTION</p><h2>A reconstructed map. A destination. Robots in motion.</h2><p>The demo connects reconstruction to downstream use in NVIDIA Isaac Sim. The reception desk is marked in the navigation map, routes are specified, and a drone, humanoid, quadruped, and wheeled robot execute the task. The same spatial reference connects a destination, route constraints, and motion control: a scene to inspect becomes a map to act with.</p>
+<blockquote class="mission-command"><span>Task intent: </span>Send the robots to the reception desk and have them line up there.</blockquote>
+<ol class="mission-flow" aria-label="Task workflow"><li><span>01</span>Task instruction</li><li><span>02</span>Locate the reception desk</li><li><span>03</span>Set routes on the map</li><li><span>04</span>Execute with controllers</li></ol>
+<figure class="demo-film"><video controls playsinline preload="none" poster="assets/embodied-demo/navigation-214924-poster.png" width="1280" height="720" aria-label="Four-robot map-based navigation demonstration"><source src="assets/embodied-demo/four-robots-214924-156s-phone.mp4" type="video/mp4">Your browser cannot play this video. Use the download link below.</video><figcaption>Map-based navigation in simulation · predefined routes · simulator-pose feedback. The authors confirm that the demonstration has been manually reviewed.</figcaption></figure>
+<div class="demo-links"><a href="assets/embodied-demo/four-robots-214924-156s-phone.mp4" download>Download mobile edition · 11.65 MB · 2:36</a><a href="assets/embodied-demo/four-robots-214924-156s-hd.mp4" download>Download HD edition · 49.69 MB · 2:36</a><a href="data/embodied_demo.json">Run &amp; media record ↗</a></div>
+<figure class="demo-map" id="demo-route-map"><a href="assets/embodied-demo/planned-route-156s.png" target="_blank" rel="noopener"><img loading="lazy" src="assets/embodied-demo/planned-route-156s.png" width="1920" height="1080" alt="Planned routes for four robots, with the reception desk and waypoints marked; static design, not execution trajectories"></a><figcaption><span>Navigation map.</span> The reception desk and planned routes share the reconstructed scene. This is a static design preview, not a recorded trajectory. Open the image to inspect the full-resolution map.</figcaption></figure>
+<details class="demo-protocol"><summary>Demo protocol and scope</summary><p>The instruction above describes the intended task, not a demonstrated language-to-plan parser. Routes were manually specified and screened against both the M4 reconstructed map and original geometry. This is not a reconstruction-only planning benchmark.</p><p>The 156-second presentation edit plays at 1×: Scene (0–6s), Reconstruction (6–12s), Planned route (12–18s), and Navigation (18–156s). Navigation uses the new run <code>20260930-214924</code>; the opening 12.32 seconds of waiting are omitted, and the later terminal hold is outside this cut. The map is a static route design, not a measured execution trace.</p><p>This demo does not evaluate online visual localization, learned instruction understanding, memory retrieval, or real-robot transfer. Those are separate capabilities; the map-based execution shown here remains the intended demonstration.</p><p>Manual presentation review is separate from the historical automated checks retained below; it does not revise their recorded verdicts.</p><p><a href="evidence/embodied-demo-audit.json">Historical automated check record ↗</a></p></details>
+</section>
+<section id="related-work"><p class="section-tag">05 / RELATED WORK</p><h2>From geometric measurement to editable Real2Sim</h2><p>AWSM connects two roles: an agent that constructs an editable scene, and downstream agents that use it as a spatial reference. SLAM and visual-inertial estimation supply pose and metric constraints; ViPE and Depth Anything 3 contribute geometric observations. Concurrent work AHa-3D explores video-driven, tool-based Real2Sim, including camera and 3D-reference estimation from video. Our emphasis is geometric grounding: bringing additional physical measurements, especially IMU-informed visual-inertial pose and metric constraints, into agentic scene construction rather than relying on video-derived evidence alone. The four-route comparison tests reconstruction systems, not the isolated effect of IMU; the demo illustrates a map-based application, not a complete autonomous embodied system.</p><p>Our evaluation therefore spans both sides of the interface: upstream systems are judged as measurement pipelines, while the Blender outputs are judged as persistent spatial artifacts. This differs from evaluating pose, depth, or novel-view synthesis in isolation.</p><span class="inline-sources"><a href="https://research.nvidia.com/labs/toronto-ai/vipe/" target="_blank" rel="noopener">ViPE</a> · <a href="https://arxiv.org/abs/2007.11898" target="_blank" rel="noopener">ORB-SLAM3</a> · <a href="https://arxiv.org/abs/2511.10647" target="_blank" rel="noopener">Depth Anything 3</a> · <a href="https://kevinxu02.github.io/real2sim-indoor-site/" target="_blank" rel="noopener">AHa-3D</a></span><div id="spatial-foundation" class="analysis-block spatial-foundation"><p class="section-tag">RESEARCH DIRECTION</p><h3>Toward persistent spaces for phygital agents</h3><p>The current outputs are editable scene snapshots, and the demo shows a scene-specific simulation integration. The longer-term goal is to make geometry-grounded agentic reconstruction a reusable process for building and maintaining spaces that phygital agents can use. Saving an editable asset is a starting point, not a demonstration of lifelong map maintenance. The directions below extend the present evidence.</p><div class="capability-grid"><div><h4>Navigate</h4><p>Use destinations, free space, and body-specific clearance in a shared geometric frame. Better reconstruction can reduce map mismatch; reliable navigation also needs localization and current observations.</p></div><div><h4>Remember & retrieve</h4><p>Link object identities and locations to source views so language, images, and past observations can refer to the same place. Retrieval and safe arrival should be evaluated separately.</p></div><div><h4>Interact & maintain</h4><p>Edit object properties and relationships, revisit changed areas, and preserve evidence and versions. Persistent memory must distinguish observed geometry, inferred completion, and unknown space.</p></div><div><h4>Simulate & learn</h4><p>Move from the current scene-specific integration to reusable simulation-ready exports with checked scale, collision geometry, and physical parameters. These can support task rehearsal, scenario variation, and training; reduced policy sim-to-real gap remains a separate question to test.</p></div></div><p>A full mesh is not required for every retrieval task: Memory Over Maps uses posed RGB-D keyframes for on-demand localization, while 3D-Mem and task-oriented scene graphs explore complementary memory and representation choices. Our proposed direction is hybrid: preserve visual evidence, refine task-relevant geometry, and build editable simulation assets where they add value. The next question is when an agent has enough evidence to act—and when it should observe again and update its map.</p><p>The immediate agenda is higher reconstruction efficiency and accuracy, reusable simulation-ready scene generation, and deeper integration with phygital agents. Together, these steps connect the reconstruction agent’s modeling and revision loop to the downstream agent’s navigation, memory, and interaction needs.</p><p class="inline-sources"><a href="https://arxiv.org/abs/2603.20530" target="_blank" rel="noopener">Memory Over Maps</a> · <a href="https://arxiv.org/abs/2411.17735" target="_blank" rel="noopener">3D-Mem</a> · <a href="https://arxiv.org/abs/2404.13696" target="_blank" rel="noopener">Clio</a> · <a href="https://arxiv.org/abs/2404.06609" target="_blank" rel="noopener">GOAT-Bench</a></p></div></section>
+<section id="limitations"><p class="section-tag">06 / LIMITATIONS</p><h2>The boundaries are part of the result</h2><ul class="limits-list"><li>One synthetic scene and one engineering run per route do not establish general superiority or statistical significance.</li><li>The four routes are complete systems, not a strict single-variable ablation; M2 versus M3 cannot be attributed to IMU alone.</li><li>M4 uses GT camera poses and is a diagnostic route, not a deployable baseline or a theoretical upper bound.</li><li>M1 model-space metrics rely on GT-assisted Sim(3) and must not be ranked as native metric recovery.</li><li>The 100-view appearance and depth sets contain modeling RGB inputs; they are not a held-out novel-view benchmark.</li><li>Appearance metrics mix geometry, material, illumination, and rendering error. M4 revision 4 has not received a complete independent visual re-review.</li><li>The reception-desk demo is map-based simulation execution, not a controlled comparison of navigation across M1–M4. Routes are predefined and localization uses simulator pose; see the demo protocol for scope.</li><li>Multimodal memory, long-term map maintenance, and reduced policy sim-to-real gap are research goals, not measured outcomes of the four-robot demo. The Office Café corner observation has no matched-view quantitative GT measurement here.</li><li>Historical downstream documentation describes the same M4 Blend hash using a different depth frontend. The frozen tables and assets are retained unchanged; their upstream naming needs reconciliation before attributing historical task outcomes to DA3. See the provenance note.</li></ul><p>Every table is generated from frozen assets. The public manifests expose model hashes, registrations, fixed views, and the exact seeded evaluation indices.</p><div class="asset-downloads"><p><strong>M1</strong> <a href="models/M1/scene.blend" download>scene.blend ↓</a> · <a href="models/M1/scene.glb" download>scene.glb ↓</a> · <a href="evidence/M1/manifest.json">manifest ↗</a></p><p><strong>M2</strong> <a href="models/M2/scene.blend" download>scene.blend ↓</a> · <a href="models/M2/scene.glb" download>scene.glb ↓</a> · <a href="evidence/M2/manifest.json">manifest ↗</a></p><p><strong>M3</strong> <a href="models/M3/scene.blend" download>scene.blend ↓</a> · <a href="models/M3/scene.glb" download>scene.glb ↓</a> · <a href="evidence/M3/manifest.json">manifest ↗</a></p><p><strong>M4</strong> <a href="models/M4/scene.blend" download>scene.blend ↓</a> · <a href="models/M4/scene.glb" download>scene.glb ↓</a> · <a href="evidence/M4/manifest.json">manifest ↗</a></p></div><p class="evidence-links"><a href="data/tables_1_7.json">Tables JSON</a> · <a href="data/scene_comparison.json">Scene manifest</a> · <a href="data/fixed_views.json">Fixed views</a> · <a href="evidence/publication_manifest.json">Publication manifest</a> · <a href="evidence/SHA256SUMS">SHA256</a></p><p><a href="evidence/provenance-note.md">Version and provenance note ↗</a></p></section>
+<section id="citation"><p class="section-tag">07 / CITATION</p><h2>Cite this work</h2><p>If this controlled study, its frozen assets, or its evaluation protocol supports your work, please cite the interactive article below.</p><pre class="citation-block"><code>@misc{awsm_2026,
+  title        = {{AWSM}: Agentic World Simulation and Mapping},
+  author       = {{Phygital AI}},
+  year         = {2026},
+  howpublished = {Interactive research article},
+  url          = {https://phygital-ai.github.io/agentic-world-simulation-and-mapping/},
+  note         = {Reconstruction benchmarks, editable scene assets, and map-based embodied demonstration}
+}</code></pre><a class="citation-download" href="data/awsm.bib" download="awsm.bib">Download BibTeX <span aria-hidden="true">↓</span></a></section>
+<section id="references"><p class="section-tag">08 / REFERENCES</p><h2>References</h2><ol class="references-list"><li><a href="https://kevinxu02.github.io/real2sim-indoor-site/" target="_blank" rel="noopener">AHa-3D: Agentic Tool Use for Real2Sim with GPT-6 Astra.</a></li><li><a href="https://arxiv.org/abs/2508.10934" target="_blank" rel="noopener">ViPE: Video Pose Engine for camera motion and near-metric geometry.</a></li><li><a href="https://arxiv.org/abs/2007.11898" target="_blank" rel="noopener">Campos et al. ORB-SLAM3: An Accurate Open-Source Library for Visual, Visual–Inertial, and Multimap SLAM. IEEE Transactions on Robotics, 2021.</a></li><li><a href="https://arxiv.org/abs/2511.10647" target="_blank" rel="noopener">Depth Anything 3: Recovering the Visual Space from Any Views.</a></li><li><a href="https://doi.org/10.1109/TIP.2003.819861" target="_blank" rel="noopener">Wang et al. Image Quality Assessment: From Error Visibility to Structural Similarity. IEEE Transactions on Image Processing, 2004.</a></li><li><a href="https://arxiv.org/abs/1801.03924" target="_blank" rel="noopener">Zhang et al. The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. CVPR, 2018.</a></li><li><a href="https://www.blender.org/" target="_blank" rel="noopener">Blender Foundation. Blender: Free and Open Source 3D Creation Suite.</a></li><li><a href="https://arxiv.org/abs/1708.03852" target="_blank" rel="noopener">Qin et al. VINS-Mono: A Robust and Versatile Monocular Visual-Inertial State Estimator. TRO, 2018.</a></li><li><a href="https://arxiv.org/abs/2509.13414" target="_blank" rel="noopener">Keetha et al. MapAnything: Universal Feed-Forward Metric 3D Reconstruction. 2025.</a></li><li><a href="https://arxiv.org/abs/2603.20530" target="_blank" rel="noopener">Zhou et al. Memory Over Maps: 3D Object Localization Without Reconstruction. 2026.</a></li><li><a href="https://arxiv.org/abs/2411.17735" target="_blank" rel="noopener">Yang et al. 3D-Mem: 3D Scene Memory for Embodied Exploration and Reasoning. 2024.</a></li><li><a href="https://arxiv.org/abs/2404.06609" target="_blank" rel="noopener">Khanna et al. GOAT-Bench: A Benchmark for Multi-Modal Lifelong Navigation. CVPR, 2024.</a></li><li><a href="https://arxiv.org/abs/2309.16650" target="_blank" rel="noopener">ConceptGraphs: Open-Vocabulary 3D Scene Graphs for Perception and Planning. ICRA, 2024.</a></li><li><a href="https://arxiv.org/abs/2404.13696" target="_blank" rel="noopener">Clio: Real-time Task-Driven Open-Set 3D Scene Graphs. RA-L, 2024.</a></li></ol></section><noscript>JavaScript is required for tables and interactive figures.</noscript></main></div><footer>Phygital AI · AWSM · frozen evidence, editable outputs · <a href="zh.html">中文</a></footer></body></html>

office-models.js

@@ -0,0 +1,395 @@
+import * as THREE from "three";
+import { GLTFLoader } from "./vendor/three/loaders/GLTFLoader.js";
+import { OrbitControls } from "./vendor/three/controls/OrbitControls.js";
+
+const stages = [...document.querySelectorAll(".office-model-stage")];
+const loader = new GLTFLoader();
+const viewers = [];
+let syncingView = false;
+
+function syncView(source) {
+  if (syncingView || !source.ready) return;
+  syncingView = true;
+  const sourceTargetOffset = source.controls.target.clone().sub(source.center);
+  const sourceCameraOffset = source.camera.position.clone().sub(source.controls.target);
+  for (const viewer of viewers) {
+    if (viewer === source || !viewer.ready) continue;
+    const scale = viewer.radius / source.radius;
+    viewer.controls.target.copy(viewer.center).addScaledVector(sourceTargetOffset, scale);
+    viewer.camera.position.copy(viewer.controls.target).addScaledVector(sourceCameraOffset, scale);
+    viewer.controls.update();
+  }
+  syncingView = false;
+}
+
+function mount(stage) {
+  if (stage.dataset.loaded) return;
+  stage.dataset.loaded = "true";
+
+  const scene = new THREE.Scene();
+  scene.background = new THREE.Color(0xe9ede5);
+
+  const camera = new THREE.PerspectiveCamera(38, 1, 0.01, 500);
+  const renderer = new THREE.WebGLRenderer({ antialias: true, powerPreference: "high-performance" });
+  renderer.setPixelRatio(Math.min(window.devicePixelRatio, 1.5));
+  renderer.outputColorSpace = THREE.SRGBColorSpace;
+  renderer.toneMapping = THREE.ACESFilmicToneMapping;
+  renderer.toneMappingExposure = 1.05;
+  stage.replaceChildren(renderer.domElement);
+
+  scene.add(new THREE.HemisphereLight(0xffffff, 0x8b948c, 2.2));
+  const sun = new THREE.DirectionalLight(0xfff4dc, 2.8);
+  sun.position.set(8, 14, 10);
+  scene.add(sun);
+
+  const controls = new OrbitControls(camera, renderer.domElement);
+  controls.enableDamping = true;
+  controls.dampingFactor = 0.08;
+  controls.screenSpacePanning = true;
+  const viewer = {
+    stage,
+    camera,
+    controls,
+    center: new THREE.Vector3(),
+    radius: 1,
+    ready: false,
+  };
+  const zh = document.documentElement.lang.startsWith("zh");
+  const toolbar = document.createElement("div");
+  toolbar.style.cssText = "position:absolute;top:10px;left:10px;display:flex;gap:8px;z-index:1";
+  const reset = document.createElement("button");
+  reset.type = "button";
+  reset.textContent = zh ? "重置视角" : "Reset view";
+  const cutaway = document.createElement("button");
+  cutaway.type = "button";
+  cutaway.textContent = zh ? "剖切视图" : "Cutaway";
+  cutaway.setAttribute("aria-pressed", "true");
+  for (const button of [reset, cutaway]) {
+    button.style.cssText = "padding:6px 10px;border:1px solid #bdcbbf;border-radius:4px;background:#fffffff0;color:#196451;font:12px sans-serif;cursor:pointer";
+    toolbar.append(button);
+  }
+  stage.append(toolbar);
+  let sectionPlane;
+  let initialDistance = 1;
+  let previousAspect = 1;
+  const direction = new THREE.Vector3(0.45, 1, 0.65).normalize();
+  const resetView = () => {
+    if (!viewer.ready) return;
+    const damping = controls.enableDamping;
+    controls.enableDamping = false;
+    controls.update();
+    controls.target.copy(viewer.center);
+    camera.position.copy(viewer.center).addScaledVector(direction, initialDistance);
+    controls.update();
+    controls.enableDamping = damping;
+    syncView(viewer);
+  };
+  reset.addEventListener("click", resetView);
+  cutaway.addEventListener("click", () => {
+    const enabled = cutaway.getAttribute("aria-pressed") !== "true";
+    cutaway.setAttribute("aria-pressed", String(enabled));
+    renderer.clippingPlanes = enabled && sectionPlane ? [sectionPlane] : [];
+  });
+  viewers.push(viewer);
+  controls.addEventListener("change", () => syncView(viewer));
+
+  const resize = () => {
+    const { width, height } = stage.getBoundingClientRect();
+    renderer.setSize(width, height, false);
+    camera.aspect = width / height;
+    if (viewer.ready) {
+      const factor = Math.max(1, 1 / camera.aspect) / Math.max(1, 1 / previousAspect);
+      initialDistance *= factor;
+      camera.position.sub(controls.target).multiplyScalar(factor).add(controls.target);
+    }
+    previousAspect = camera.aspect;
+    camera.updateProjectionMatrix();
+  };
+
+  loader.load(
+    stage.dataset.src,
+    ({ scene: model }) => {
+      scene.add(model);
+      const fullBox = new THREE.Box3().setFromObject(model);
+      const sectionHeight = fullBox.min.y + (fullBox.max.y - fullBox.min.y) * 0.64;
+      const box = new THREE.Box3();
+      model.traverse((node) => {
+        if (!node.isMesh) return;
+        const meshBox = new THREE.Box3().setFromObject(node);
+        if (meshBox.min.y > sectionHeight) return;
+        meshBox.max.y = Math.min(meshBox.max.y, sectionHeight);
+        box.union(meshBox);
+      });
+      const center = box.getCenter(new THREE.Vector3());
+      const size = box.getSize(new THREE.Vector3());
+      const radius = box.getBoundingSphere(new THREE.Sphere()).radius;
+      center.y = box.min.y + size.y * 0.25;
+      viewer.center.copy(center);
+      viewer.radius = radius;
+      sectionPlane = new THREE.Plane(new THREE.Vector3(0, -1, 0), sectionHeight);
+      renderer.clippingPlanes = cutaway.getAttribute("aria-pressed") === "true" ? [sectionPlane] : [];
+      initialDistance = radius / Math.sin(THREE.MathUtils.degToRad(camera.fov / 2)) * Math.max(1, 1 / camera.aspect);
+      viewer.ready = true;
+      camera.near = Math.max(radius / 1000, 0.01);
+      camera.far = radius * 20;
+      camera.updateProjectionMatrix();
+      const readySource = viewers.find((candidate) => candidate !== viewer && candidate.ready);
+      if (readySource) syncView(readySource);
+      else resetView();
+      stage.dataset.ready = "true";
+      stage.sceneDiagnostics = () => ({
+        ready: viewer.ready,
+        camera: camera.position.toArray(),
+        target: controls.target.toArray(),
+        cutaway: renderer.clippingPlanes.length > 0,
+        sectionHeight: sectionPlane.constant,
+        bounds: { min: box.min.toArray(), max: box.max.toArray() },
+      });
+    },
+    undefined,
+    () => {
+      stage.classList.add("is-error");
+      stage.insertAdjacentHTML("beforeend", "<p>Model failed to load.</p>");
+    },
+  );
+
+  new ResizeObserver(resize).observe(stage);
+  resize();
+
+  renderer.setAnimationLoop(() => {
+    controls.update();
+    renderer.render(scene, camera);
+  });
+}
+
+if (stages.length) {
+  const observer = new IntersectionObserver(
+    (entries) => {
+      for (const entry of entries) {
+        if (entry.isIntersecting) {
+          mount(entry.target);
+          observer.unobserve(entry.target);
+        }
+      }
+    },
+    { rootMargin: "300px" },
+  );
+  stages.forEach((stage) => observer.observe(stage));
+}
+
+const syncedVideos = [...document.querySelectorAll("[data-synced-video]")];
+const videoPlay = document.querySelector(".office-video-play");
+const videoProgress = document.querySelector(".office-video-progress");
+const videoTime = document.querySelector(".office-video-time");
+const videoStatus = document.querySelector(".office-video-status");
+
+if (syncedVideos.length === 3 && videoPlay && videoProgress && videoTime && videoStatus) {
+  const primary = syncedVideos[0];
+  const zh = document.documentElement.lang.startsWith("zh");
+  const copy = zh ? {
+    play: "播放全部三个视频", pause: "暂停全部三个视频",
+    ready: "同步播放 · 70.1 秒 · 30 fps", waiting: "正在加载，三个视频就绪后将同步播放…",
+    interrupted: "播放已中断，请点击播放继续。",
+    failedVideo: (name) => `${name}加载失败，请点击播放重试。`,
+  } : {
+    play: "Play all three videos", pause: "Pause all three videos",
+    ready: "Shared playback · 70.1 seconds · 30 fps", waiting: "Loading — all three videos will start together…",
+    interrupted: "Playback was interrupted. Press play to continue.",
+    failedVideo: (name) => `Unable to load ${name}. Press play to retry.`,
+  };
+  let wanted = false;
+  let waiting = false;
+  let starting = false;
+  let generation = 0;
+  let targetTime = 0;
+  let finished = false;
+  let startRetries = 0;
+  let retryAt = 0;
+  const duration = () => Number.isFinite(primary.duration) ? primary.duration : 70.1;
+  const formatTime = (seconds) => {
+    const safe = Number.isFinite(seconds) ? seconds : 0;
+    return `${String(Math.floor(safe / 60)).padStart(2, "0")}:${String(Math.floor(safe % 60)).padStart(2, "0")}`;
+  };
+  const updateProgress = () => {
+    const time = waiting ? targetTime : primary.currentTime;
+    if (!videoProgress.matches(":active")) videoProgress.value = String((time / duration()) * 1000);
+    videoTime.textContent = `${formatTime(time)} / ${formatTime(duration())}`;
+    videoProgress.setAttribute("aria-valuetext", videoTime.textContent);
+  };
+  const setPlaying = (playing) => {
+    videoPlay.textContent = playing ? "❚❚" : "▶";
+    videoPlay.setAttribute("aria-label", playing ? copy.pause : copy.play);
+    videoPlay.setAttribute("aria-pressed", String(playing));
+  };
+  const pauseAll = () => syncedVideos.forEach((video) => {
+    video.pause();
+    video.playbackRate = 1;
+  });
+  const align = () => syncedVideos.forEach((video) => {
+    if (Number.isFinite(video.duration) && Math.abs(video.currentTime - targetTime) > 0.04) {
+      video.currentTime = Math.min(targetTime, video.duration);
+    }
+  });
+  const stop = (message = copy.ready) => {
+    wanted = waiting = starting = false;
+    generation += 1;
+    pauseAll();
+    setPlaying(false);
+    videoStatus.textContent = message;
+    updateProgress();
+  };
+  const ready = () => syncedVideos.every((video) => !video.seeking && video.readyState >= 3);
+  const bufferedAhead = (video) => {
+    for (let i = 0; i < video.buffered.length; i += 1) {
+      if (video.buffered.start(i) <= video.currentTime + 0.05 && video.buffered.end(i) > video.currentTime) {
+        return video.buffered.end(i) - video.currentTime;
+      }
+    }
+    return 0;
+  };
+  // Some browsers stop preloading after a small byte budget. Let an idle
+  // decoder start so play() can resume the download; active downloads can
+  // build a buffer first. This also avoids a play/pause loop on slow links.
+  const buffered = () => syncedVideos.every((video) => {
+    // A paused decoder may stay at HAVE_CURRENT_DATA until play() is called,
+    // even with future bytes buffered. Requiring canplay here can deadlock.
+    if (video.seeking || video.error || video.readyState < HTMLMediaElement.HAVE_CURRENT_DATA) return false;
+    if (video.networkState === HTMLMediaElement.NETWORK_IDLE) return true;
+    return bufferedAhead(video) >= Math.min(0.5, Math.max(0, video.duration - video.currentTime - 0.05));
+  });
+  const resume = async () => {
+    if (!wanted || !waiting || starting || performance.now() < retryAt || !buffered()) return;
+    const attempt = ++generation;
+    waiting = false;
+    starting = true;
+    const results = await Promise.allSettled(syncedVideos.map((video) => video.play()));
+    if (attempt !== generation) return;
+    starting = false;
+    const rejected = results.find((result) => result.status === "rejected");
+    if (rejected) {
+      // pause(), seeking and load() can cancel play() without a media failure.
+      const mediaFailure = syncedVideos.find((video) => video.error);
+      if (!mediaFailure && rejected.reason?.name === "AbortError" && startRetries < 2) {
+        startRetries += 1;
+        targetTime = Math.min(...syncedVideos.map((video) => video.currentTime));
+        pauseAll();
+        align();
+        waiting = true;
+        retryAt = performance.now() + 250;
+        videoStatus.textContent = copy.waiting;
+        return;
+      }
+      stop(mediaFailure ? copy.failedVideo(mediaFailure.getAttribute("aria-label")) : copy.interrupted);
+      return;
+    }
+    startRetries = 0;
+    videoStatus.textContent = copy.ready;
+  };
+  const hold = (time = primary.currentTime) => {
+    generation += 1;
+    starting = false;
+    waiting = true;
+    targetTime = time;
+    pauseAll();
+    align();
+    if (wanted) videoStatus.textContent = copy.waiting;
+    updateProgress();
+    void resume();
+  };
+  const waitForBuffer = () => hold(Math.min(...syncedVideos.map((video) => video.currentTime)));
+
+  syncedVideos.forEach((video) => {
+    video.muted = true;
+    video.playsInline = true;
+    for (const event of ["loadedmetadata", "loadeddata", "canplay", "seeked", "progress"]) {
+      video.addEventListener(event, () => {
+        if (waiting) align();
+        updateProgress();
+        void resume();
+      });
+    }
+    video.addEventListener("waiting", () => {
+      if (wanted && !waiting && !ready()) waitForBuffer();
+    });
+    video.addEventListener("error", () => {
+      // Ignore an old queued error after a successful load() retry.
+      if (video.error) stop(copy.failedVideo(video.getAttribute("aria-label")));
+    });
+    video.addEventListener("ended", () => {
+      if (!wanted) return;
+      finished = true;
+      const end = duration();
+      stop();
+      // Seeking exactly to duration may wrap a WebKit decoder back to zero.
+      syncedVideos.forEach((peer) => {
+        if (Math.abs(peer.currentTime - end) > 0.08) peer.currentTime = Math.max(0, end - 1 / 30);
+      });
+      updateProgress();
+    });
+  });
+  videoProgress.addEventListener("input", () => {
+    finished = false;
+    hold((Number(videoProgress.value) / 1000) * duration());
+  });
+  videoPlay.addEventListener("click", () => {
+    if (wanted) {
+      stop();
+      return;
+    }
+    wanted = true;
+    startRetries = 0;
+    retryAt = 0;
+    setPlaying(true);
+    syncedVideos.forEach((video) => {
+      video.preload = "auto";
+      if (video.error || video.networkState === HTMLMediaElement.NETWORK_NO_SOURCE) video.load();
+    });
+    const time = waiting ? targetTime : primary.currentTime;
+    const replay = finished || syncedVideos.some((video) => video.ended) || time >= duration() - 0.08;
+    finished = false;
+    hold(replay ? 0 : time);
+  });
+  document.addEventListener("visibilitychange", () => {
+    if (document.hidden) stop();
+  });
+  const tick = () => {
+    if (wanted && waiting) void resume();
+    if (wanted && !waiting && !starting) {
+      if (!ready()) {
+        waitForBuffer();
+      } else {
+        for (const video of syncedVideos.slice(1)) {
+          const drift = video.currentTime - primary.currentTime;
+          if (Math.abs(drift) > 0.12) {
+            hold();
+            break;
+          }
+          // Small rate adjustments avoid repeated seeks for sub-frame drift.
+          video.playbackRate = Math.abs(drift) > 0.025 ? (drift > 0 ? 0.96 : 1.04) : 1;
+        }
+      }
+    }
+    updateProgress();
+    requestAnimationFrame(tick);
+  };
+  requestAnimationFrame(tick);
+  updateProgress();
+}
+
+const officeExternalSlot = document.querySelector('#office-external-slot');
+for (const button of document.querySelectorAll('[data-office-external]')) {
+  button.addEventListener('click', () => {
+    const source = button.dataset.officeExternal;
+    if (!['https://office-cafe-vipe.hiwtishere.chatgpt.site/?lang=en', 'https://office-cafe-vipe.hiwtishere.chatgpt.site/shake?lang=en'].includes(source)) return;
+    const frame = document.createElement('iframe');
+    frame.src = source;
+    frame.lang = 'en';
+    frame.title = source.includes('/shake') ? 'Office Café shake experiment — English requested' : 'Office Café interactive model — English requested';
+    frame.allow = 'fullscreen';
+    frame.allowFullscreen = true;
+    frame.referrerPolicy = 'no-referrer';
+    frame.setAttribute('sandbox', 'allow-scripts allow-same-origin allow-pointer-lock');
+    officeExternalSlot.replaceChildren(frame);
+    officeExternalSlot.hidden = false;
+  });
+}

office.css

@@ -0,0 +1,153 @@
+.office-model-grid,
+.office-video-grid {
+  display: grid;
+  grid-template-columns: repeat(2, minmax(0, 1fr));
+  gap: 20px;
+  margin: 28px 0;
+}
+
+.office-external-grid { display: grid; grid-template-columns: repeat(2, minmax(0, 1fr)); gap: 20px; }
+.office-external-card { margin: 0; overflow: hidden; border: 1px solid var(--line); border-radius: 6px; background: var(--white); }
+.office-external-card img { width: 100%; aspect-ratio: 16 / 9; object-fit: cover; }
+.office-external-card figcaption { padding: 14px; margin: 0; }
+.office-external-card button { max-width: 100%; padding: 8px 12px; border: 1px solid var(--line); border-radius: 4px; background: var(--paper); color: var(--green); font: inherit; cursor: pointer; }
+.office-external-note { font-size: 12px; color: var(--muted); }
+.office-external-slot iframe { width: 100%; height: 560px; border: 1px solid var(--line); border-radius: 6px; background: var(--white); }
+@media (max-width: 700px) { .office-external-grid { grid-template-columns: 1fr; } .office-external-slot iframe { height: 460px; } }
+
+.office-model-card,
+.office-video-card {
+  margin: 0;
+  overflow: hidden;
+  border: 1px solid var(--line);
+  border-radius: 6px;
+  background: var(--white);
+}
+
+.office-model-stage {
+  position: relative;
+  width: 100%;
+  aspect-ratio: 4 / 3;
+  overflow: hidden;
+  background: #e9ede5;
+}
+
+.office-model-stage canvas {
+  display: block;
+  width: 100%;
+  height: 100%;
+  touch-action: none;
+}
+
+.office-model-card figcaption,
+.office-video-card figcaption {
+  margin: 0;
+  padding: 14px 16px 16px;
+}
+
+.office-model-card figcaption strong,
+.office-video-card figcaption strong {
+  display: block;
+  margin-bottom: 4px;
+  color: var(--ink);
+  font-size: 14px;
+}
+
+.office-model-card figcaption span,
+.office-video-card figcaption span {
+  display: block;
+  line-height: 1.55;
+}
+
+.office-video-card video {
+  display: block;
+  width: 100%;
+  aspect-ratio: 16 / 9;
+  background: #111;
+  object-fit: cover;
+}
+
+.office-video-card-wide {
+  grid-column: 1 / -1;
+}
+
+.office-video-sync {
+  grid-template-columns: repeat(3, minmax(0, 1fr));
+  margin-bottom: 0;
+}
+
+.office-video-status {
+  grid-column: 1 / -1;
+  margin: 0;
+  color: var(--muted);
+  font-size: 12px;
+}
+
+.office-video-controls {
+  display: grid;
+  grid-template-columns: auto minmax(0, 1fr) auto;
+  gap: 12px;
+  align-items: center;
+  margin: 0 0 28px;
+  padding: 12px 14px;
+  border: 1px solid var(--line);
+  border-top: 0;
+  border-radius: 0 0 6px 6px;
+  background: var(--white);
+}
+
+.office-video-play {
+  width: 38px;
+  height: 34px;
+  padding: 0;
+  color: var(--green);
+  border: 1px solid var(--line);
+  border-radius: 3px;
+  background: var(--paper);
+  cursor: pointer;
+}
+
+.office-video-progress {
+  width: 100%;
+  accent-color: var(--green);
+}
+
+.office-video-time {
+  color: var(--muted);
+  font: 11px/1.2 ui-monospace, SFMono-Regular, Consolas, monospace;
+  white-space: nowrap;
+}
+
+.office-model-note {
+  color: var(--muted);
+  font-size: 12px;
+}
+
+@media (max-width: 700px) {
+  .office-model-grid,
+  .office-video-grid {
+    grid-template-columns: 1fr;
+  }
+
+  .office-video-card-wide {
+    grid-column: auto;
+  }
+
+  .office-video-sync {
+    grid-template-columns: repeat(3, minmax(0, 1fr));
+    gap: 6px;
+  }
+
+  .office-video-sync figcaption {
+    padding: 8px 6px;
+  }
+
+  .office-video-sync figcaption strong {
+    font-size: 11px;
+    overflow-wrap: anywhere;
+  }
+
+  .office-video-sync figcaption span {
+    display: none;
+  }
+}

requirements.txt

@@ -0,0 +1 @@
+playwright>=1.40,<2

scene-compare.js

@@ -0,0 +1,9 @@
+import * as THREE from 'three';import {GLTFLoader} from './vendor/three/loaders/GLTFLoader.js';import {OrbitControls} from './vendor/three/controls/OrbitControls.js';import {RoomEnvironment} from './vendor/three/environments/RoomEnvironment.js';
+const panel=document.querySelector('#figure-3');if(panel){const stage=panel.querySelector('.scene-stage'),select=panel.querySelector('#model-select'),mode=panel.querySelector('#scene-mode'),slider=panel.querySelector('#scene-split'),divider=panel.querySelector('.scene-divider'),status=panel.querySelector('.scene-status'),left=panel.querySelector('.scene-label-left'),right=panel.querySelector('.scene-label-right'),zh=document.documentElement.lang.startsWith('zh');let started=false,active=false,renderer,camera,controls,environment,manifest,selectedScene,truthScene,resizeObserver,split=.5,version=0;const cache=new Map();const setStatus=(text,error=false)=>{status.textContent=text;status.hidden=!text;status.classList.toggle('error',error);panel.dataset.state=error?'error':text?'loading':'ready'};
+function render(){if(!renderer||!selectedScene)return;const w=stage.clientWidth,h=stage.clientHeight;renderer.setScissorTest(false);renderer.setViewport(0,0,w,h);renderer.clear();if(mode.value==='single'||select.value==='GT'||!truthScene){renderer.render(selectedScene,camera);return}const x=Math.round(w*split);renderer.setScissorTest(true);renderer.setScissor(0,0,x,h);renderer.render(selectedScene,camera);renderer.setScissor(x,0,w-x,h);renderer.render(truthScene,camera);renderer.setScissorTest(false)}
+function updateMode(){const comparing=mode.value==='compare'&&select.value!=='GT';panel.classList.toggle('is-comparing',comparing);slider.disabled=!comparing;divider.hidden=!comparing;right.hidden=!comparing;left.textContent=select.value;render()}function updateSplit(value){split=Math.max(.02,Math.min(.98,Number(value)/100));slider.value=String(Math.round(split*100));stage.style.setProperty('--split',`${split*100}%`);divider.setAttribute('aria-valuenow',slider.value);render()}
+function resetCamera(){if(!camera||!manifest)return;const v=manifest.camera;controls.target.fromArray(v.target);camera.position.fromArray(v.position).sub(controls.target).multiplyScalar(Math.max(1,1.5/(stage.clientWidth/stage.clientHeight))).add(controls.target);camera.near=v.near;camera.far=v.far;camera.fov=v.fov;camera.updateProjectionMatrix();controls.update();render()}function resize(){if(!renderer)return;const w=stage.clientWidth,h=stage.clientHeight;renderer.setSize(w,h,false);const oldFit=Math.max(1,1.5/camera.aspect);camera.aspect=w/h;if(controls)camera.position.sub(controls.target).multiplyScalar(Math.max(1,1.5/camera.aspect)/oldFit).add(controls.target);camera.updateProjectionMatrix();render()}
+function loadScene(id){if(!cache.has(id)){const promise=new GLTFLoader().loadAsync(manifest.models[id].url).then(gltf=>{const scene=new THREE.Scene();scene.background=new THREE.Color('#e9ede5');scene.environment=environment;scene.add(new THREE.HemisphereLight(0xffffff,0x68746b,2));const sun=new THREE.DirectionalLight(0xffffff,2.3);sun.position.set(12,25,18);scene.add(sun);let hiddenCutaway=0;if(id!=='GT'){const rows=manifest.models[id].runtime_matrix_gltf_y_up;gltf.scene.applyMatrix4(new THREE.Matrix4().set(...rows.flat()));gltf.scene.traverse(node=>{const name=node.name||'',isShell=/(ceiling|roof|wall|facade)/i.test(name),isDetail=/(emblem|art|grass|metal)/i.test(name);if(isShell&&!isDetail){node.visible=false;hiddenCutaway++}})}scene.add(gltf.scene);const bounds=new THREE.Box3().setFromObject(gltf.scene);scene.userData.method=id;scene.userData.hiddenCutaway=hiddenCutaway;scene.userData.bounds={min:bounds.min.toArray(),max:bounds.max.toArray()};return scene}).catch(e=>{cache.delete(id);throw e});cache.set(id,promise)}return cache.get(id)}
+async function selectScene(){if(!manifest)return;const request=++version,id=select.value;setStatus(zh?`正在加载 ${id}…`:`Loading ${id}…`);try{const [model,gt]=await Promise.all([loadScene(id),loadScene('GT')]);if(request!==version)return;selectedScene=model;truthScene=gt;setStatus('');updateMode();panel.dataset.loadedModel=id;panel.dataset.gtLoaded='true'}catch(e){if(request!==version)return;console.error('Scene asset load failed',e);setStatus(zh?'场景加载失败，请重试。':'Scene loading failed. Retry.',true)}}
+async function start(){if(started)return;started=true;try{setStatus(zh?'正在加载对比场景…':'Loading comparison scenes…');const response=await fetch('data/scene_comparison.json', {cache:'no-store'});if(!response.ok)throw Error(`manifest HTTP ${response.status}`);manifest=await response.json();renderer=new THREE.WebGLRenderer({antialias:true});renderer.setPixelRatio(Math.min(devicePixelRatio,1.5));renderer.toneMapping=THREE.ACESFilmicToneMapping;renderer.outputColorSpace=THREE.SRGBColorSpace;renderer.domElement.setAttribute('aria-label',zh?'重建与 GT 三维场景':'Reconstruction and GT 3D scene');renderer.domElement.tabIndex=0;stage.prepend(renderer.domElement);camera=new THREE.PerspectiveCamera(42,1,.05,500);controls=new OrbitControls(camera,renderer.domElement);controls.enableDamping=false;controls.minDistance=1;controls.maxDistance=200;controls.addEventListener('change',render);const pmrem=new THREE.PMREMGenerator(renderer),room=new RoomEnvironment();environment=pmrem.fromScene(room,.04).texture;room.dispose();pmrem.dispose();resizeObserver=new ResizeObserver(resize);resizeObserver.observe(stage);resize();resetCamera();panel.sceneDiagnostics=()=>({selected:select.value,loaded:selectedScene?.userData.method,gtLoaded:!!truthScene,hiddenCutaway:selectedScene?.userData.hiddenCutaway||0,bounds:selectedScene?.userData.bounds,camera:camera.position.toArray(),target:controls.target.toArray(),split,mode:mode.value,singleCamera:true,singleViewport:true,coordinateConversion:manifest.coordinate_conversion,cachedModels:[...cache.keys()]});await selectScene()}catch(e){console.error('Scene comparison initialization failed',e);resizeObserver?.disconnect();controls?.dispose();environment?.dispose();renderer?.dispose();renderer?.domElement.remove();renderer=camera=controls=environment=manifest=undefined;selectedScene=truthScene=undefined;cache.clear();started=false;setStatus(zh?'无法启动三维场景，请重试。':'Unable to start 3D viewer. Retry.',true)}}
+select.addEventListener('change',()=>{updateMode();selectScene()});mode.addEventListener('change',updateMode);slider.addEventListener('input',()=>updateSplit(slider.value));panel.querySelector('#scene-reset').addEventListener('click',resetCamera);panel.querySelector('#scene-retry').addEventListener('click',()=>manifest&&renderer?selectScene():start());divider.addEventListener('pointerdown',e=>{e.preventDefault();active=true;divider.setPointerCapture(e.pointerId);if(controls)controls.enabled=false});divider.addEventListener('pointermove',e=>{if(!active)return;const b=stage.getBoundingClientRect();updateSplit((e.clientX-b.left)/b.width*100)});const stop=()=>{active=false;if(controls)controls.enabled=true};divider.addEventListener('pointerup',stop);divider.addEventListener('pointercancel',stop);divider.addEventListener('keydown',e=>{const d={ArrowLeft:-2,ArrowRight:2,PageDown:-10,PageUp:10};if(e.key in d){e.preventDefault();updateSplit(split*100+d[e.key])}});updateSplit(50);updateMode();const lazy=new IntersectionObserver(es=>{if(es.some(e=>e.isIntersecting)){lazy.disconnect();start()}},{rootMargin:'300px'});lazy.observe(panel)}

scripts/browser_qa.py

@@ -0,0 +1,6 @@
+#!/usr/bin/env python3
+"""Public entry point for optional browser QA."""
+import runpy
+from pathlib import Path
+
+runpy.run_path(str(Path(__file__).with_name("site_browser_qa.py")), run_name="__main__")

scripts/build.py

@@ -0,0 +1,6 @@
+#!/usr/bin/env python3
+"""Public entry point for the deterministic site builder."""
+import runpy
+from pathlib import Path
+
+runpy.run_path(str(Path(__file__).with_name("render.py")), run_name="__main__")

scripts/editorial.py

@@ -0,0 +1,208 @@
+"""Bilingual narrative and demo additions around the frozen research article."""
+
+from hashlib import sha256
+import json
+from pathlib import Path
+
+SITE_TITLE = "AWSM: Agentic World Simulation and Mapping"
+BASE_URL = "https://phygital-ai.github.io/agentic-world-simulation-and-mapping/"
+BIBTEX = r"""@misc{awsm_2026,
+  title        = {{AWSM}: Agentic World Simulation and Mapping},
+  author       = {{Phygital AI}},
+  year         = {2026},
+  howpublished = {Interactive research article},
+  url          = {https://phygital-ai.github.io/agentic-world-simulation-and-mapping/},
+  note         = {Reconstruction benchmarks, editable scene assets, and map-based embodied demonstration}
+}
+"""
+
+REFERENCES = [
+    ("Qin et al. VINS-Mono: A Robust and Versatile Monocular Visual-Inertial State Estimator. TRO, 2018.", "https://arxiv.org/abs/1708.03852"),
+    ("Keetha et al. MapAnything: Universal Feed-Forward Metric 3D Reconstruction. 2025.", "https://arxiv.org/abs/2509.13414"),
+    ("Zhou et al. Memory Over Maps: 3D Object Localization Without Reconstruction. 2026.", "https://arxiv.org/abs/2603.20530"),
+    ("Yang et al. 3D-Mem: 3D Scene Memory for Embodied Exploration and Reasoning. 2024.", "https://arxiv.org/abs/2411.17735"),
+    ("Khanna et al. GOAT-Bench: A Benchmark for Multi-Modal Lifelong Navigation. CVPR, 2024.", "https://arxiv.org/abs/2404.06609"),
+    ("ConceptGraphs: Open-Vocabulary 3D Scene Graphs for Perception and Planning. ICRA, 2024.", "https://arxiv.org/abs/2309.16650"),
+    ("Clio: Real-time Task-Driven Open-Set 3D Scene Graphs. RA-L, 2024.", "https://arxiv.org/abs/2404.13696"),
+]
+
+
+def reconstruction_reductions():
+    tables = json.loads((Path(__file__).resolve().parents[1] / "data/tables_1_7.json").read_text())["tables"]
+    reductions = []
+    for table, field in (("table3", "bidirectional_mean_m"), ("table2", "model_depth_absrel")):
+        methods = {row["method"]: row[field] for row in tables[table]}
+        reductions.append(100 * (methods["M1"] - methods["M4"]) / methods["M1"])
+    return tuple(reductions)
+
+
+def narrative_copy(copy, en):
+    surface_reduction, depth_reduction = reconstruction_reductions()
+    copy["metric1_reduction"] = f"≈{surface_reduction:.0f}% relative reduction · M1 → M4" if en else f"相对降低约 {surface_reduction:.0f}% · M1 → M4"
+    copy["metric2_reduction"] = f"≈{depth_reduction:.0f}% relative reduction · M1 → M4" if en else f"相对降低约 {depth_reduction:.0f}% · M1 → M4"
+    copy["metric_scope"] = "Single-scene M1 → M4 diagnostic comparison: M1 uses GT-assisted Sim(3) alignment; M4 uses GT camera poses. Relative reductions compare complete pipelines, not IMU alone." if en else "单场景 M1 → M4 诊断性对比：M1 使用真值辅助的 Sim(3) 配准，M4 使用真值相机位姿。相对降幅比较的是完整管线，而非 IMU 的独立贡献。"
+    if en:
+        copy.update({
+            "tldr": f"AWSM uses large-model agents to turn visual observations into editable 3D scenes, grounded in camera-pose and depth estimates—including IMU-informed visual–inertial constraints that tie geometry to physical scale. In one controlled Isaac Sim scene, four reconstruction routes test spatial fidelity, not just visual plausibility. The aligned comparison from RGB-only M1 to GT-pose-conditioned M4 shows ≈{surface_reduction:.0f}% lower bidirectional surface error and ≈{depth_reduction:.0f}% lower model-depth AbsRel across 180 evaluation views. A multi-robot demo then illustrates goal-directed navigation and task execution on the reconstructed map: follow predefined routes to a marked reception desk and line up there. Next steps are higher efficiency and accuracy, reusable simulation-ready assets, and deeper integration with phygital agents for persistent spatial memory, object search, and interaction.",
+            "intro1": "Agentic reconstruction uses a large model to inspect observations, call modeling tools, and iteratively build and check an editable scene. The challenge is not only to produce a convincing room, but to preserve the scale, shape, and spatial relationships that make it useful beyond rendering. AWSM asks how geometric evidence can constrain this process so that its outputs become more faithful spatial references for downstream agents.",
+            "motivation_title": "From plausible reconstruction to a usable spatial reference",
+            "motivation1": "A reconstruction agent can assemble a plausible room while getting a corner, distance, or passage wrong. For an embodied agent, those are not merely visual defects: they change where a destination lies and which route a body can follow. Geometry grounding means constraining the modeling process with pose, depth, and metric evidence where available, rather than relying on visual plausibility alone. The question is how those constraints survive the conversion from observations into editable objects.",
+            "motivation2": "The reception-desk task makes this connection concrete: mark a destination in the reconstructed map, specify routes, and execute them with robot controllers. We use ‘scene representation’ for the editable objects and geometry, ‘map’ for their navigation role, and ‘simulation assets’ for their integration into a simulator. The ‘world’ in AWSM refers to this reusable spatial environment, not a learned dynamics predictor. A persistent spatial foundation connecting navigation, memory, interaction, and simulation is the longer-term goal; the experiments and demo examine its geometric basis and one downstream use.",
+            "contributions": [
+                "A four-route study of geometry-grounded agentic reconstruction, keeping the scene, 180 RGB identities, modeling objective, and editable output format fixed while comparing complete pipelines with different geometric evidence.",
+                "An evaluation protocol that separates trajectory error, native depth, final-scene depth, surface geometry, and rendered appearance, with frozen Blend/GLB assets, reproducible evaluation subsets, public hashes, and interactive comparisons.",
+                "A map-based multi-robot simulation example connecting a reconstructed scene to destination marking, predefined routes, and controller execution. It illustrates downstream use rather than a controlled navigation comparison across reconstruction routes.",
+            ],
+            "method_title": "One agentic workflow, four reconstruction routes",
+            "interactive_title": "Inspect the reconstructed scenes, not just aggregate scores",
+            "discussion_title": "What geometric grounding changes—and what it does not",
+            "positioning": "AWSM connects two roles: an agent that constructs an editable scene, and downstream agents that use it as a spatial reference. SLAM and visual-inertial estimation supply pose and metric constraints; ViPE and Depth Anything 3 contribute geometric observations. Concurrent work AHa-3D explores video-driven, tool-based Real2Sim, including camera and 3D-reference estimation from video. Our emphasis is geometric grounding: bringing additional physical measurements, especially IMU-informed visual-inertial pose and metric constraints, into agentic scene construction rather than relying on video-derived evidence alone. The four-route comparison tests reconstruction systems, not the isolated effect of IMU; the demo illustrates a map-based application, not a complete autonomous embodied system.",
+        })
+        copy["method_intro"] = "The reconstruction agent follows a shared observe–build–verify loop: inspect the available evidence, write Blender Python to construct objects, render review views, and revise the scene before freezing it for evaluation. This tool-using modeling process is what ‘agentic’ describes here; the downstream robot controllers have a separate role."
+        copy["limits"].extend([
+            "The reception-desk demo is map-based simulation execution, not a controlled comparison of navigation across M1–M4. Routes are predefined and localization uses simulator pose; see the demo protocol for scope.",
+            "Multimodal memory, long-term map maintenance, and reduced policy sim-to-real gap are research goals, not measured outcomes of the four-robot demo. The Office Café corner observation has no matched-view quantitative GT measurement here.",
+            "Historical downstream documentation describes the same M4 Blend hash using a different depth frontend. The frozen tables and assets are retained unchanged; their upstream naming needs reconciliation before attributing historical task outcomes to DA3. See the provenance note.",
+        ])
+        copy["office1"] += " An author-observed failure—a curved real-world corner simplified into a square one—motivates separating local shape from global scale. A correct scale alone cannot repair that shape error; this is a qualitative observation, not another measured benchmark."
+    else:
+        copy.update({
+            "tldr": f"AWSM 让大模型智能体将视觉观测转化为可编辑的 3D 场景，并以估计的相机位姿与深度进行几何锚定，其中包括融合 IMU 的视觉惯性约束，将几何与真实物理尺度绑定。在一个受控的 Isaac Sim 场景中，四条重建路线检验的不只是视觉合理性，更是空间忠实度：经配准的纯 RGB 路线 M1 与真值位姿条件路线 M4 对比，双向表面误差降低约 {surface_reduction:.0f}%，180 个评估视角上的模型深度绝对相对误差（AbsRel）降低约 {depth_reduction:.0f}%。多机器人 demo 进一步展示基于重建地图的目标导航与任务执行：沿预设路线前往地图上已标记的前台，并在那里排队。下一步是提升效率与精度、形成可复用的仿真就绪资产，并进一步与虚实融合智能体（phygital agents）集成，支持持久空间记忆、寻物与交互。",
+            "intro1": "Agentic 重建由大模型检查观测、调用建模工具，并迭代构建和验证可编辑场景。挑战不只是生成一个令人信服的房间，还在于保留尺度、形状与空间关系，让结果在渲染之外仍然有用。AWSM 研究如何用几何证据约束这一过程，让输出成为下游智能体更忠实的空间参照。",
+            "motivation_title": "从看起来合理的重建，到可以使用的空间参照",
+            "motivation1": "负责重建的智能体可以搭建一个看似合理的房间，却重建错转角、距离或通道。对具身智能体而言，这些不只是视觉缺陷，还会改变目标的位置与身体可以通过的路线。几何锚定，是在信息可用时以位姿、深度与米制证据约束建模过程，而非仅依赖视觉合理性。关键问题是：这些约束如何在观测转化为可编辑对象的过程中得到保留？",
+            "motivation2": "前台任务让这种联系变得具体：在重建地图中标记目标、设定路线，再由机器人控制器执行。本文用“场景表示”指可编辑的对象与几何，用“地图”描述它在导航中的作用，用“仿真资产”描述它与仿真器的集成。AWSM 中的“世界”指这种可复用的空间环境，而非学习得到的动力学预测模型。连接导航、记忆、交互与仿真的持久空间基础是长期目标；当前实验和 demo 分别研究其几何基础，并展示一种下游用途。",
+            "contributions": [
+                "对几何锚定的 agentic 重建开展四路线研究：固定场景、同一组 180 帧 RGB 输入、建模目标与可编辑输出格式，比较具有不同几何证据的完整管线。",
+                "将轨迹误差、原生深度、最终场景深度、表面几何和渲染外观分开评测，并提供冻结 Blend/GLB 资产、可复现评测子集、公开哈希与交互式对照。",
+                "通过基于地图的多机器人仿真示例，将重建场景连接到目标标记、预设路线和控制器执行；它展示下游用途，而非不同重建路线之间的受控导航对比。",
+            ],
+            "method_title": "一个 agentic 工作流，四条重建路线",
+            "interactive_title": "不只看汇总指标，也检查重建场景",
+            "discussion_title": "几何锚定改变了什么，又没有解决什么",
+            "positioning": "AWSM 连接两类角色：负责构建可编辑场景的智能体，以及将它作为空间参照使用的下游智能体。SLAM 与视觉惯性估计提供位姿和米制约束，ViPE 与 Depth Anything 3 提供几何观测。同期工作 AHa-3D 探索视频驱动、工具辅助的 Real2Sim，包括从视频估计相机与三维参照。我们的重点是几何锚定：将额外的物理测量，尤其是融合 IMU 的视觉惯性位姿与米制约束，引入智能体场景构建，而不只依赖视频推断的证据。四路线对比评估的是重建系统，并非 IMU 独立作用的消融实验；demo 展示的是地图应用，而不是完整的自主具身系统。",
+        })
+        copy["method_intro"] = "负责重建的智能体遵循共同的“观察—构建—验证”循环：检查可用证据，编写 Blender Python 构建对象，渲染检查视图，并在冻结评测前迭代修改。这里的 agentic 指这种工具驱动的建模过程；下游机器人控制器承担另一种角色。"
+        copy["limits"].extend([
+            "前台 demo 展示基于地图的仿真执行，不是 M1–M4 的受控导航对比。路线预先设定，定位使用仿真器位姿；具体范围见演示条件。",
+            "多模态空间记忆、长期地图维护与策略 sim-to-real 差距的降低是研究目标，不是这段四机器人视频已量化的结果。本页也没有 Office Café 拐角的同视角 GT 定量测量。",
+            "历史下游文档对同一 M4 Blend 哈希使用了不同的深度前端描述。本页冻结表格与资产保持不变；将旧任务收益归因于 DA3 之前，需要核对上游方法命名，详见来源说明。",
+        ])
+        copy["office1"] += " 作者观察到的一类失败是：真实的弯曲转角被简化成方正结构。它提醒我们区分局部形状与全局尺度——尺度正确并不能自动修复形状错误；这里将其作为定性观察，而非另一组定量结果。"
+
+
+def demo_section(en):
+    title = "A reconstructed map. A destination. Robots in motion." if en else "一张重建地图，一个目标，一次具身执行。"
+    lead = (
+        "The demo connects reconstruction to downstream use in NVIDIA Isaac Sim. The reception desk is marked in the navigation map, routes are specified, and a drone, humanoid, quadruped, and wheeled robot execute the task. The same spatial reference connects a destination, route constraints, and motion control: a scene to inspect becomes a map to act with."
+        if en else
+        "这个 demo 在 NVIDIA Isaac Sim 中将重建连接到下游使用：在导航地图中标出前台，设定路线，再让无人机、人形、四足和轮式机器人执行任务。同一个空间参照连接目标、路径约束与运动控制：从可以查看的重建场景，走向可以用于行动的地图。"
+    )
+    command = "Send the robots to the reception desk and have them line up there." if en else "让机器人前往前台，并在那里排成一列。"
+    labels = ("Task instruction", "Locate the reception desk", "Set routes on the map", "Execute with controllers") if en else ("任务指令", "地图上定位前台", "设定导航路线", "控制器执行")
+    steps = "".join(f'<li><span>{index:02d}</span>{label}</li>' for index, label in enumerate(labels, 1))
+    caption = (
+        "Map-based navigation in simulation · predefined routes · simulator-pose feedback. The authors confirm that the demonstration has been manually reviewed."
+        if en else
+        "基于重建地图的仿真导航 · 预设路线 · 仿真器位姿反馈。作者确认，本演示已完成人工审阅。"
+    )
+    details = (
+        '<p>The instruction above describes the intended task, not a demonstrated language-to-plan parser. Routes were manually specified and screened against both the M4 reconstructed map and original geometry. This is not a reconstruction-only planning benchmark.</p>'
+        '<p>The 156-second presentation edit plays at 1×: Scene (0–6s), Reconstruction (6–12s), Planned route (12–18s), and Navigation (18–156s). Navigation uses the new run <code>20260930-214924</code>; the opening 12.32 seconds of waiting are omitted, and the later terminal hold is outside this cut. The map is a static route design, not a measured execution trace.</p>'
+        '<p>This demo does not evaluate online visual localization, learned instruction understanding, memory retrieval, or real-robot transfer. Those are separate capabilities; the map-based execution shown here remains the intended demonstration.</p>'
+        if en else
+        '<p>上方指令说明任务意图，不代表本次演示实现了语言到计划的自动解析。路线人工设定，并同时参考 M4 重建地图和原始几何筛选；这不是只依赖重建图的规划 benchmark。</p>'
+        '<p>成片为 156 秒、1× 原速：场景（0–6秒）、重建（6–12秒）、路线设计（12–18秒）、导航（18–156秒）。导航来自新运行 <code>20260930-214924</code>，略去开头等待的 12.32 秒，后续终点保持过程不在本剪辑中。地图是静态路线设计图，不是实测执行轨迹。</p>'
+        '<p>本 demo 不评测在线视觉定位、学习型指令理解、记忆检索或真实机器人迁移。这些是独立能力，不改变本演示所表达的“利用重建地图执行导航任务”。</p>'
+    )
+    return f'''<section id="embodied-demo" class="embodied-demo">
+<p class="section-tag">DEMO / MAP-BASED EMBODIED EXECUTION</p><h2>{title}</h2><p>{lead}</p>
+<blockquote class="mission-command"><span>{"Task intent: " if en else "任务意图："}</span>{command}</blockquote>
+<ol class="mission-flow" aria-label="{"Task workflow" if en else "任务流程"}">{steps}</ol>
+<figure class="demo-film"><video controls playsinline preload="none" poster="assets/embodied-demo/navigation-214924-poster.png" width="1280" height="720" aria-label="{"Four-robot map-based navigation demonstration" if en else "四机器人地图导航演示"}"><source src="assets/embodied-demo/four-robots-214924-156s-phone.mp4" type="video/mp4">{"Your browser cannot play this video. Use the download link below." if en else "浏览器无法播放此视频，请使用下方下载链接。"}</video><figcaption>{caption}</figcaption></figure>
+<div class="demo-links"><a href="assets/embodied-demo/four-robots-214924-156s-phone.mp4" download>{"Download mobile edition" if en else "下载手机版"} · 11.65 MB · 2:36</a><a href="assets/embodied-demo/four-robots-214924-156s-hd.mp4" download>{"Download HD edition" if en else "下载高清版"} · 49.69 MB · 2:36</a><a href="data/embodied_demo.json">{"Run &amp; media record" if en else "运行与视频记录"} ↗</a></div>
+<figure class="demo-map" id="demo-route-map"><a href="assets/embodied-demo/planned-route-156s.png" target="_blank" rel="noopener"><img loading="lazy" src="assets/embodied-demo/planned-route-156s.png" width="1920" height="1080" alt="{"Planned routes for four robots, with the reception desk and waypoints marked; static design, not execution trajectories" if en else "标出前台与航点的四机器人路线设计图；静态设计，非执行轨迹"}"></a><figcaption><span>{"Navigation map." if en else "导航地图。"}</span> {"The reception desk and planned routes share the reconstructed scene. This is a static design preview, not a recorded trajectory. Open the image to inspect the full-resolution map." if en else "在重建场景中标出前台与规划路线。这是静态设计图，而非实跑轨迹；点击查看原尺寸地图。"}</figcaption></figure>
+<details class="demo-protocol"><summary>{"Demo protocol and scope" if en else "演示条件与范围"}</summary>{details}<p>{"Manual presentation review is separate from the historical automated checks retained below; it does not revise their recorded verdicts." if en else "人工展示审阅与下方保留的历史自动检查属于不同验证层次，不改变其原始判定。"}</p><p><a href="evidence/embodied-demo-audit.json">{"Historical automated check record" if en else "历史自动检查记录"} ↗</a></p></details>
+</section>'''
+
+
+def grounding_section(en):
+    if en:
+        return '''<div id="scale-and-sensors" class="analysis-block grounding-block"><h3>Geometry grounding: scale is necessary, but not sufficient</h3>
+<p>Monocular projection geometry leaves a global scale ambiguity; learned metric depth adds a useful prior, not an independent measurement of every real-world dimension. Calibrated RGB-D, a known baseline or length, and visual-inertial estimation can supply metric constraints. IMU fusion can recover metric motion under sufficient excitation and reliable initialization, bias estimation, synchronization, and calibration. An IMU is not a direct distance sensor, and its presence alone does not guarantee accurate scale.</p>
+<p>Scale, local shape, camera pose, and navigable connectivity must be checked separately. A curved corner simplified into a square cannot be repaired by rescaling alone. Pose-conditioned depth can improve reconstruction, but neither the DA3 ablation nor our M2/M3 results imply that every depth metric improves monotonically. The four routes below are system comparisons, not an isolated IMU ablation.</p>
+<p class="inline-sources"><a href="https://arxiv.org/abs/1708.03852" target="_blank" rel="noopener">VINS-Mono</a> · <a href="https://arxiv.org/html/2511.10647v1#S7.SS2.SSS3" target="_blank" rel="noopener">DA3 pose-conditioning ablation</a></p></div>'''
+    return '''<div id="scale-and-sensors" class="analysis-block grounding-block"><h3>几何锚定：真实尺度是必要约束，但不是全部答案</h3>
+<p>纯单目投影几何存在全局尺度歧义；学习型米制深度提供有用先验，却不是对每个真实尺寸的独立测量。标定 RGB-D、已知基线或长度，以及视觉惯性估计，都可以提供米制约束。IMU 融合能在运动激励充分、初始化、偏置估计、时间同步与标定可靠时恢复米制运动；IMU 不是直接测距传感器，“带 IMU”不等于“尺度一定准确”。</p>
+<p>尺度、局部形状、相机位姿和可通行连通性需要分别检查。弯曲转角被简化成直角，不能只靠缩放修复。位姿条件可以改善深度重建，但 DA3 消融和本文 M2/M3 的结果都不支持“所有深度指标必然同步改善”；下文四条路线是完整系统比较，而非只切换 IMU 的单变量实验。</p>
+<p class="inline-sources"><a href="https://arxiv.org/abs/1708.03852" target="_blank" rel="noopener">VINS-Mono</a> · <a href="https://arxiv.org/html/2511.10647v1#S7.SS2.SSS3" target="_blank" rel="noopener">DA3 位姿条件消融</a></p></div>'''
+
+
+def outlook_section(en):
+    if en:
+        title = "Toward persistent spaces for phygital agents"
+        intro = "The current outputs are editable scene snapshots, and the demo shows a scene-specific simulation integration. The longer-term goal is to make geometry-grounded agentic reconstruction a reusable process for building and maintaining spaces that phygital agents can use. Saving an editable asset is a starting point, not a demonstration of lifelong map maintenance. The directions below extend the present evidence."
+        cards = [
+            ("Navigate", "Use destinations, free space, and body-specific clearance in a shared geometric frame. Better reconstruction can reduce map mismatch; reliable navigation also needs localization and current observations."),
+            ("Remember & retrieve", "Link object identities and locations to source views so language, images, and past observations can refer to the same place. Retrieval and safe arrival should be evaluated separately."),
+            ("Interact & maintain", "Edit object properties and relationships, revisit changed areas, and preserve evidence and versions. Persistent memory must distinguish observed geometry, inferred completion, and unknown space."),
+            ("Simulate & learn", "Move from the current scene-specific integration to reusable simulation-ready exports with checked scale, collision geometry, and physical parameters. These can support task rehearsal, scenario variation, and training; reduced policy sim-to-real gap remains a separate question to test."),
+        ]
+        closing = "A full mesh is not required for every retrieval task: Memory Over Maps uses posed RGB-D keyframes for on-demand localization, while 3D-Mem and task-oriented scene graphs explore complementary memory and representation choices. Our proposed direction is hybrid: preserve visual evidence, refine task-relevant geometry, and build editable simulation assets where they add value. The next question is when an agent has enough evidence to act—and when it should observe again and update its map."
+        next_steps = "The immediate agenda is higher reconstruction efficiency and accuracy, reusable simulation-ready scene generation, and deeper integration with phygital agents. Together, these steps connect the reconstruction agent’s modeling and revision loop to the downstream agent’s navigation, memory, and interaction needs."
+        label = "RESEARCH DIRECTION"
+    else:
+        title = "走向虚实融合智能体可持续使用的空间"
+        intro = "当前输出是可编辑的场景快照，demo 展示了针对特定场景的仿真集成。长期目标是把几何锚定的 agentic 重建发展为可复用流程，构建并维护虚实融合智能体能够使用的空间。保存可编辑资产是起点，不等于已经验证长期地图维护；以下方向是在当前证据基础上的延伸。"
+        cards = [
+            ("导航", "在同一几何坐标中表达目标、自由空间与不同机器人的通行净空。更忠实的重建可以减少地图失配，可靠导航还需要定位与实时观测。"),
+            ("记忆与多模态检索", "将对象身份、位置和源视图关联，让语言、图像与过去的观察指向同一个地点。检索到对象与安全到达对象，需要分别评测。"),
+            ("交互与长期维护", "编辑对象属性与关系，重访变化区域，保留证据和版本。持久记忆应区分实际观察、推断补全和未知空间，而非把所有生成内容都当作事实。"),
+            ("仿真与学习", "从当前特定场景的集成，走向可复用的仿真就绪导出流程，检查尺度、碰撞几何与物理参数。它可支持任务预演、场景变体和训练；是否降低策略 sim-to-real 差距，仍需单独检验。"),
+        ]
+        closing = "并非每次寻物都需要完整网格：Memory Over Maps 用带位姿 RGB-D 关键帧按需定位，3D-Mem 与任务相关场景图也探索了不同记忆和表示方式。我们主张混合路线：保留视觉证据，细化任务相关几何，在有价值的地方构建可编辑仿真资产。下一步的关键问题是：什么时候证据已经足够支持行动，什么时候 agent 应重新观察并更新地图？"
+        next_steps = "近期重点是提升重建效率与精度、形成可复用的仿真就绪场景生成流程，以及进一步与虚实融合智能体集成，让重建智能体的建模与修正循环连接到下游智能体的导航、记忆和交互需求。"
+        label = "研究方向"
+    items = "".join(f'<div><h4>{name}</h4><p>{text}</p></div>' for name, text in cards)
+    return f'''<div id="spatial-foundation" class="analysis-block spatial-foundation"><p class="section-tag">{label}</p><h3>{title}</h3><p>{intro}</p><div class="capability-grid">{items}</div><p>{closing}</p><p>{next_steps}</p><p class="inline-sources"><a href="https://arxiv.org/abs/2603.20530" target="_blank" rel="noopener">Memory Over Maps</a> · <a href="https://arxiv.org/abs/2411.17735" target="_blank" rel="noopener">3D-Mem</a> · <a href="https://arxiv.org/abs/2404.13696" target="_blank" rel="noopener">Clio</a> · <a href="https://arxiv.org/abs/2404.06609" target="_blank" rel="noopener">GOAT-Bench</a></p></div>'''
+
+
+def editorial_page(page, en):
+    page = page.replace('01 / RESULTS', '01 / OVERVIEW')
+    page = page.replace('04 / MORE RESULTS AND ANALYSIS', '04 / RESULTS AND ANALYSIS')
+    style_version = sha256((Path(__file__).resolve().parents[1] / "editorial.css").read_bytes()).hexdigest()[:12]
+    page = page.replace('<link rel="stylesheet" href="style.css">', f'<link rel="stylesheet" href="style.css"><link rel="stylesheet" href="editorial.css?v={style_version}">')
+    base_style_version = sha256((Path(__file__).resolve().parents[1] / "style.css").read_bytes()).hexdigest()[:12]
+    page = page.replace('href="style.css"', f'href="style.css?v={base_style_version}"')
+    office_version = sha256((Path(__file__).resolve().parents[1] / "office.css").read_bytes()).hexdigest()[:12]
+    page = page.replace('</head>', f'<link rel="stylesheet" href="office.css?v={office_version}"></head>')
+    for asset in ("app.js", "scene-compare.js", "office-models.js"):
+        version = sha256((Path(__file__).resolve().parents[1] / asset).read_bytes()).hexdigest()[:12]
+        page = page.replace(f'src="{asset}"', f'src="{asset}?v={version}"')
+    if en:
+        page = page.replace(f'<link rel="canonical" href="{BASE_URL}index.html">', f'<link rel="canonical" href="{BASE_URL}">')
+    page = page.replace(f'hreflang="zh-CN" href="{BASE_URL}index.html"', f'hreflang="zh-CN" href="{BASE_URL}zh.html"')
+    page = page.replace(f'hreflang="en" href="{BASE_URL}en.html"', f'hreflang="en" href="{BASE_URL}"')
+    page = page.replace('</head>', f'<link rel="alternate" hreflang="x-default" href="{BASE_URL}"></head>')
+    page = page.replace('<a href="index.html" ', '<a href="zh.html" ', 1)
+    page = page.replace('<a href="en.html" ', '<a href="index.html" ', 1)
+    page = page.replace('<a class="brand" href="index.html">ASTRA / WORLD MODELS</a>', '<a class="brand" href="index.html">PHYGITAL AI / AWSM</a>')
+    subtitle = 'From real spaces to <strong class="tagline-emphasis">worlds phygital agents can use.</strong>' if en else '把真实空间，变成<strong class="tagline-emphasis">虚实融合智能体可以使用的世界。</strong>'
+    pronunciation = 'pronounced “awesome”' if en else '读作 “awesome”'
+    brand_note = 'Phygital = physical + digital.' if en else 'Phygital = physical（物理）+ digital（数字），即虚实融合。'
+    before_header, header_start, remainder = page.partition('<header>')
+    old_header, header_end, after_header = remainder.partition('</header>')
+    title_name = "Agentic World Simulation and Mapping" if en else "智能体世界仿真与建图"
+    heading = f'<p class="title-name">{title_name}</p><div class="title-lockup"><h1 aria-label="AWSM: {title_name}">AWSM</h1><p class="pronunciation">{pronunciation}</p></div>'
+    publication = 'Published <time datetime="2026-10-01">October 1, 2026</time>' if en else '发布于 <time datetime="2026-10-01">2026年10月1日</time>'
+    page = before_header + header_start + heading + f'<p class="lead">{subtitle}</p><p class="brand-note">{brand_note}</p><p class="publication-date">{publication}</p>' + header_end + after_header
+    page = page.replace('</head>', '<meta property="article:published_time" content="2026-10-01"></head>')
+    download_label = "Download BibTeX" if en else "下载 BibTeX 引用"
+    page = page.replace('</code></pre></section>', f'</code></pre><a class="citation-download" href="data/awsm.bib" download="awsm.bib">{download_label} <span aria-hidden="true">↓</span></a></section>', 1)
+    page = page.replace('<section id="related-work">', demo_section(en) + '\n<section id="related-work">', 1)
+    page = page.replace('</section>\n<section id="workflow">', grounding_section(en) + '</section>\n<section id="workflow">', 1)
+    page = page.replace('</section>\n<section id="limitations">', outlook_section(en) + '</section>\n<section id="limitations">', 1)
+    page = page.replace('href="#related-work">', 'href="#embodied-demo">' + ('Embodied demo' if en else '具身演示') + '</a><a href="#related-work">', 1)
+    evidence_link = '<p><a href="evidence/provenance-note.md">' + ('Version and provenance note' if en else '版本与资产来源说明') + ' ↗</a></p>'
+    page = page.replace('</section>\n<section id="citation">', evidence_link + '</section>\n<section id="citation">', 1)
+    return page.replace('Agentic World · frozen evidence, editable outputs', 'Phygital AI · AWSM · frozen evidence, editable outputs')

scripts/publish.py

@@ -0,0 +1,96 @@
+#!/usr/bin/env python3
+"""Publish a validated, committed site to the Phygital AI organization."""
+
+import json
+import os
+import subprocess
+import sys
+from pathlib import Path
+
+
+ROOT = Path(__file__).resolve().parents[1]
+OWNER = "Phygital-AI"
+REPO = "agentic-world-simulation-and-mapping"
+REMOTE_NAME = "phygital"
+REMOTE_URL = f"https://github.com/{OWNER}/{REPO}.git"
+SITE_URL = f"https://{OWNER.lower()}.github.io/{REPO}/"
+
+
+def git(*args):
+    return subprocess.run(["git", "-C", str(ROOT), *args], check=True, capture_output=True, text=True).stdout.strip()
+
+
+def api(path, method="GET", body=None):
+    command = ["gh", "api", path, "--method", method]
+    if body is not None:
+        command.extend(["--input", "-"])
+    result = subprocess.run(command, input=json.dumps(body) if body is not None else None, capture_output=True, text=True)
+    try:
+        response = json.loads(result.stdout) if result.stdout.strip() else {}
+    except json.JSONDecodeError:
+        raise SystemExit(f"GitHub API returned an invalid response for {path}")
+    return result.returncode, response
+
+
+def require_api(path, method="GET", body=None):
+    code, response = api(path, method, body)
+    if code:
+        raise SystemExit(f"GitHub API failed for {path}: {response.get('message', 'request failed')}")
+    return response
+
+
+def main():
+    subprocess.run([sys.executable, str(ROOT / "scripts/render.py"), "--check"], check=True)
+    subprocess.run([sys.executable, str(ROOT / "scripts/validate.py")], check=True)
+    if git("status", "--porcelain"):
+        raise SystemExit("Commit the validated revision before publishing.")
+    user = require_api("user")
+    if user.get("login") != "Yaofang-Liu":
+        raise SystemExit("Authenticate GitHub CLI as Yaofang-Liu before publishing.")
+    path = f"repos/{OWNER}/{REPO}"
+    code, repository = api(path)
+    if code and str(repository.get("status")) == "404":
+        repository = require_api(f"orgs/{OWNER}/repos", "POST", {
+            "name": REPO,
+            "description": "Geometry-grounded agentic scene reconstruction, map-based embodied execution, and persistent spatial representations for memory, interaction, and simulation.",
+            "private": False,
+            "auto_init": False,
+            "homepage": SITE_URL,
+        })
+    elif code:
+        raise SystemExit("Cannot inspect the intended organization repository.")
+    if repository.get("full_name") != f"{OWNER}/{REPO}" or repository.get("private") or not repository.get("permissions", {}).get("push"):
+        raise SystemExit("Repository identity, visibility, or push permissions do not match the publication target.")
+    if REMOTE_NAME not in git("remote").splitlines():
+        git("remote", "add", REMOTE_NAME, REMOTE_URL)
+    if git("remote", "get-url", REMOTE_NAME) != REMOTE_URL:
+        raise SystemExit("Unexpected publication remote; no push attempted.")
+    push = subprocess.run([
+        "git", "-C", str(ROOT), "-c", "credential.helper=", "-c", "credential.helper=!gh auth git-credential",
+        "push", "-u", REMOTE_NAME, "HEAD:main",
+    ], capture_output=True, text=True, env={**os.environ, "GIT_TERMINAL_PROMPT": "0"})
+    if push.returncode:
+        raise SystemExit("Non-force push failed. Inspect branch state and GitHub CLI authentication; no forced overwrite was attempted.")
+    require_api(path, "PATCH", {"homepage": SITE_URL, "default_branch": "main"})
+    code, pages = api(path + "/pages")
+    configuration = {"build_type": "legacy", "source": {"branch": "main", "path": "/"}}
+    if code and str(pages.get("status")) == "404":
+        require_api(path + "/pages", "POST", configuration)
+    elif code:
+        raise SystemExit("Code pushed, but GitHub Pages configuration could not be inspected.")
+    elif pages.get("build_type") != "legacy" or pages.get("source") != configuration["source"]:
+        require_api(path + "/pages", "PUT", configuration)
+    build_code, build = api(path + "/pages/builds", "POST", {})
+    record = {
+        "repository": REMOTE_URL.removesuffix(".git"),
+        "commit": git("rev-parse", "HEAD"),
+        "url": SITE_URL,
+        "pages_build_requested": build_code == 0,
+        "pages_build_url": build.get("url"),
+    }
+    (ROOT / ".publication.json").write_text(json.dumps(record, indent=2) + "\n", encoding="utf-8")
+    print(json.dumps(record))
+
+
+if __name__ == "__main__":
+    main()

scripts/render.py

@@ -0,0 +1,82 @@
+#!/usr/bin/env python3
+"""Render narrative-only updates without rebuilding any frozen experiment asset."""
+
+import argparse
+import hashlib
+import json
+from pathlib import Path
+
+from site_builder_interactive import academic_page
+from editorial import BIBTEX
+
+
+ROOT = Path(__file__).resolve().parents[1]
+PUBLIC_ROOT_FILES = (
+    ".nojekyll",
+    "index.html",
+    "en.html",
+    "zh.html",
+    "style.css",
+    "editorial.css",
+    "office.css",
+    "app.js",
+    "scene-compare.js",
+    "office-models.js",
+)
+PUBLIC_DIRECTORIES = ("assets", "data", "evidence", "models", "vendor")
+GENERATED_MANIFESTS = ("evidence/publication_manifest.json", "evidence/SHA256SUMS")
+
+
+def sha256(path):
+    return hashlib.sha256(path.read_bytes()).hexdigest()
+
+
+def publication_records():
+    paths = [ROOT / name for name in PUBLIC_ROOT_FILES]
+    for directory in PUBLIC_DIRECTORIES:
+        paths.extend(path for path in (ROOT / directory).rglob("*") if path.is_file())
+    return [
+        {"path": path.relative_to(ROOT).as_posix(), "sha256": sha256(path), "bytes": path.stat().st_size}
+        for path in sorted(paths)
+        if path.relative_to(ROOT).as_posix() not in GENERATED_MANIFESTS
+    ]
+
+
+def main():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--check", action="store_true", help="Verify generated pages and hashes without writing")
+    args = parser.parse_args()
+    for name, english in (("index.html", True), ("en.html", True), ("zh.html", False)):
+        content = academic_page(english) + "\n"
+        target = ROOT / name
+        if args.check:
+            if target.read_text(encoding="utf-8") != content:
+                raise SystemExit(f"Stale generated page: {name}")
+        else:
+            target.write_text(content, encoding="utf-8")
+    citation_path = ROOT / "data/awsm.bib"
+    if args.check:
+        if citation_path.read_text(encoding="utf-8") != BIBTEX:
+            raise SystemExit("Stale downloadable citation")
+    else:
+        citation_path.write_text(BIBTEX, encoding="utf-8")
+    records = publication_records()
+    manifest = json.dumps({"schema_version": 2, "source_contract": "data/source_contract.json", "files": records}, ensure_ascii=False, indent=2) + "\n"
+    manifest_path = ROOT / GENERATED_MANIFESTS[0]
+    if args.check:
+        if manifest_path.read_text(encoding="utf-8") != manifest:
+            raise SystemExit("Stale publication manifest")
+    else:
+        manifest_path.write_text(manifest, encoding="utf-8")
+    sums = "\n".join(f'{record["sha256"]}  {record["path"]}' for record in records)
+    sums += f"\n{sha256(manifest_path)}  {GENERATED_MANIFESTS[0]}\n"
+    if args.check:
+        if (ROOT / GENERATED_MANIFESTS[1]).read_text(encoding="utf-8") != sums:
+            raise SystemExit("Stale SHA256SUMS")
+    else:
+        (ROOT / GENERATED_MANIFESTS[1]).write_text(sums, encoding="utf-8")
+    print(json.dumps({"status": "CHECKED" if args.check else "RENDERED", "pages": 3, "published_files": len(records), "frozen_experiment_assets_rebuilt": False}))
+
+
+if __name__ == "__main__":
+    main()

scripts/site_browser_qa.py

@@ -0,0 +1,6 @@
+#!/usr/bin/env python3
+"""Run desktop/mobile browser QA for both interactive figures."""
+import runpy
+from pathlib import Path
+
+runpy.run_path(str(Path(__file__).with_name("site_browser_qa_interactive.py")), run_name="__main__")

scripts/site_browser_qa_interactive.py

@@ -0,0 +1,188 @@
+#!/usr/bin/env python3
+"""Exercise both interactive figures on desktop/mobile and zh/en."""
+from pathlib import Path
+import argparse, json, tempfile
+try:
+    from playwright.sync_api import sync_playwright
+except ImportError:
+    print(json.dumps({"status":"SKIP","reason":"Playwright is not installed; run: pip install playwright && playwright install chromium"}))
+    raise SystemExit(0)
+
+ap=argparse.ArgumentParser(); ap.add_argument("--url",default="http://127.0.0.1:8765/")
+ap.add_argument("--quick",action="store_true",help="Check representative models and fixed views across all four language/viewport combinations")
+ap.add_argument("--output",type=Path,default=None)
+ap.add_argument("--executable-path",type=Path,default=None); a=ap.parse_args()
+out=a.output or Path(tempfile.mkdtemp(prefix="agentic-world-blog-qa-")); out.mkdir(parents=True,exist_ok=True)
+root=Path(__file__).resolve().parents[1]
+demo_metadata=json.loads((root/"data/embodied_demo.json").read_text(encoding="utf-8"))
+demo_media=demo_metadata["media"]
+office_origin="https://office-cafe-vipe.hiwtishere.chatgpt.site"
+results=[]
+with sync_playwright() as pw:
+    try: browser=pw.chromium.launch(channel="chromium",headless=True,executable_path=a.executable_path)
+    except Exception as e:
+        print(json.dumps({"status":"SKIP","reason":f"Playwright browser unavailable: {e}"})); raise SystemExit(0)
+    for viewport,size in (("desktop",{"width":1440,"height":1000}),("mobile",{"width":390,"height":844})):
+        for lang,path in (("zh","zh.html"),("en","index.html")):
+            page=browser.new_page(viewport=size); errors=[]; failed=[]
+            page.route(office_origin+"/**",lambda route: route.fulfill(
+                status=200,
+                content_type="text/html",
+                body="<!doctype html><title>External not tested</title><p>external-not-tested</p>",
+            ))
+            page.on("console",lambda m: errors.append(f"console {m.type}: {m.text}") if m.type=="error" else None)
+            page.on("pageerror",lambda e: errors.append(str(e)))
+            def record_failure(request):
+                if request.url.startswith(office_origin):
+                    return
+                if request.resource_type=="media" and request.failure=="net::ERR_ABORTED":
+                    return
+                if "/assets/fixed_views/" in request.url and request.failure=="net::ERR_ABORTED":
+                    return
+                failed.append(f"{request.url}: {request.failure}")
+            page.on("requestfailed",record_failure)
+            page.goto(a.url.rstrip("/")+"/"+path,wait_until="domcontentloaded",timeout=120_000)
+            assert page.locator("h1").text_content()=="AWSM"
+            assert page.locator("header .title-name").text_content()==("智能体世界仿真与建图" if lang=="zh" else "Agentic World Simulation and Mapping")
+            assert page.locator("header time").get_attribute("datetime")=="2026-10-01"
+            assert page.locator(".tagline-emphasis").evaluate("element => getComputedStyle(element).textDecorationLine")=="none"
+            assert page.locator('nav a').last.get_attribute('href')=='https://github.com/wentingw/AWSM'
+            assert page.locator('nav a').last.inner_text()=='GitHub ↗'
+            assert page.locator('figure[id^="figure-"]').evaluate_all("elements => elements.map(element => element.id)")==[f'figure-{number}' for number in range(1,6)]
+            for number in range(1,6):
+                assert page.locator(f'#figure-{number} > figcaption').inner_text().startswith(f'Figure {number}.')
+            assert page.locator('#figure-2 > img').get_attribute('src')=='assets/fixed_five_view_comparison_m1_m4.jpg'
+            assert page.locator('.scene-comparison').get_attribute('id')=='figure-3'
+            assert page.locator('.fixed-comparison').get_attribute('id')=='figure-4'
+            assert not page.locator('.office-external-slot iframe').count()
+            for button in page.locator('[data-office-external]').all():
+                button.click()
+                assert page.locator('.office-external-slot iframe').get_attribute('src')==button.get_attribute('data-office-external')
+                assert page.locator('.office-external-slot iframe').get_attribute('lang')=='en'
+            page.locator('.office-external-grid').screenshot(path=str(out/f'{viewport}-{lang}-external-preview.png'))
+            assert page.locator(".citation-download").get_attribute("href")=="data/awsm.bib"
+            assert page.locator('nav a[aria-current="page"]').get_attribute("href")==path
+            assert page.locator('nav').get_by_role("link",name="中文",exact=True).get_attribute("href")=="zh.html"
+            assert page.locator('nav').get_by_role("link",name="EN",exact=True).get_attribute("href")=="index.html"
+            for section_id in ("results","motivation","workflow","more-results","related-work","limitations","citation","references"):
+                assert page.locator(f"#{section_id}").count()==1
+            assert page.locator("#results .insight-strip > div").count()==2
+            assert page.locator("#results .tldr").count()==1
+            assert page.locator("#motivation .contributions > li").count()==3
+            assert page.locator("#workflow .workflow-strip > div").count()==5
+            assert page.locator("#related-work > .inline-sources > a").count()==4
+            assert page.locator("#citation .citation-block").count()==1
+            assert page.locator("#references .references-list > li").count()>=7
+            assert page.locator('meta[name="twitter:card"]').get_attribute("content")=="summary_large_image"
+            demo=page.locator("#embodied-demo")
+            assert demo.count()==1
+            video=demo.locator("video")
+            assert video.count()==1
+            assert video.get_attribute("controls") is not None
+            assert video.get_attribute("autoplay") is None
+            assert video.get_attribute("poster")==demo_media["poster"]
+            assert video.locator("source").count()==1
+            assert video.locator("source").get_attribute("src")==demo_media["video"]
+            assert video.locator("source").get_attribute("type")=="video/mp4"
+            route_map=demo.locator("#demo-route-map img")
+            assert route_map.get_attribute("src")==demo_media["map"]["image"]
+            route_map.scroll_into_view_if_needed()
+            page.wait_for_function("image=>image.complete&&image.naturalWidth===1920",arg=route_map.element_handle())
+            assert demo.locator('a[download]').count()==2
+            assert demo.locator(f'a[href="{demo_media["hd"]["video"]}"]').count()==1
+            details=demo.locator("details.demo-protocol")
+            assert details.count()==1 and not details.get_attribute("open")
+            details.locator("summary").click()
+            assert details.get_attribute("open") is not None
+            h264_support=video.evaluate("v=>v.canPlayType('video/mp4; codecs=\"avc1.42E01E\"')")
+            if h264_support:
+                video.evaluate("v=>v.load()")
+                page.wait_for_function("v=>v.readyState>=1||v.error",arg=video.element_handle(),timeout=30_000)
+                video_state=video.evaluate("v=>({status:'PASS',error:v.error&&`${v.error.code}: ${v.error.message}`,duration:v.duration,width:v.videoWidth,height:v.videoHeight,readyState:v.readyState})")
+                assert not video_state["error"],f"demo video metadata/H.264 load failed: {video_state['error']}"
+                assert video_state["readyState"]>=1,"demo video metadata did not load"
+                assert abs(video_state["duration"]-demo_media["duration_s"])<0.25,video_state
+                assert video_state["width"]==demo_media["width"] and video_state["height"]==demo_media["height"],video_state
+                video.evaluate("async v=>{v.muted=true;await v.play()}")
+                page.wait_for_function("v=>v.currentTime>0.1",arg=video.element_handle(),timeout=10_000)
+                video.evaluate("v=>v.pause()")
+                video_state["video_playback"]="PASS"
+            else:
+                video_state={"status":"UNSUPPORTED","reason":"browser reports no H.264/AVC MP4 support; metadata not tested"}
+            page.wait_for_selector("#table-7 table"); assert page.locator(".table-figure table").count()==5
+            table2_headers=page.locator("#table-2 thead th").all_text_contents()
+            assert len(table2_headers)==5 and all("配准" not in x and "Alignment" not in x for x in table2_headers)
+            table3_headers=page.locator("#table-3 thead th").all_text_contents()
+            assert len(table3_headers)==4 and all("SHA256" not in x and "配准" not in x and "Alignment" not in x for x in table3_headers)
+            assert page.locator("#table-2").evaluate("(x)=>x.compareDocumentPosition(document.querySelector('#figure-1')) & Node.DOCUMENT_POSITION_FOLLOWING")
+            for table_index in (2,3,6,7):
+                assert page.locator(f"#table-{table_index} tbody strong").count()>0
+            assert page.locator("#overview-video > video").count()==1
+            assert page.locator("#overview-video > video > source").get_attribute("src")=="assets/awsm-promo-v5.mp4"
+            assert page.locator("#figure-2").evaluate("(x)=>x.compareDocumentPosition(document.querySelector('#figure-3')) & Node.DOCUMENT_POSITION_FOLLOWING")
+            assert page.locator("#more-results").evaluate("(x)=>x.compareDocumentPosition(document.querySelector('#embodied-demo')) & Node.DOCUMENT_POSITION_FOLLOWING")
+            figure3=page.locator("#figure-3"); figure3.scroll_into_view_if_needed()
+            page.wait_for_function("document.querySelector('#figure-3').dataset.state === 'ready'",timeout=120_000)
+            assert page.locator("#model-select").input_value()=="M4"
+            assert page.locator("#figure-3").get_attribute("data-loaded-model")=="M4"
+            scene_models=("M4","M3","GT") if a.quick else ("M1","M2","M3","M4","GT")
+            for method in scene_models:
+                page.select_option("#model-select",method)
+                page.wait_for_function("(m)=>{const p=document.querySelector('#figure-3');const d=p.sceneDiagnostics?.();return p.dataset.state==='ready'&&p.dataset.loadedModel===m&&d?.loaded===m&&d.gtLoaded&&d.singleCamera&&d.singleViewport&&(m==='GT'||d.hiddenCutaway>0)}",arg=method,timeout=120_000)
+                if not a.quick or method=="M4":
+                    figure3.screenshot(path=str(out/f"{viewport}-{lang}-{method}.png"))
+                page.select_option("#scene-mode","single")
+                if not a.quick:
+                    figure3.screenshot(path=str(out/f"{viewport}-{lang}-{method}-single.png"))
+                page.select_option("#scene-mode","compare")
+                if method=="M4":
+                    bounds=page.evaluate("document.querySelector('#figure-3').sceneDiagnostics().bounds")
+                    assert bounds["min"][0]<18<bounds["max"][0]
+                    assert bounds["min"][1]<1.2<bounds["max"][1]
+                    assert bounds["min"][2]<-21<bounds["max"][2]
+            page.select_option("#model-select","M2"); page.select_option("#scene-mode","compare")
+            before=page.evaluate("document.querySelector('#figure-3').sceneDiagnostics()")
+            page.locator(".scene-divider").focus(); page.keyboard.press("ArrowRight")
+            after=page.evaluate("document.querySelector('#figure-3').sceneDiagnostics()")
+            assert after["split"]>before["split"] and after["singleCamera"] and after["singleViewport"]
+            page.select_option("#scene-mode","single"); assert page.locator("#scene-split").is_disabled()
+            page.click("#scene-reset")
+            loaded_images=[]
+            for method in (("M1","M4") if a.quick else ("M1","M2","M3","M4")):
+                page.select_option("#compare-method",method)
+                for frame in (("0","144") if a.quick else ("0","36","72","108","144")):
+                    page.select_option("#compare-frame",frame)
+                    page.wait_for_function("([m,f])=>{const x=document.querySelector('#figure-4'),a=document.querySelector('#compare-pred'),b=document.querySelector('#compare-gt'),id=String(f).padStart(3,'0');return x.dataset.method===m&&x.dataset.frame===f&&a.currentSrc.endsWith(`/assets/fixed_views/${m}/${id}.png`)&&b.currentSrc.endsWith(`/assets/fixed_views/GT/${id}.png`)&&a.complete&&a.naturalWidth>0&&b.complete&&b.naturalWidth>0}",arg=[method,frame],timeout=30_000)
+                    loaded_images.append(f"{method}/{int(frame):03d}")
+            office=page.locator(".office-model-grid")
+            office.scroll_into_view_if_needed()
+            page.wait_for_function("[...document.querySelectorAll('.office-model-stage')].every(stage => stage.dataset.ready === 'true')",timeout=120_000)
+            office.screenshot(path=str(out/f"{viewport}-{lang}-office.png"))
+            page.wait_for_function("[...document.querySelectorAll('#office-cafe video')].every(video => video.readyState >= 1 && !video.error)",timeout=60_000)
+            for stage in page.locator(".office-model-stage").all():
+                initial=stage.evaluate("stage => stage.sceneDiagnostics()")
+                assert initial["ready"] and initial["cutaway"]
+                assert initial["camera"][1]>initial["bounds"]["max"][1]
+                toggle=stage.locator('button[aria-pressed]')
+                toggle.click()
+                assert not stage.evaluate("stage => stage.sceneDiagnostics().cutaway")
+                toggle.click()
+                bounds=stage.locator('canvas').bounding_box()
+                page.mouse.move(bounds['x']+bounds['width']*0.5,bounds['y']+bounds['height']*0.6)
+                page.mouse.down()
+                page.mouse.move(bounds['x']+bounds['width']*0.7,bounds['y']+bounds['height']*0.65,steps=8)
+                page.mouse.up()
+                stage.locator('button').first.click()
+                page.wait_for_function("([selector,expected]) => document.querySelector(selector).sceneDiagnostics().camera.every((value,index) => Math.abs(value-expected[index]) < 0.001)",arg=[f'.office-model-stage[data-src="{stage.get_attribute("data-src")}"]',initial['camera']],timeout=10_000)
+                assert stage.evaluate("stage => stage.sceneDiagnostics().cutaway")
+            overflow=page.evaluate("document.documentElement.scrollWidth > window.innerWidth")
+            assert not overflow,f"horizontal overflow: {viewport}/{lang}"
+            assert not errors,errors; assert not failed,failed
+            page.locator("header").screenshot(path=str(out/f"{viewport}-{lang}.png"))
+            run_status="PASS" if video_state["status"]=="PASS" else "PARTIAL"
+            results.append({"viewport":viewport,"language":lang,"status":run_status,"scene_models":scene_models,"fixed_views":len(loaded_images),"tables":5,"divider_changed":True,"video_metadata":video_state,"external_app":"403 in separate live check; embed wiring tested with fixture", "office_models":2,"overflow":overflow})
+            page.close()
+    browser.close()
+(out/"report.json").write_text(json.dumps(results,ensure_ascii=False,indent=2)+"\n")
+overall_status="PASS" if all(run["status"]=="PASS" for run in results) else "PARTIAL"
+print(json.dumps({"status":overall_status,"output":str(out),"runs":results},ensure_ascii=False))

scripts/site_builder.py

@@ -0,0 +1,6 @@
+#!/usr/bin/env python3
+"""Deterministically build the interactive bilingual publication."""
+import runpy
+from pathlib import Path
+
+runpy.run_path(str(Path(__file__).with_name("site_builder_interactive.py")), run_name="__main__")

scripts/site_builder_interactive.py

@@ -0,0 +1,442 @@
+#!/usr/bin/env python3
+"""Build the interactive bilingual publication from frozen sources."""
+from pathlib import Path
+from urllib.request import Request, urlopen
+import argparse, hashlib, html, json, shutil, tempfile
+from editorial import BASE_URL, BIBTEX, SITE_TITLE, REFERENCES, narrative_copy, editorial_page
+
+ROOT = Path(__file__).resolve().parents[1]
+DEFAULT_SOURCE = ROOT.parent / "world_lobby_four_trajectory_20260929"
+REFERENCE = ROOT.parent / "world_model_blog/private_repo/sceneweft_sync_20260925/blog"
+TABLE_SOURCE = "astra_blender2/evaluation/blog_extended_100_20260930/tables_1_7.json"
+VIEW_SOURCE = "astra_blender2/report/space/results/blog_tables_1_5/reference_style_figures.json"
+REG_DIR = "astra_blender2/evaluation/reference_style_figures"
+METHODS, FRAMES = ("M1", "M2", "M3", "M4"), (0, 36, 72, 108, 144)
+GT_URL = "https://wentingw.github.io/astra-world-model-blog/assets/comparison/GT.glb"
+GT_SHA = "6c28ab067c63c5b1fce19c6a972f3d66c7ac4d9bd9ce3b5a57e2e18f5641fe8f"
+GT_BYTES = 24816108
+CODE_URL = "https://github.com/wentingw/AWSM"
+OFFICE_URL = "https://office-cafe-vipe.hiwtishere.chatgpt.site/"
+
+def sha(path): return hashlib.sha256(path.read_bytes()).hexdigest()
+def write(path, text):
+    path.parent.mkdir(parents=True, exist_ok=True); path.write_text(text, encoding="utf-8")
+def dump(path, value): write(path, json.dumps(value, ensure_ascii=False, indent=2) + "\n")
+def copy_checked(src, dst, expected=None, source_label=None):
+    if not src.is_file(): raise FileNotFoundError(src)
+    actual = sha(src)
+    if expected and actual != expected: raise ValueError(f"SHA256 mismatch: {src}: {actual} != {expected}")
+    dst.parent.mkdir(parents=True, exist_ok=True); shutil.copy2(src, dst)
+    return {"source": source_label or str(src), "published": dst.relative_to(ROOT).as_posix(),
+            "sha256": actual, "bytes": src.stat().st_size}
+def mm(a, b): return [[sum(a[i][k]*b[k][j] for k in range(4)) for j in range(4)] for i in range(4)]
+
+def clean():
+    assets = ROOT / "assets"
+    if assets.exists():
+        for path in assets.iterdir():
+            if path.name != "office-cafe":
+                shutil.rmtree(path) if path.is_dir() else path.unlink()
+    for name in ("data", "evidence", "models", "vendor", "qa", "content"):
+        if (ROOT/name).exists(): shutil.rmtree(ROOT/name)
+    for name in ("index.html","en.html","app.js","scene-compare.js","style.css",
+                 "article.zh.md","article.en.md",".publication.json"):
+        (ROOT/name).unlink(missing_ok=True)
+
+def download_gt():
+    target=ROOT/"assets/comparison/GT.glb"; target.parent.mkdir(parents=True,exist_ok=True)
+    request=Request(GT_URL,headers={"User-Agent":"agentic-world-blog-builder/2"})
+    with tempfile.NamedTemporaryFile(dir=target.parent,delete=False) as temp:
+        tmp=Path(temp.name)
+        with urlopen(request,timeout=180) as response: shutil.copyfileobj(response,temp)
+    try:
+        if tmp.stat().st_size != GT_BYTES: raise ValueError(f"GT bytes: {tmp.stat().st_size} != {GT_BYTES}")
+        if sha(tmp) != GT_SHA: raise ValueError(f"GT SHA256: {sha(tmp)} != {GT_SHA}")
+        tmp.replace(target)
+    finally: tmp.unlink(missing_ok=True)
+    return {"source_url":GT_URL,"published":"assets/comparison/GT.glb","sha256":GT_SHA,"bytes":GT_BYTES}
+
+def table(n):
+    return f'<figure class="table-figure" id="table-{n}" data-table-key="table{n}"><figcaption><span>Table {n}.</span> <span class="table-title"></span></figcaption><div class="table-scroll table-mount" role="region" tabindex="0" aria-label="Table {n}"></div></figure>'
+
+def hero_figure(en):
+    label = "AWSM introduction video" if en else "AWSM 介绍视频"
+    resources_label = "AWSM project resources" if en else "AWSM 项目资源"
+    github_label = "GitHub (opens in a new window)" if en else "GitHub（在新窗口打开）"
+    links = f'''<div class="project-links" role="group" aria-label="{resources_label}"><a class="project-resource" href="{CODE_URL}" target="_blank" rel="noopener noreferrer" aria-label="{github_label}" title="{github_label}"><img src="assets/icons/github.svg" width="24" height="24" alt="">GitHub<svg class="project-external-arrow" viewBox="0 0 16 16" width="12" height="12" fill="none" stroke="currentColor" stroke-width="1.6" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><path d="M4 12 12 4M4 4h8v8"/></svg></a><span class="project-resource"><img src="assets/icons/huggingface.svg" width="24" height="24" alt="">Hugging Face (Coming soon)</span></div>'''
+    return f'''<div class="hero" id="overview-video"><video controls playsinline preload="metadata" poster="assets/awsm-promo-v5-poster.jpg" aria-label="{label}"><source src="assets/awsm-promo-v5.mp4" type="video/mp4"><a href="assets/awsm-promo-v5.mp4">{label}</a></video>{links}</div>'''
+
+def overview_figure(en):
+    return f'''<figure id="figure-2"><img loading="lazy" src="assets/fixed_five_view_comparison_m1_m4.jpg" alt="M1 to M4 and input GT across five fixed views"><figcaption><span>Figure 2.</span> {"M1–M4 and input GT across five fixed views." if en else "M1–M4 与输入 GT 在五个固定视角下的并排比较。"}</figcaption></figure>'''
+
+def scene_figure(en):
+    options="".join(f'<option value="{m}"{" selected" if m=="M4" else ""}>{m}</option>' for m in METHODS)+'<option value="GT">GT</option>'
+    return f'''<figure class="scene-comparison" id="figure-3" data-state="idle"><h3>{"Rotate, zoom, and inspect the frozen scenes" if en else "旋转、缩放，检查冻结场景"}</h3>
+<div class="scene-toolbar"><label class="method-control">{"Reconstruction / GT" if en else "重建方法 / GT"}<select id="model-select">{options}</select></label><label>{"Display" if en else "显示方式"}<select id="scene-mode"><option value="compare">{"Split comparison with GT" if en else "中轴线对比 GT"}</option><option value="single">{"View selected scene alone" if en else "单独查看所选场景"}</option></select></label><button id="scene-reset" type="button">{"Reset view" if en else "重置视角"}</button></div>
+<div class="scene-stage"><span class="scene-label scene-label-left">M4</span><span class="scene-label scene-label-right">GT</span><button class="scene-divider" type="button" role="slider" aria-label="Comparison divider" aria-valuemin="2" aria-valuemax="98" aria-valuenow="50"></button><input id="scene-split" class="scene-split" type="range" min="2" max="98" value="50" aria-label="Comparison split"><p class="scene-status">{"Scroll here to load the 3D comparison." if en else "滚动到此处加载三维对比。"}</p></div><button id="scene-retry" class="scene-retry" type="button">{"Retry" if en else "重试"}</button>
+<figcaption><span>Figure 3.</span> {"One camera and one full viewport compare the selected frozen model with GT. Ceilings and surrounding walls are hidden at runtime for the same cutaway-style inspection; source files remain unchanged." if en else "所选冻结模型与 GT 共用同一相机和完整 viewport；网页运行时隐藏顶棚与四周墙体，以相同剖视方式检查，源模型文件保持不变。"}</figcaption></figure>'''
+
+def fixed_figure(en):
+    methods="".join(f'<option value="{m}">{m}</option>' for m in METHODS)
+    frames="".join(f'<option value="{f}">{f}</option>' for f in FRAMES)
+    return f'''<figure class="fixed-comparison" id="figure-4"><div class="controlrow"><label>{"Method" if en else "方法"}<select id="compare-method" disabled>{methods}</select></label><label>{"Fixed view" if en else "固定视角"}<select id="compare-frame" disabled>{frames}</select></label></div><div class="compare-pair"><div><img id="compare-pred" src="assets/fixed_views/M1/000.png" alt="Selected frozen model rendering"><p id="compare-label" class="compare-subcaption">M1 / {"VIEW" if en else "视角"} 0</p></div><div><img id="compare-gt" src="assets/fixed_views/GT/000.png" alt="Input GT RGB"><p class="compare-subcaption">INPUT / GT RGB</p></div></div><figcaption><span>Figure 4.</span> {"Frozen-model rendering and input GT RGB from exactly the same view." if en else "同一所选视角下的冻结模型渲染与输入 GT RGB。"}</figcaption></figure>'''
+
+def downloads():
+    return '<div class="asset-downloads">'+"".join(f'<p><strong>{m}</strong> <a href="models/{m}/scene.blend" download>scene.blend ↓</a> · <a href="models/{m}/scene.glb" download>scene.glb ↓</a> · <a href="evidence/{m}/manifest.json">manifest ↗</a></p>' for m in METHODS)+'</div>'
+
+def office_videos(en):
+    media_root = "assets/office-cafe"
+    titles = ("Input video", "Procedural model", "Point cloud") if en else ("输入视频", "程序化模型", "点云")
+    descriptions = ("Original phone walkthrough.", "Model rendered along the source-camera trajectory.", "Scan rendered along the same trajectory.") if en else ("原始手机拍摄视频。", "沿原视频相机轨迹渲染的模型。", "沿相同轨迹渲染的扫描点云。")
+    cards = "".join(
+        f'<figure class="office-video-card"><video data-synced-video muted preload="metadata" playsinline poster="assets/office-cafe/{stem}-poster.jpg" src="{media_root}/{stem}-web.mp4?v={sha(ROOT / media_root / (stem + "-web.mp4"))[:12]}" aria-label="{title}"></video><figcaption><strong>{title}</strong><span>{description}</span></figcaption></figure>'
+        for stem, title, description in zip(("office-input", "office-pmodel", "office-scan"), titles, descriptions)
+    )
+    return f'''<div class="office-video-grid office-video-sync">{cards}</div>
+<div class="office-video-controls"><button class="office-video-play" type="button" aria-label="{"Play all three videos" if en else "播放全部三个视频"}" aria-pressed="false">▶</button><input class="office-video-progress" type="range" min="0" max="1000" value="0" aria-label="{"Shared video progress" if en else "三个视频的共同播放进度"}"><output class="office-video-time">00:00 / 01:10</output><p class="office-video-status" role="status" aria-live="polite">{"Shared playback · 70.1 seconds · 30 fps" if en else "同步播放 · 70.1 秒 · 30 fps"}</p></div>'''
+
+
+def office_external(en):
+    heading = "Explore the original interactive apps" if en else "打开原始交互网页"
+    titles = ("Interactive model", "Shake experiment") if en else ("交互模型", "摇晃实验")
+    captions = ("Local preview from the published model-render video.", "Local preview from the published strong-shake replay.") if en else ("来自已发布模型渲染视频的本地预览。", "来自已发布较强摇晃回放的本地预览。")
+    load = "Load external app · English requested" if en else "加载外部网页 · 请求英文版"
+    notice = "Links request English with lang=en; the external host controls language and availability. These local content previews are not screenshots of the external website. If embedding is unavailable, use the full-screen links or the local models and videos below." if en else "链接通过 lang=en 请求英文，实际语言与可用性由外站控制。这里展示的是本地内容预览，并非外站页面截图。若嵌入不可用，可使用全屏外链，或查看下方本地模型与视频。"
+    cards = "".join(
+        f'<figure class="office-external-card"><a href="{OFFICE_URL}{suffix}" target="_blank" rel="noopener noreferrer"><img loading="lazy" src="assets/office-cafe/{poster}" alt="{title}: {caption}"></a><figcaption><strong>{title}</strong><p>{caption}</p><button type="button" data-office-external="{OFFICE_URL}{suffix}" aria-controls="office-external-slot">{load}</button></figcaption></figure>'
+        for suffix, poster, title, caption in zip(("?lang=en", "shake?lang=en"), ("office-pmodel-poster.jpg", "office-shake-heavy-poster.jpg"), titles, captions)
+    )
+    return f'<div class="office-external"><h4>{heading}</h4><div class="office-external-grid">{cards}</div><p class="office-external-note">{notice}</p><div id="office-external-slot" class="office-external-slot" hidden></div></div>'
+
+
+def office_case(en):
+    if en:
+        return f'''<div id="office-cafe" class="analysis-block external-case"><h3>Beyond the lobby: what survives in a phone-captured space?</h3><p>Office Café is not a second run of the World Lobby benchmark and is not quantitative validation under the same protocol. It is an external, real-capture case study showing how an object-centric scene, a solved source-camera trajectory, video-to-model comparison, and physics replays can coexist in one inspectable artifact. An author-observed failure—a curved real-world corner simplified into a square one—motivates separating local shape from global scale. A correct scale alone cannot repair that shape error; this is a qualitative observation, not another measured benchmark.</p><p class="office-links"><a href="https://office-cafe-vipe.hiwtishere.chatgpt.site/?lang=en" target="_blank" rel="noopener">Open full-screen model ↗</a><a href="https://office-cafe-vipe.hiwtishere.chatgpt.site/shake?lang=en" target="_blank" rel="noopener">Open shake experiment ↗</a></p>{office_external(en)}<div class="office-model-grid"><figure class="office-model-card"><div class="office-model-stage" data-src="assets/office-cafe/scene.glb" aria-label="Interactive Office Café ViPE model"></div><figcaption><strong>Current model</strong><span>Video plus estimated depth as input; drag to orbit, right-drag to pan, and scroll to zoom.</span></figcaption></figure><figure class="office-model-card"><div class="office-model-stage" data-src="assets/office-cafe/atlas-scene.glb" aria-label="Interactive legacy Office Café model"></div><figcaption><strong>Legacy model</strong><span>Video-only input; this legacy layout was not recalibrated against the ViPE scan.</span></figcaption></figure></div><p class="office-model-note">Both models open in an elevated cutaway view: upper geometry is clipped only for display to expose the interior. Toggle Cutaway to inspect the complete model; Reset view restores the overview. Authored coordinates, materials, and model files are unchanged. Linked orbit controls compare layouts, not registered geometry.</p>{office_videos(en)}<div class="office-video-grid"><figure class="office-video-card office-video-card-wide"><video controls preload="metadata" playsinline poster="assets/office-cafe/office-shake-heavy-poster.jpg" src="assets/office-cafe/office-shake-heavy.mp4"></video><figcaption><strong>Strong earthquake effect</strong><span>Every object in the scene is interactable, and cabinet doors can open; this replay shows the physical response under strong shaking.</span></figcaption></figure></div><script type="module" src="office-models.js"></script></div>'''
+    return f'''<div id="office-cafe" class="analysis-block external-case"><h3>走出仿真大厅：手机采集空间中还剩下什么？</h3><p>Office Café 不是 World Lobby 基准的第二次运行，也不构成同协议的定量验证。它是一个真实采集的外部案例，用于展示对象化场景、求解后的原视频相机轨迹、视频—模型对照与物理回放如何组合成一个可检查的空间产物。作者观察到的一类失败是：真实的弯曲转角被简化成方正结构。它提醒我们区分���部形状与全局尺度——尺度正确并不能自动修复形状错误；这里将其作为定性观察，而非另一组定量结果。</p><p class="office-links"><a href="https://office-cafe-vipe.hiwtishere.chatgpt.site/?lang=en" target="_blank" rel="noopener">全屏打开模型 ↗</a><a href="https://office-cafe-vipe.hiwtishere.chatgpt.site/shake?lang=en" target="_blank" rel="noopener">打开摇晃实验 ↗</a></p>{office_external(en)}<div class="office-model-grid"><figure class="office-model-card"><div class="office-model-stage" data-src="assets/office-cafe/scene.glb" aria-label="可旋转查看的 Office Café 原有模型"></div><figcaption><strong>原有模型</strong><span>输入为视频与估计深度；拖动旋转、右键平移、滚轮缩放。</span></figcaption></figure><figure class="office-model-card"><div class="office-model-stage" data-src="assets/office-cafe/atlas-scene.glb" aria-label="可旋转查看的 Office Café 旧模型"></div><figcaption><strong>旧模型</strong><span>输入仅为视频；该旧版布局未按 ViPE 扫描重新标定。</span></figcaption></figure></div><p class="office-model-note">两个模型默认采用俯视剖切视图，仅在显示时裁去上部遮挡，以展示室内布局；点击“剖切视图”可切换完整模型，“重置视角”可恢复总览。原始坐标、材质和模型文件保持不变。联动旋转用于观察布局，不代表几何配准。</p>{office_videos(en)}<div class="office-video-grid"><figure class="office-video-card office-video-card-wide"><video controls preload="metadata" playsinline poster="assets/office-cafe/office-shake-heavy-poster.jpg" src="assets/office-cafe/office-shake-heavy.mp4"></video><figcaption><strong>较强地震效果</strong><span>场景中的物体均可交互，柜门可以打开；此段展示较强摇晃下的物理响应。</span></figcaption></figure></div><script type="module" src="office-models.js"></script></div>'''
+
+def office_section(en):
+    return f'''<section id="office-cafe"><p class="section-tag">08 / OFFICE CAFÉ</p><h2>{"From a phone walkthrough to an interactive twin" if en else "从手机视频到可交互空间"}</h2><p>{"The complete Office Café viewer is embedded below. It retains model-only, scan ↔ model, and source-video ↔ model comparison modes; playback follows all 2,103 solved source-camera poses. The linked shake experiment retains control, gentle, and strong MuJoCo replays." if en else "下方嵌入完整 Office Café 查看器：保留“模型”“扫描 ↔ 模型”“原视频 ↔ 模型”三种对照方式，并可沿 2,103 个求解后的原视频相机位姿播放。摇晃实验继续提供零摇晃、轻度和较强三组 MuJoCo 回放。"}</p><p class="office-links"><a href="{OFFICE_URL}" target="_blank" rel="noopener">{"Open full viewer" if en else "全屏打开模型"} ↗</a><a href="{OFFICE_URL}shake.html" target="_blank" rel="noopener">{"Open shake experiment" if en else "打开摇晃实验"} ↗</a></p><div class="office-cafe-embed"><iframe src="{OFFICE_URL}" title="Office Café interactive spatial twin" loading="lazy" allow="fullscreen" allowfullscreen></iframe></div></section>'''
+
+def legacy_page(en=False):
+    title="Four paths to an editable world" if en else "四条路径，重建一个可编辑世界"
+    desc=("A measured comparison of four Astra–Blender reconstructions: pose, depth, geometry, appearance, and novel-view depth." if en else "对四种 Astra–Blender 重建的克制比较：位姿、深度、几何、外观和同场景新视角深度。")
+    sections=(["Question & four methods","Pose and native/model depth","Editable M1–M4","Geometry","Appearance","Same-scene novel-view depth","Limits & evidence","Office Café twin"] if en else ["问题与四种方法","位姿与原生/模型深度","可编辑 M1–M4","几何","外观","同场景新视角深度","局限与证据","Office Café 模型"])
+    ids=("question","pose-depth","editable","geometry","appearance","novel-depth","limitations","office-cafe")
+    toc="".join(f'<a href="#{i}">{t}</a>' for i,t in zip(ids,sections)); current,other=("en.html","index.html") if en else ("index.html","en.html")
+    methods=("M1 uses sampled RGB and an unmeasured layout. M2 uses RGB-only ViPE poses and pose-conditioned DA3 depth. M3 uses monocular-inertial ORB-SLAM3 poses and pose-conditioned DA3 depth. M4 uses GT camera poses and pose-conditioned DA3 depth; GT mesh and depth are not modelling inputs." if en else "M1 使用采样 RGB 与无测量尺度的布局；M2 使用 RGB-only ViPE 位姿与 pose-conditioned DA3 深度；M3 使用 ORB-SLAM3 单目惯性位姿与 pose-conditioned DA3 深度；M4 使用 GT 相机位姿与 pose-conditioned DA3 深度，建模时不读取 GT 网格或 GT 深度。")
+    m1=("M1 uses one frozen GT-assisted Sim(3) from five manual camera associations. It is diagnostic only and does not represent native metric recovery." if en else "M1 使用由 5 个手工相机关联得到、冻结不变的 GT-assisted Sim(3)；仅作诊断，不表示原生 metric recovery。")
+    m4=("M4 is revision 4: 74 semantic objects, 37 static AABB colliders, and 180 cameras. The independent reviewer examined revision 2 only. Revisions 3–4 fixed identified issues, but revision 4 has not received a complete independent visual re-review; final visual approval is not claimed." if en else "M4 为 revision 4：74 个 semantic objects、37 个 static AABB colliders、180 台 cameras。独立 reviewer 只审阅了 rev2；rev3/4 修复了已指出的问题，但 rev4 尚无完整独立视觉复审，因此不声称最终视觉通过。")
+    copy_notice=("These display copies and runtime registrations do not alter the frozen models. M1's alignment is diagnostic only." if en else "这些展示副本和运行时配准不改变冻结模型；M1 配准仅作诊断。")
+    return f'''<!doctype html><html lang="{"en" if en else "zh-CN"}"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"><title>{title} — Agentic World</title><meta name="description" content="{desc}"><meta property="og:type" content="article"><meta property="og:title" content="{title}"><meta property="og:description" content="{desc}"><meta property="og:image" content="{BASE_URL}assets/fixed_five_view_comparison_m1_m4.jpg"><meta property="og:url" content="{BASE_URL}{current}"><link rel="canonical" href="{BASE_URL}{current}"><link rel="alternate" hreflang="zh-CN" href="{BASE_URL}index.html"><link rel="alternate" hreflang="en" href="{BASE_URL}en.html"><link rel="stylesheet" href="style.css"><script type="importmap">{{"imports":{{"three":"./vendor/three/three.module.js"}}}}</script><script src="app.js" defer></script><script type="module" src="scene-compare.js"></script></head>
+<body data-lang="{"en" if en else "zh"}"><a class="skip" href="#main">{"Skip to article" if en else "跳到正文"}</a><nav><a class="brand" href="index.html">ASTRA / WORLD MODELS</a><div><a href="index.html" {"aria-current='page'" if not en else ""}>中文</a><a href="en.html" {"aria-current='page'" if en else ""}>EN</a><a href="{CODE_URL}" target="_blank" rel="noopener noreferrer">GitHub ↗</a></div></nav><header><p class="eyebrow">RESEARCH JOURNAL · 30 SEPTEMBER 2026</p><h1>{title}</h1><p class="lead">{desc}</p><div class="meta"><span>1 scene</span><span>4 editable models</span><span>7 frozen tables</span></div></header>
+{hero_figure(en)}<div class="layout"><aside class="toc"><p>ON THIS PAGE</p>{toc}</aside><main id="main">
+<section id="question"><p class="section-tag">01 / QUESTION</p><h2>{sections[0]}</h2><p class="standfirst">{"M4 is strongest overall in this case, but this single-scene study does not establish general superiority." if en else "M4 在本案例中综合最好，但单场景工程实验不能证明一般性优势。"}</p><p>{methods}</p>{table(1)}</section>
+<section id="pose-depth"><p class="section-tag">02 / MEASUREMENT</p><h2>{sections[1]}</h2><p>{"Native depth evaluates the upstream pipeline. Model depth ray-casts the final frozen Blender scene." if en else "原生深度评估上游链路；模型深度从评测相机向最终冻结 Blender 场景投射射线。"}</p><aside class="notice"><strong>M1 diagnostic limitation.</strong> {m1}</aside>{table(2)}<figure id="figure-1"><img loading="lazy" src="assets/figure14_abc.png" alt="Trajectory and depth composite"><figcaption><span>Figure 1.</span> {"Trajectory, native depth, and model depth." if en else "轨迹、原生深度与模型深度。"}</figcaption></figure></section>
+<section id="editable"><p class="section-tag">03 / EDITABLE ASSETS</p><h2>{sections[2]}</h2><p>{m4}</p><p>{copy_notice}</p>{overview_figure(en)}{scene_figure(en)}{fixed_figure(en)}{downloads()}</section>
+<section id="geometry"><p class="section-tag">04 / GEOMETRY</p><h2>{sections[3]}</h2><p>{"Bidirectional distances use 100,000 seeded samples per side under the declared global alignment; no mesh ICP is fitted." if en else "双向表面距离每侧使用 100,000 个固定随机种子样本，并采用已声明的全局配准；不额外拟合 mesh ICP。"}</p>{table(3)}</section>
+<section id="appearance"><p class="section-tag">05 / APPEARANCE</p><h2>{sections[4]}</h2><p>{"Table 4 uses the five fixed 640×480 checks (0, 36, 72, 108, 144). Table 5 uses one seeded set of 100 views sampled from the other 175 modeling frames. These 100 views are new evaluation views relative to the fixed-five check, but they were still available as modeling RGB inputs. No crop, mask, or per-view fitting is used." if en else "Table 4 使用 5 个固定 640×480 检查视角（0、36、72、108、144）。Table 5 从其余 175 张建模帧中按固定随机种子抽取同一组 100 个视角；它们只相对于固定 5 帧是新的检查视角，仍属于建模 RGB 输入，并非未见输入。全部评测均不裁剪、不遮罩、不逐视角拟合。"}</p>{table(4)}{table(5)}</section>
+<section id="novel-depth"><p class="section-tag">06 / NOVEL VIEWS</p><h2>{sections[5]}</h2><p>{"Table 6 evaluates model optical-Z on the five fixed reference views. Table 7 uses exactly the same seeded 100-view set as Table 5 and aggregates all GT-valid pixels." if en else "Table 6 在固定 5 个参考视角上评测模型 optical-Z；Table 7 使用与 Table 5 完全相同的固定随机 100 视角，并在全部 GT 有效像素上汇总。"}</p><figure id="figure-5"><img loading="lazy" src="assets/model_depth_error_five_views_m1_m4.png" alt="Model depth error for M1 to M4"><figcaption><span>Figure 5.</span> {"Model-depth error over five fixed views; teal denotes missing or invalid samples." if en else "M1–M4 五个固定视角的模型深度误差；青绿色表示缺失或无效样本。"}</figcaption></figure>{table(6)}{table(7)}<noscript>JavaScript is required for the tables and interactive figures.</noscript></section>
+<section id="limitations"><p class="section-tag">07 / LIMITS & EVIDENCE</p><h2>{sections[6]}</h2><p>{"One synthetic scene and one run per route are insufficient for broad claims." if en else "单一合成场景、每条路线一次工程运行，不足以支持广泛结论。"}</p><p><a href="data/tables_1_7.json">Tables JSON</a> · <a href="data/scene_comparison.json">Scene manifest</a> · <a href="data/fixed_views.json">Fixed views</a> · <a href="evidence/publication_manifest.json">Manifest</a> · <a href="evidence/SHA256SUMS">SHA256</a></p></section>{office_section(en)}</main></div><footer>Agentic World · frozen local publication · <a href="{other}">{"中文" if en else "English"}</a></footer></body></html>'''
+
+STYLE=r''':root{--paper:#f7f6f1;--white:#fffefa;--ink:#182c29;--muted:#60706a;--green:#196451;--line:#d9dfd7;--gold:#b87e40}*{box-sizing:border-box}html{scroll-behavior:smooth}body{margin:0;background:var(--paper);color:var(--ink);font:17px/1.8 Inter,"Noto Sans CJK SC",system-ui,sans-serif}a{color:var(--green);text-underline-offset:4px}img{display:block;max-width:100%;height:auto}.skip{position:absolute;left:-9999px}.skip:focus{left:12px;top:12px;background:white;padding:8px;z-index:9}nav{max-width:1420px;margin:auto;padding:22px 4vw;display:flex;justify-content:space-between;border-bottom:1px solid var(--line);font-size:12px;letter-spacing:.09em}nav div{display:flex;gap:20px}.brand{font-weight:750;text-decoration:none}header{max-width:1120px;margin:56px auto 36px;padding:0 32px}.eyebrow,.section-tag{font-size:11px;letter-spacing:.16em;color:var(--green);font-weight:700}h1,h2{font-family:Georgia,"Noto Serif CJK SC",serif;font-weight:500}h1{font-size:clamp(46px,7vw,88px);line-height:1.12;letter-spacing:-.035em;margin:18px 0}.lead{max-width:800px;color:var(--muted);font-size:20px}.meta{display:flex;gap:24px;flex-wrap:wrap;font-size:12px;color:var(--muted)}.hero{max-width:1320px;margin:35px auto 68px;padding:0 28px}.hero img,figure>img{width:100%;background:#e8ebe5}.hero-pair-grid{display:grid;grid-template-columns:1fr 1fr;gap:3px;background:var(--ink)}.hero-pair-grid img{width:100%;aspect-ratio:4/3;object-fit:cover}.hero-pair-grid p{margin:0;padding:7px 12px;background:var(--ink);color:#f7f6f1;font-size:10px;letter-spacing:.11em}figcaption{font-size:12px;color:var(--muted);margin-top:10px}figcaption span{font-weight:700;color:var(--green)}.layout{max-width:1300px;margin:auto;padding:0 30px;display:grid;grid-template-columns:190px minmax(0,920px);gap:60px;justify-content:center}.toc{position:sticky;top:24px;align-self:start;display:flex;flex-direction:column;gap:9px;font-size:12px}.toc a{text-decoration:none;color:var(--muted)}main{min-width:0}section{padding:28px 0 38px;border-top:1px solid var(--line)}section:first-child{border-top:0}.standfirst{font-size:23px;line-height:1.55;color:var(--green)}h2{font-size:32px;margin:7px 0 22px}.notice{border-left:3px solid var(--gold);padding:10px 18px;margin:24px 0;color:var(--muted);background:#fff9}.table-figure{margin:30px 0}.table-scroll{overflow:auto;border:1px solid var(--line);background:var(--white);margin-top:9px}table{width:100%;border-collapse:collapse;white-space:nowrap;font-size:12px;font-variant-numeric:tabular-nums}th,td{padding:11px 12px;border-bottom:1px solid var(--line);text-align:left;vertical-align:top}th{background:#edf0e9;color:var(--muted)}td.wrap{white-space:normal;min-width:240px}.scene-comparison,.fixed-comparison{margin:34px 0;padding:20px;border:1px solid var(--line);border-radius:6px;background:var(--white)}.scene-comparison h3{margin:0 0 18px}.scene-toolbar,.controlrow{display:flex;gap:14px 20px;align-items:end;flex-wrap:wrap}.scene-toolbar label,.controlrow label{display:flex;flex-direction:column;font-size:12px;color:var(--muted)}.method-control{flex:1 1 280px}select,button{font:inherit}.scene-toolbar button,.scene-retry{font-size:13px;padding:9px 13px;color:var(--green);border:1px solid var(--line);border-radius:3px;background:var(--paper);cursor:pointer}.scene-stage{position:relative;height:540px;width:100%;background:#e9ede5;margin-top:18px;overflow:hidden;border-radius:4px;--split:50%}.scene-stage canvas{display:block;width:100%;height:100%;touch-action:none}.scene-label{position:absolute;top:12px;z-index:3;padding:4px 8px;background:#fffde;color:var(--ink);font-size:11px;pointer-events:none}.scene-label-left{left:12px}.scene-label-right{right:12px}.scene-divider{position:absolute;z-index:4;left:var(--split);top:0;width:3px;height:100%;padding:0;border:0;background:white;box-shadow:0 0 0 1px #182c2955;transform:translateX(-50%);cursor:ew-resize;touch-action:none}.scene-divider:after{content:"↔";position:absolute;top:50%;left:50%;transform:translate(-50%,-50%);width:30px;height:30px;line-height:27px;border-radius:50%;background:white;color:var(--green);border:1px solid var(--line)}.scene-split{position:absolute;width:1px;height:1px;opacity:0}.scene-status{position:absolute;left:50%;top:50%;transform:translate(-50%,-50%);padding:9px 12px;background:#fffef0;z-index:5;font-size:13px}.scene-status.error{color:#8d251d}.scene-retry{display:none;margin-top:10px}.scene-comparison[data-state=error] .scene-retry{display:inline-block}.scene-comparison:not(.is-comparing) .scene-divider{display:none}.compare-pair{display:grid;grid-template-columns:1fr 1fr;gap:14px;margin-top:16px}.compare-pair img{width:100%;aspect-ratio:4/3;object-fit:contain;background:#e9ede5}.compare-subcaption{margin:6px 0 0;font-size:11px;color:var(--muted);letter-spacing:.04em}.asset-downloads{display:grid;grid-template-columns:1fr 1fr;gap:0 20px;font-size:12px}.asset-downloads p{margin:4px 0}.office-cafe-embed{margin-top:24px;border:1px solid var(--line);border-radius:6px;overflow:hidden;background:#111}.office-cafe-embed iframe{display:block;width:100%;height:min(880px,85vh);min-height:680px;border:0;background:#111}.office-links{display:flex;gap:18px;flex-wrap:wrap;font-size:12px}footer{max-width:1120px;margin:70px auto 0;padding:30px;border-top:1px solid var(--line);font-size:12px;color:var(--muted)}@media(max-width:950px){.layout{grid-template-columns:1fr}.toc{position:static;display:grid;grid-template-columns:1fr 1fr}}@media(max-width:600px){body{font-size:16px}nav{padding:17px 18px}.brand{font-size:10px}nav div{gap:12px}header{padding:0 20px;margin-top:35px}h1{font-size:43px}.lead{font-size:17px}.hero{padding:0 12px;margin-bottom:40px}.layout{padding:0 20px}.toc{grid-template-columns:1fr}.standfirst{font-size:20px}h2{font-size:27px}.scene-comparison,.fixed-comparison{padding:12px}.scene-stage{height:360px}.compare-pair,.asset-downloads{grid-template-columns:1fr}.office-cafe-embed iframe{height:760px;min-height:620px}}'''
+
+APP=r'''const C={zh:{t:["输入与建模路线","位姿、原生深度与模型深度","冻结模型几何","固定 5 视角外观","同场景 100 视角外观","固定 5 视角模型深度","同场景 100 视角模型深度"],view:"视角",l:{method:"方法",allowed_inputs:"允许输入",geometry_source:"几何来源",modeling_role:"建模角色",ate_m:"ATE (m) ↓",rotation_deg:"旋转误差 (°) ↓",native_depth_absrel:"原生深度 AbsRel ↓",model_depth_absrel:"模型深度 AbsRel ↓",alignment:"配准 / 限制",model_to_gt_mean_m:"模型→GT 均值 (m) ↓",observed_gt_to_model_mean_m:"观测 GT→模型均值 (m) ↓",bidirectional_mean_m:"双向平均 (m) ↓",model_sha256:"模型 SHA256",psnr_db:"PSNR (dB) ↑",ssim:"SSIM ↑",lpips_alex_v01:"LPIPS ↓",absrel:"AbsRel ↓",rmse_m:"RMSE (m) ↓",coverage:"覆盖率 ↑",penalized_mae_m:"缺失惩罚 MAE (m) ↓"}},en:{t:["Inputs and modelling routes","Pose, native depth, and model depth","Frozen-model geometry","Fixed-five appearance","Same-scene 100-view appearance","Fixed-five model depth","Same-scene 100-view model depth"],view:"VIEW",l:{method:"Method",allowed_inputs:"Allowed inputs",geometry_source:"Geometry source",modeling_role:"Modelling role",ate_m:"ATE (m) ↓",rotation_deg:"Rotation (°) ↓",native_depth_absrel:"Native depth AbsRel ↓",model_depth_absrel:"Model depth AbsRel ↓",alignment:"Alignment / limitation",model_to_gt_mean_m:"Model→GT mean (m) ↓",observed_gt_to_model_mean_m:"Observed GT→model mean (m) ↓",bidirectional_mean_m:"Bidirectional mean (m) ↓",model_sha256:"Model SHA256",psnr_db:"PSNR (dB) ↑",ssim:"SSIM ↑",lpips_alex_v01:"LPIPS ↓",absrel:"AbsRel ↓",rmse_m:"RMSE (m) ↓",coverage:"Coverage ↑",penalized_mae_m:"Penalized MAE (m) ↓"}}};const c=C[document.body.dataset.lang],directions={ate_m:"min",rotation_deg:"min",native_depth_absrel:"min",model_depth_absrel:"min",model_to_gt_mean_m:"min",observed_gt_to_model_mean_m:"min",bidirectional_mean_m:"min",psnr_db:"max",ssim:"max",lpips_alex_v01:"min",absrel:"min",rmse_m:"min",coverage:"max",penalized_mae_m:"min"},hidden={2:new Set(["alignment"]),3:new Set(["model_sha256","alignment"])},fmt=(k,v)=>v===null?"—":typeof v!=="number"?String(v):["native_depth_absrel","model_depth_absrel","absrel","coverage"].includes(k)?(100*v).toFixed(2)+"%":v.toFixed(4);fetch("data/tables_1_7.json").then(r=>{if(!r.ok)throw Error(r.status);return r.json()}).then(d=>{[1,2,3,4,5,6,7].forEach(i=>{const f=document.querySelector("#table-"+i);if(!f)return;const rows=d.tables["table"+i],keys=Object.keys(rows[0]).filter(k=>!hidden[i]?.has(k)),best={};keys.forEach(k=>{if(!directions[k])return;const values=rows.map(r=>r[k]).filter(v=>typeof v==="number"&&Number.isFinite(v));if(values.length)best[k]=(directions[k]==="max"?Math.max:Math.min)(...values)});f.querySelector(".table-title").textContent=c.t[i-1];const t=document.createElement("table"),h=t.createTHead().insertRow();keys.forEach(k=>{const x=document.createElement("th");x.scope="col";x.textContent=c.l[k]||k;h.append(x)});const b=t.createTBody();rows.forEach(r=>{const tr=b.insertRow();keys.forEach((k,j)=>{const x=document.createElement(j?"td":"th");if(!j)x.scope="row";if(["allowed_inputs","geometry_source","modeling_role","alignment"].includes(k))x.className="wrap";const text=fmt(k,r[k]);if(typeof r[k]==="number"&&r[k]===best[k]){const strong=document.createElement("strong");strong.textContent=text;x.append(strong)}else x.textContent=text;tr.append(x)})});f.querySelector(".table-mount").append(t)})}).catch(e=>{document.querySelectorAll(".table-mount").forEach(x=>x.textContent="Table load failed: "+e.message);console.error(e)});const fixedPreloads=[];for(const m of ["M1","M2","M3","M4","GT"])for(const f of [0,36,72,108,144]){const image=new Image();image.src=`assets/fixed_views/${m}/${String(f).padStart(3,"0")}.png`;fixedPreloads.push(image)}const method=document.querySelector("#compare-method"),frame=document.querySelector("#compare-frame");function refresh(){const id=String(frame.value).padStart(3,"0"),m=method.value;document.querySelector("#compare-pred").src=`assets/fixed_views/${m}/${id}.png`;document.querySelector("#compare-gt").src=`assets/fixed_views/GT/${id}.png`;document.querySelector("#compare-label").textContent=`${m} / ${c.view} ${frame.value}`;document.querySelector("#figure-4").dataset.method=m;document.querySelector("#figure-4").dataset.frame=frame.value}method?.addEventListener("change",refresh);frame?.addEventListener("change",refresh);refresh();'''
+
+SCENE_JS=r'''import * as THREE from 'three';import {GLTFLoader} from './vendor/three/loaders/GLTFLoader.js';import {OrbitControls} from './vendor/three/controls/OrbitControls.js';import {RoomEnvironment} from './vendor/three/environments/RoomEnvironment.js';
+const panel=document.querySelector('#figure-3');if(panel){const stage=panel.querySelector('.scene-stage'),select=panel.querySelector('#model-select'),mode=panel.querySelector('#scene-mode'),slider=panel.querySelector('#scene-split'),divider=panel.querySelector('.scene-divider'),status=panel.querySelector('.scene-status'),left=panel.querySelector('.scene-label-left'),right=panel.querySelector('.scene-label-right'),zh=document.documentElement.lang.startsWith('zh');let started=false,active=false,renderer,camera,controls,environment,manifest,selectedScene,truthScene,resizeObserver,split=.5,version=0;const cache=new Map();const setStatus=(text,error=false)=>{status.textContent=text;status.hidden=!text;status.classList.toggle('error',error);panel.dataset.state=error?'error':text?'loading':'ready'};
+function render(){if(!renderer||!selectedScene)return;const w=stage.clientWidth,h=stage.clientHeight;renderer.setScissorTest(false);renderer.setViewport(0,0,w,h);renderer.clear();if(mode.value==='single'||select.value==='GT'||!truthScene){renderer.render(selectedScene,camera);return}const x=Math.round(w*split);renderer.setScissorTest(true);renderer.setScissor(0,0,x,h);renderer.render(selectedScene,camera);renderer.setScissor(x,0,w-x,h);renderer.render(truthScene,camera);renderer.setScissorTest(false)}
+function updateMode(){const comparing=mode.value==='compare'&&select.value!=='GT';panel.classList.toggle('is-comparing',comparing);slider.disabled=!comparing;divider.hidden=!comparing;right.hidden=!comparing;left.textContent=select.value;render()}function updateSplit(value){split=Math.max(.02,Math.min(.98,Number(value)/100));slider.value=String(Math.round(split*100));stage.style.setProperty('--split',`${split*100}%`);divider.setAttribute('aria-valuenow',slider.value);render()}
+function resetCamera(){if(!camera||!manifest)return;const v=manifest.camera;controls.target.fromArray(v.target);camera.position.fromArray(v.position).sub(controls.target).multiplyScalar(Math.max(1,1.5/(stage.clientWidth/stage.clientHeight))).add(controls.target);camera.near=v.near;camera.far=v.far;camera.fov=v.fov;camera.updateProjectionMatrix();controls.update();render()}function resize(){if(!renderer)return;const w=stage.clientWidth,h=stage.clientHeight;renderer.setSize(w,h,false);const oldFit=Math.max(1,1.5/camera.aspect);camera.aspect=w/h;if(controls)camera.position.sub(controls.target).multiplyScalar(Math.max(1,1.5/camera.aspect)/oldFit).add(controls.target);camera.updateProjectionMatrix();render()}
+function loadScene(id){if(!cache.has(id)){const promise=new GLTFLoader().loadAsync(manifest.models[id].url).then(gltf=>{const scene=new THREE.Scene();scene.background=new THREE.Color('#e9ede5');scene.environment=environment;scene.add(new THREE.HemisphereLight(0xffffff,0x68746b,2));const sun=new THREE.DirectionalLight(0xffffff,2.3);sun.position.set(12,25,18);scene.add(sun);let hiddenCutaway=0;if(id!=='GT'){const rows=manifest.models[id].runtime_matrix_gltf_y_up;gltf.scene.applyMatrix4(new THREE.Matrix4().set(...rows.flat()));gltf.scene.traverse(node=>{const name=node.name||'',isShell=/(ceiling|roof|wall|facade)/i.test(name),isDetail=/(emblem|art|grass|metal)/i.test(name);if(isShell&&!isDetail){node.visible=false;hiddenCutaway++}})}scene.add(gltf.scene);const bounds=new THREE.Box3().setFromObject(gltf.scene);scene.userData.method=id;scene.userData.hiddenCutaway=hiddenCutaway;scene.userData.bounds={min:bounds.min.toArray(),max:bounds.max.toArray()};return scene}).catch(e=>{cache.delete(id);throw e});cache.set(id,promise)}return cache.get(id)}
+async function selectScene(){if(!manifest)return;const request=++version,id=select.value;setStatus(zh?`正在加载 ${id}…`:`Loading ${id}…`);try{const [model,gt]=await Promise.all([loadScene(id),loadScene('GT')]);if(request!==version)return;selectedScene=model;truthScene=gt;setStatus('');updateMode();panel.dataset.loadedModel=id;panel.dataset.gtLoaded='true'}catch(e){if(request!==version)return;console.error('Scene asset load failed',e);setStatus(zh?'场景加载失败，请重试。':'Scene loading failed. Retry.',true)}}
+async function start(){if(started)return;started=true;try{setStatus(zh?'正在加载对比场景…':'Loading comparison scenes…');const response=await fetch('data/scene_comparison.json', {cache:'no-store'});if(!response.ok)throw Error(`manifest HTTP ${response.status}`);manifest=await response.json();renderer=new THREE.WebGLRenderer({antialias:true});renderer.setPixelRatio(Math.min(devicePixelRatio,1.5));renderer.toneMapping=THREE.ACESFilmicToneMapping;renderer.outputColorSpace=THREE.SRGBColorSpace;renderer.domElement.setAttribute('aria-label',zh?'重建与 GT 三维场景':'Reconstruction and GT 3D scene');renderer.domElement.tabIndex=0;stage.prepend(renderer.domElement);camera=new THREE.PerspectiveCamera(42,1,.05,500);controls=new OrbitControls(camera,renderer.domElement);controls.enableDamping=false;controls.minDistance=1;controls.maxDistance=200;controls.addEventListener('change',render);const pmrem=new THREE.PMREMGenerator(renderer),room=new RoomEnvironment();environment=pmrem.fromScene(room,.04).texture;room.dispose();pmrem.dispose();resizeObserver=new ResizeObserver(resize);resizeObserver.observe(stage);resize();resetCamera();panel.sceneDiagnostics=()=>({selected:select.value,loaded:selectedScene?.userData.method,gtLoaded:!!truthScene,hiddenCutaway:selectedScene?.userData.hiddenCutaway||0,bounds:selectedScene?.userData.bounds,camera:camera.position.toArray(),target:controls.target.toArray(),split,mode:mode.value,singleCamera:true,singleViewport:true,coordinateConversion:manifest.coordinate_conversion,cachedModels:[...cache.keys()]});await selectScene()}catch(e){console.error('Scene comparison initialization failed',e);resizeObserver?.disconnect();controls?.dispose();environment?.dispose();renderer?.dispose();renderer?.domElement.remove();renderer=camera=controls=environment=manifest=undefined;selectedScene=truthScene=undefined;cache.clear();started=false;setStatus(zh?'无法启动三维场景，请重试。':'Unable to start 3D viewer. Retry.',true)}}
+select.addEventListener('change',()=>{updateMode();selectScene()});mode.addEventListener('change',updateMode);slider.addEventListener('input',()=>updateSplit(slider.value));panel.querySelector('#scene-reset').addEventListener('click',resetCamera);panel.querySelector('#scene-retry').addEventListener('click',()=>manifest&&renderer?selectScene():start());divider.addEventListener('pointerdown',e=>{e.preventDefault();active=true;divider.setPointerCapture(e.pointerId);if(controls)controls.enabled=false});divider.addEventListener('pointermove',e=>{if(!active)return;const b=stage.getBoundingClientRect();updateSplit((e.clientX-b.left)/b.width*100)});const stop=()=>{active=false;if(controls)controls.enabled=true};divider.addEventListener('pointerup',stop);divider.addEventListener('pointercancel',stop);divider.addEventListener('keydown',e=>{const d={ArrowLeft:-2,ArrowRight:2,PageDown:-10,PageUp:10};if(e.key in d){e.preventDefault();updateSplit(split*100+d[e.key])}});updateSplit(50);updateMode();const lazy=new IntersectionObserver(es=>{if(es.some(e=>e.isIntersecting)){lazy.disconnect();start()}},{rootMargin:'300px'});lazy.observe(panel)}'''
+
+def academic_page(en=False):
+    current, other = ("index.html", "zh.html") if en else ("zh.html", "index.html")
+    if en:
+        title = SITE_TITLE
+        desc = "From real spaces to worlds phygital agents can use."
+        sections = [
+            ("results", "Overview"),
+            ("motivation", "Motivation"),
+            ("workflow", "Workflow"),
+            ("more-results", "Results & analysis"),
+            ("related-work", "Related work"),
+            ("limitations", "Limitations"),
+            ("citation", "Citation"),
+            ("references", "References"),
+        ]
+        copy = {
+            "hook": "A scene can look plausible and still be spatially wrong.",
+            "tldr": "We isolate how camera pose and depth evidence propagate through an agentic Blender reconstruction pipeline. Stronger geometric evidence consistently improves the frozen scene’s geometry and model depth, but appearance metrics remain split—showing that spatial fidelity and visual similarity are distinct objectives.",
+            "intro1": "Image synthesis rewards a convincing frame. A spatial system has a harder obligation: scale, location, occlusion, and object boundaries must remain coherent when the camera moves—and the result must still be editable after rendering ends.",
+            "intro2": "We reconstruct World Lobby, a simulated scene in NVIDIA Isaac Sim, through four complete engineering routes, from RGB-only modeling to ground-truth-pose-conditioned modeling. Each route produces a frozen Blender scene, not merely a point cloud or a novel-view renderer. The experiment asks what additional geometric evidence buys, where it fails to help, and which errors survive object-centric reconstruction.",
+            "contributions": [
+                "A controlled four-route study that keeps the scene, 180 RGB identities, modeling objective, and editable output format fixed while varying the geometric evidence.",
+                "A measurement contract that separates trajectory error, native depth, frozen-scene depth, surface geometry, and rendered appearance instead of collapsing them into one score.",
+                "Frozen Blend/GLB assets, deterministic evaluation subsets, public hashes, and interactive comparisons that let readers inspect where aggregate metrics agree—and where they do not.",
+            ],
+            "metric1": "0.376 → 0.072 m",
+            "metric1_label": "bidirectional surface error",
+            "metric2": "16.44% → 7.60%",
+            "metric2_label": "model-depth AbsRel · 180 evaluation views",
+            "metric3": "No universal winner",
+            "metric3_label": "PSNR, SSIM, and LPIPS disagree",
+            "motivation_title": "From a good-looking render to a usable spatial representation",
+            "motivation1": "An editable world model is useful because its objects can be selected, moved, hidden, assigned materials, queried for collisions, and revisited from known cameras. Those affordances matter for design, robotics, and simulation. They also expose errors that a single attractive image can conceal.",
+            "motivation2": "Our working hypothesis is that measurement does not replace generative reasoning; it constrains it. Pose and depth should reduce the space of plausible layouts, while the modeling system converts noisy observations into a compact semantic scene. That conversion can regularize noise, but it is also lossy: omitted surfaces and simplified materials do not reappear simply because the camera trajectory improves.",
+            "positioning": "ViPE estimates camera motion and near-metric geometry from RGB video; ORB-SLAM3 provides a classical visual-inertial trajectory; Depth Anything 3 predicts pose-conditioned multi-view depth. These systems produce measurements, not editable objects. Concurrent work AHa-3D explores video-driven agentic Real2Sim; Our emphasis is additional physical evidence, including IMU-informed visual-inertial constraints, for geometric grounding. Our question is complementary and deliberately narrower: when the modeling objective is held fixed, which measurement channel changes the final editable world—and which errors remain downstream?",
+            "method_title": "Four routes, one frozen evaluation contract",
+            "method1": "All routes receive the same 180 RGB identities and the same Astra–Blender modeling objective. M1 tests how far visual priors can go without metric measurements. M2 adds RGB-only ViPE poses and pose-conditioned DA3 depth. M3 uses monocular-inertial ORB-SLAM3 poses with the same class of DA3 observations. M4 supplies ground-truth camera poses to diagnose the remaining modeling error; it does not receive the GT mesh or GT depth as modeling input.",
+            "workflow": [
+                ("Observe", "Inspect all 180 RGB frames and, where allowed, pose-conditioned DA3 geometry."),
+                ("Build", "Write Blender Python that decomposes the room into named objects, materials, cameras, and collision proxies."),
+                ("Verify", "Compare paired RGB and optical-Z renders at fixed review cameras; revise within a fixed version budget."),
+                ("Freeze", "Export Blend/GLB, manifests, semantic IDs, and hashes before any GT score is visible to the author."),
+                ("Evaluate", "Measure pose, native depth, model depth, surface distance, and appearance under one frozen protocol."),
+            ],
+            "method2": "Ground truth is introduced only after the scenes are frozen. M2 and M3 are evaluated under one global SE(3) alignment with scale fixed to one; no mesh ICP or per-view fitting is used. M1 has no native metric trajectory, so its model-space numbers use one frozen GT-assisted Sim(3) from five manual camera associations and remain diagnostic rather than evidence of metric recovery.",
+            "results_title": "Three findings matter more than a leaderboard",
+            "finding1_title": "1. Better pose does not mechanically imply better native depth",
+            "finding1": "M3 improves trajectory accuracy over M2—ATE falls from 0.167 m to 0.121 m and rotation error from 0.31° to 0.19°. Yet its native DA3 AbsRel is slightly worse (20.38% versus 19.06%). After modeling, the ordering reverses: M3 reaches 8.37% model-depth AbsRel versus 9.29% for M2. Upstream metrics and downstream scene fidelity are related, but they are not interchangeable.",
+            "finding2_title": "2. Geometric evidence turns a plausible layout into a more faithful space",
+            "finding2": "Under the declared alignment and without mesh ICP, bidirectional surface error falls from 0.376 m for the RGB-only diagnostic baseline to 0.118 m with ViPE, 0.082 m with ORB-SLAM3, and 0.072 m with GT poses. M4 is strongest on geometry in this case, while M3 closes most of the gap using an estimated trajectory.",
+            "finding3_title": "3. Qualitative comparison",
+            "interactive_title": "Do not trust the aggregate alone—inspect the frozen worlds",
+            "interactive1": "Use the shared camera to compare each reconstruction with GT. Start with M1 to see how a semantically plausible room can drift in metric layout; compare M2 and M3 to find regions where geometry improves without a matching gain in RGB similarity; then inspect M4 to isolate errors that remain even when camera pose is no longer the bottleneck.",
+            "interactive2": "The browser hides ceilings and surrounding walls only for cutaway inspection. Runtime registration and display choices do not modify the downloadable frozen GLB or Blend files.",
+            "discussion_title": "What this case suggests about editable world models",
+            "discussion1_title": "Modeling behaves like a structured—and lossy—regularizer",
+            "discussion1": "For M2–M4, the final scene depth is substantially closer to GT than native DA3 depth on the shared 180-view protocol: AbsRel falls by roughly 51–60%. One plausible interpretation is that object-level aggregation suppresses inconsistent local predictions. This is a hypothesis, not a controlled causal result: the same abstraction can also erase real surfaces and fine geometry.",
+            "discussion2_title": "Measurement constrains generation; it does not solve appearance",
+            "discussion2": "The strongest geometric route does not dominate PSNR or SSIM. Pose removes one source of uncertainty, but appearance still depends on material estimation, lighting, object detail, and renderer mismatch. Evaluating only RGB would miss spatial progress; evaluating only geometry would miss perceptual failure.",
+            "discussion3_title": "The 100-view check measures consistency, not unseen-view generalization",
+            "discussion3": "Table 7 uses a seeded set of 100 cameras drawn from the 175 modeling frames outside the fixed-five check. These are additional evaluation views relative to that check, but their RGB images were available during modeling. On this set, M4 has the lowest depth AbsRel (7.47%), while M3 has the lowest RMSE (1.090 m).",
+            "limits_title": "The boundaries are part of the result",
+            "limits": [
+                "One synthetic scene and one engineering run per route do not establish general superiority or statistical significance.",
+                "The four routes are complete systems, not a strict single-variable ablation; M2 versus M3 cannot be attributed to IMU alone.",
+                "M4 uses GT camera poses and is a diagnostic route, not a deployable baseline or a theoretical upper bound.",
+                "M1 model-space metrics rely on GT-assisted Sim(3) and must not be ranked as native metric recovery.",
+                "The 100-view appearance and depth sets contain modeling RGB inputs; they are not a held-out novel-view benchmark.",
+                "Appearance metrics mix geometry, material, illumination, and rendering error. M4 revision 4 has not received a complete independent visual re-review.",
+            ],
+            "evidence": "Every table is generated from frozen assets. The public manifests expose model hashes, registrations, fixed views, and the exact seeded evaluation indices.",
+            "office_title": "Beyond the lobby: what survives in a phone-captured space?",
+            "office1": "Office Café is not a second run of the World Lobby benchmark and is not quantitative validation under the same protocol. It is an external, real-capture case study showing how an object-centric scene, a solved source-camera trajectory, video-to-model comparison, and physics replays can coexist in one inspectable artifact.",
+            "open_viewer": "Open full viewer",
+            "open_shake": "Open shake experiment",
+        }
+    else:
+        title = "AWSM：智能体世界仿真与建图"
+        desc = "把真实空间，变成虚实融合智能体可以使用的世界。"
+        sections = [
+            ("results", "概览"),
+            ("motivation", "研究动机"),
+            ("workflow", "工作流"),
+            ("more-results", "结果与分析"),
+            ("related-work", "相关工��"),
+            ("limitations", "局限性"),
+            ("citation", "引用"),
+            ("references", "参考文献"),
+        ]
+        copy = {
+            "hook": "一个场景可以看起来合理，却在空间上是错的。",
+            "tldr": "我们研究相机位姿与深度证据如何穿过 agentic Blender 重建流程，最终影响冻结场景。更强的几何证据持续改善模型几何与模型深度，但外观指标依然分化——空间忠实度与视觉相似度是两类不同目标。",
+            "intro1": "图像生成奖励一张令人信服的画面；空间系统承担更严格的义务：当相机移动时，尺度、位置、遮挡与物体边界仍需保持一致，而且渲染结束后，结果还必须能够被编辑。",
+            "intro2": "我们对 NVIDIA Isaac Sim 中的 World Lobby 仿真场景，沿四条完整工程路径进行重建：从纯 RGB 建模，逐步加入估计位姿、深度与真实相机位姿。每条路径交付的是冻结的 Blender 场景，而不只是点云或新视角渲染器。实验关注的不是谁生成了最漂亮的单帧，而是几何证据究竟带来了什么、没有解决什么，以及哪些误差会穿过对象化建模继续存在。",
+            "contributions": [
+                "一项受控的四路径研究：固定场景、180 张 RGB 身份、建模目标与可编辑交付格式，只改变可用的几何证据。",
+                "一份分解式测量契约：分别评估轨迹、原生深度、冻结场景深度、表面几何与渲染外观，而不是把它们压缩成一个总分。",
+                "可下载的冻结 Blend/GLB、确定性评测子集、公开哈希与交互对照，使读者能够检查平均指标在哪些区域成立、又在哪里失效。",
+            ],
+            "metric1": "0.376 → 0.072 m",
+            "metric1_label": "双向表面误差",
+            "metric2": "16.44% → 7.60%",
+            "metric2_label": "模型深度绝对相对误差（AbsRel）· 180 个评估视角",
+            "metric3": "没有单一赢家",
+            "metric3_label": "PSNR、SSIM 与 LPIPS 结论分化",
+            "motivation_title": "从“看起来不错”到“可以使用”的空间表示",
+            "motivation1": "可编辑世界模型的价值，在于其中的对象能够被选择、移动、隐藏、赋予材质、查询碰撞，并从已知相机重新访问。这些能力服务于设计、机器人与仿真，也会暴露一张漂亮图片可以掩盖的空间错误。",
+            "motivation2": "我们的工作假设是：测量不是生成推理的替代品，而是它的约束。位姿与深度应当缩小合理布局的解空间，建模系统则把噪声观测压缩为语义场景。这种转换可能正则化噪声，但同样是有损的：遗漏的表面、近似的材质，不会仅仅因为相机轨迹更准确而自动恢复。",
+            "positioning": "ViPE 从 RGB 视频估计相机运动与近米制几何，ORB-SLAM3 提供经典的视觉—惯性轨迹，Depth Anything 3 预测位姿条件下的多视图深度；它们输出的是测量，而不是可编辑对象。同期工作 AHa-3D 探索视频驱动的 agentic Real2Sim。我们强调利用额外物理证据，尤其是融合 IMU 的视觉惯性约束，进行几何锚定。本文的问题与之互补且更窄：当建模目标保持不变时，哪一种测量通道真正改变了最终可编辑世界，哪些误差仍会留在下游？",
+            "method_title": "四条路径，一份冻结评测契约",
+            "method1": "四条路径使用相同的 180 张 RGB 身份和相同的 Astra–Blender 建模目标。M1 检验没有米制测量时视觉先验能走多远；M2 加入 RGB-only ViPE 位姿与 pose-conditioned DA3 深度；M3 使用 ORB-SLAM3 单目惯性位姿与同类 DA3 观测；M4 提供 GT 相机位姿，用于诊断剩余建模误差，但建模时不读取 GT 网格或 GT 深度。",
+            "workflow": [
+                ("观察", "检查全部 180 张 RGB；在允许的方法中读取位姿条件下的 DA3 几何。"),
+                ("建模", "编写 Blender Python，将房间拆解为命名对象、材质、相机与碰撞代理。"),
+                ("验证", "在固定修订相机上比较成对 RGB 与 optical-Z 渲染，并在固定版本预算内修改。"),
+                ("冻结", "在作者看到任何 GT 评分前，导出 Blend/GLB、manifest、语义 ID 与哈希。"),
+                ("评测", "在同一冻结协议下分别测量位姿、原生深度、模型深度、表面距离与外观。"),
+            ],
+            "method2": "所有场景冻结后才引入真值。M2/M3 只使用一次全局 SE(3) 配准，尺度固定为 1；不进行 mesh ICP 或逐视角拟合。M1 没有原生米制轨迹，其模型空间指标使用由 5 个手工相机关联得到、冻结不变的 GT-assisted Sim(3)，因此只能解释形状与显示误差，不能视为米制恢复。",
+            "results_title": "比排行榜更重要的���个发现",
+            "finding1_title": "1. 更好的位姿不会机械地带来更好的原生深度",
+            "finding1": "M3 的轨迹优于 M2：ATE 从 0.167 m 降至 0.121 m，旋转误差从 0.31° 降至 0.19°；但 M3 的原生 DA3 AbsRel 反而略差（20.38% 对 19.06%）。经过建模后，排序再次反转：M3 的模型深度 AbsRel 为 8.37%，优于 M2 的 9.29%。上游指标与最终场景忠实度相关，却不能彼此替代。",
+            "finding2_title": "2. 几何证据把“合理布局”约束成更忠实的空间",
+            "finding2": "在声明配准且不使用 mesh ICP 的条件下，双向表面误差从纯 RGB 诊断基线的 0.376 m，依次降至 ViPE 的 0.118 m、ORB-SLAM3 的 0.082 m 与 GT 位姿的 0.072 m。M4 在本案例中拥有最强几何，而 M3 依靠估计轨迹已经缩小了大部分差距。",
+            "finding3_title": "3. 定性比较",
+            "interactive_title": "不要只相信平均值——亲自检查冻结场景",
+            "interactive1": "在共享相机中将每个重建与 GT 对照：先从 M1 观察语义上合理的房间如何偏离米制布局；再比较 M2/M3，寻找几何改善但 RGB 相似度未同步提升的区域；最后检查 M4，把相机位姿不再是瓶颈后仍然存在的建模与材质误差分离出来。",
+            "interactive2": "网页只在运行时隐藏顶棚与四周墙体，以便剖视检查；运行时配准与显示选择不会修改可下载的冻结 GLB 或 Blend 文件。",
+            "discussion_title": "这个案例对可编辑世界模型意味着什么",
+            "discussion1_title": "建模像一种结构化、同时有损的正则化",
+            "discussion1": "在共享的 180 视角协议下，M2–M4 的最终场景深度都显著接近 GT：相较原生 DA3，AbsRel 约下降 51–60%。一种可能解释是，对象级聚合抑制了局部不一致预测；但这仍是假说，而非受控因果结论。同一种抽象也可能删除真实表面与细节。",
+            "discussion2_title": "测量约束生成，却不会自动解决外观",
+            "discussion2": "几何最强的路径并未统治 PSNR 或 SSIM。位姿消除了一类不确定性，外观仍受材质估计、照明、物体细节与渲染器差异支配。只评 RGB 会漏掉空间进步，只评几何也会漏掉感知失败。",
+            "discussion3_title": "100 视角检查衡量一致性，而不是未见视角泛化",
+            "discussion3": "Table 7 使用一组固定随机种子的 100 个相机，它们从固定 5 帧之外的 175 张建模帧中抽取。这些视角相对固定检查是额外视角，但其 RGB 在建模阶段可见。在该集合上，M4 的深度 AbsRel 最低（7.47%），M3 的 RMSE 最低（1.090 m）。",
+            "limits_title": "边界本身也是结果的一部分",
+            "limits": [
+                "一个合成场景、每条路径一次工程运行，不能建立一般优越性或统计显著性。",
+                "四条路径是完整系统，而非严格的单变量消融；M2 与 M3 的差异不能只归因于 IMU。",
+                "M4 使用 GT 相机位姿，是诊断路线，不是可部署基线，也不是理论上界。",
+                "M1 模型空间指标依赖 GT-assisted Sim(3)，不能与原生米制恢复混为一谈。",
+                "100 视角外观与深度集合包含建模 RGB 输入，不是 held-out novel-view benchmark。",
+                "外观指标混合了几何、材质、照明与渲染误差；M4 revision 4 尚未完成完整独立视觉复审。",
+            ],
+            "evidence": "所有表格都由冻结资产生成；公开 manifest 提供模型哈希、配准、固定视角与固定随机评测索引。",
+            "office_title": "走出仿真大厅：手机采集空间中还剩下什么？",
+            "office1": "Office Café 不是 World Lobby 基准的第二次运行，也不构成同协议的定量验证。它是一个真实采集的外部案例，用于展示对象化场景、求解后的原视频相机轨迹、视频—模型对照与物理回放如何组合成一个可检查的空间产物。",
+            "open_viewer": "全屏打开模型",
+            "open_shake": "打开摇晃实验",
+        }
+    narrative_copy(copy, en)
+    toc = "".join(f'<a href="#{anchor}">{label}</a>' for anchor, label in sections)
+    limits = "".join(f"<li>{item}</li>" for item in copy["limits"])
+    contributions = "".join(f"<li>{item}</li>" for item in copy["contributions"])
+    workflow = "".join(
+        f'<div><strong>{index:02d} · {name}</strong><span>{description}</span></div>'
+        for index, (name, description) in enumerate(copy["workflow"], 1)
+    )
+    related_links = (
+        '<span class="inline-sources">'
+        '<a href="https://research.nvidia.com/labs/toronto-ai/vipe/" target="_blank" rel="noopener">ViPE</a> · '
+        '<a href="https://arxiv.org/abs/2007.11898" target="_blank" rel="noopener">ORB-SLAM3</a> · '
+        '<a href="https://arxiv.org/abs/2511.10647" target="_blank" rel="noopener">Depth Anything 3</a> · '
+        '<a href="https://kevinxu02.github.io/real2sim-indoor-site/" target="_blank" rel="noopener">AHa-3D</a>'
+        '</span>'
+    )
+    citation = html.escape(BIBTEX.rstrip())
+    references = [
+        ('AHa-3D: Agentic Tool Use for Real2Sim with GPT-6 Astra.',
+         'https://kevinxu02.github.io/real2sim-indoor-site/'),
+        ('ViPE: Video Pose Engine for camera motion and near-metric geometry.',
+         'https://arxiv.org/abs/2508.10934'),
+        ('Campos et al. ORB-SLAM3: An Accurate Open-Source Library for Visual, Visual–Inertial, and Multimap SLAM. IEEE Transactions on Robotics, 2021.',
+         'https://arxiv.org/abs/2007.11898'),
+        ('Depth Anything 3: Recovering the Visual Space from Any Views.',
+         'https://arxiv.org/abs/2511.10647'),
+        ('Wang et al. Image Quality Assessment: From Error Visibility to Structural Similarity. IEEE Transactions on Image Processing, 2004.',
+         'https://doi.org/10.1109/TIP.2003.819861'),
+        ('Zhang et al. The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. CVPR, 2018.',
+         'https://arxiv.org/abs/1801.03924'),
+        ('Blender Foundation. Blender: Free and Open Source 3D Creation Suite.',
+         'https://www.blender.org/'),
+    ]
+    references.extend(REFERENCES)
+    reference_list = "".join(
+        f'<li><a href="{url}" target="_blank" rel="noopener">{text}</a></li>'
+        for text, url in references
+    )
+    page = f'''<!doctype html><html lang="{"en" if en else "zh-CN"}"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"><title>{title} — Phygital AI</title><meta name="description" content="{desc}"><meta property="og:type" content="article"><meta property="og:locale" content="{"en_US" if en else "zh_CN"}"><meta property="og:title" content="{title}"><meta property="og:description" content="{desc}"><meta property="og:image" content="{BASE_URL}assets/teaser_originals.png"><meta property="og:url" content="{BASE_URL}{current}"><meta name="twitter:card" content="summary_large_image"><meta name="twitter:title" content="{title}"><meta name="twitter:description" content="{desc}"><meta name="twitter:image" content="{BASE_URL}assets/teaser_originals.png"><link rel="canonical" href="{BASE_URL}{current}"><link rel="alternate" hreflang="zh-CN" href="{BASE_URL}index.html"><link rel="alternate" hreflang="en" href="{BASE_URL}en.html"><link rel="stylesheet" href="style.css"><script type="importmap">{{"imports":{{"three":"./vendor/three/three.module.js"}}}}</script><script src="app.js" defer></script><script type="module" src="scene-compare.js"></script></head>
+<body data-lang="{"en" if en else "zh"}"><a class="skip" href="#main">{"Skip to article" if en else "跳到正文"}</a><nav><a class="brand" href="index.html">ASTRA / WORLD MODELS</a><div><a href="index.html" {"aria-current='page'" if not en else ""}>中文</a><a href="en.html" {"aria-current='page'" if en else ""}>EN</a><a href="{CODE_URL}" target="_blank" rel="noopener noreferrer">GitHub ↗</a></div></nav><header><p class="eyebrow">RESEARCH ESSAY · EDITABLE WORLD MODELS</p><h1>{title}</h1><p class="lead">{desc}</p><div class="meta"><span>{"1 controlled scene" if en else "1 个受控场景"}</span><span>{"4 reconstruction routes" if en else "4 条重建路径"}</span><span>{"editable Blender outputs" if en else "可编辑 Blender 交付"}</span></div></header>
+{hero_figure(en)}<div class="layout"><aside class="toc"><p>ON THIS PAGE</p>{toc}</aside><main id="main">
+<section id="results"><p class="section-tag">01 / RESULTS</p><aside class="tldr"><strong>TL;DR</strong><p>{copy["tldr"]}</p></aside><h2>{copy["hook"]}</h2><p class="standfirst">{copy["intro1"]}</p><p>{copy["intro2"]}</p><div class="insight-strip"><div><strong>{copy["metric1"]}</strong><span>{copy["metric1_label"]}</span><small class="metric-reduction">{copy["metric1_reduction"]}</small></div><div><strong>{copy["metric2"]}</strong><span>{copy["metric2_label"]}</span><small class="metric-reduction">{copy["metric2_reduction"]}</small></div></div><p class="metric-caveat">{copy["metric_scope"]}</p></section>
+<section id="motivation"><p class="section-tag">02 / MOTIVATION</p><h2>{copy["motivation_title"]}</h2><p>{copy["motivation1"]}</p><p>{copy["motivation2"]}</p><aside class="claim"><strong>{"Research question" if en else "研究问题"}</strong><p>{"How does progressively stronger geometric evidence change the fidelity—and the remaining failure modes—of an editable object-centric world?" if en else "逐步增强的几何证据，如何改变一个对象化可编辑世界的忠实度，以及它仍然保留的失败模式？"}</p></aside><h3>{"Contributions" if en else "本文贡献"}</h3><ol class="contributions">{contributions}</ol></section>
+<section id="workflow"><p class="section-tag">03 / WORKFLOW</p><h2>{copy["method_title"]}</h2><p>{copy["method_intro"]}</p><p>{copy["method1"]}</p>{table(1)}<h3>{"Agentic reconstruction loop" if en else "Agentic 重建闭环"}</h3><div class="workflow-strip">{workflow}</div><p>{copy["method2"]}</p><aside class="notice"><strong>M1 diagnostic limitation.</strong> {"Its Sim(3)-aligned geometry and depth numbers describe shape after GT-assisted display alignment, not native metric recovery." if en else "其 Sim(3) 配准后的几何与深度数值描述 GT-assisted 显示配准后的形状，不代表原生米制恢复。"}</aside></section>
+<section id="more-results"><p class="section-tag">04 / MORE RESULTS AND ANALYSIS</p><h2>{copy["results_title"]}</h2><article class="finding"><p class="finding-kicker">FINDING 01</p><h3>{copy["finding1_title"]}</h3><p>{copy["finding1"]}</p></article>{table(2)}<figure id="figure-1"><img loading="lazy" src="assets/figure14_abc.png" alt="Trajectory, native depth, and model depth"><figcaption><span>Figure 1.</span> {"The upstream trajectory and native-depth ranking does not map one-to-one onto frozen-scene depth." if en else "上游轨迹与原生深度的排序，并不会一一映射为冻结场景的深度排序。"}</figcaption></figure><article class="finding"><p class="finding-kicker">FINDING 02</p><h3>{copy["finding2_title"]}</h3><p>{copy["finding2"]}</p></article>{table(3)}<article class="finding finding-comparison"><p class="finding-kicker">FINDING 03</p><h3>{copy["finding3_title"]}</h3></article><div class="analysis-block"><h3>{copy["interactive_title"]}</h3><p>{copy["interactive1"]}</p>{overview_figure(en)}{scene_figure(en)}{fixed_figure(en)}<p>{copy["interactive2"]}</p></div><div class="analysis-block"><h3>{copy["discussion_title"]}</h3><h4>{copy["discussion1_title"]}</h4><p>{copy["discussion1"]}</p><h4>{copy["discussion2_title"]}</h4><p>{copy["discussion2"]}</p><h4>{copy["discussion3_title"]}</h4><p>{copy["discussion3"]}</p><figure id="figure-5"><img loading="lazy" src="assets/model_depth_error_five_views_m1_m4.png" alt="Model-depth error at five fixed evaluation views"><figcaption><span>Figure 5.</span> {"Model-depth error at the fixed-five check; teal marks missing or invalid predictions." if en else "固定 5 视角检查中的模型深度误差；青绿色表示缺失或无效预测。"}</figcaption></figure>{table(6)}{table(7)}</div>{office_case(en)}</section>
+<section id="related-work"><p class="section-tag">05 / RELATED WORK</p><h2>{"From geometric measurement to editable Real2Sim" if en else "从几何测量到可编辑 Real2Sim"}</h2><p>{copy["positioning"]}</p><p>{"Our evaluation therefore spans both sides of the interface: upstream systems are judged as measurement pipelines, while the Blender outputs are judged as persistent spatial artifacts. This differs from evaluating pose, depth, or novel-view synthesis in isolation." if en else "因此，我们的评测同时覆盖接口两端：上游系统作为测量管线接受检验，Blender 输出则作为持久化空间产物接受检验；这不同于孤立地评估位姿、深度或新视角合成。"}</p>{related_links}</section>
+<section id="limitations"><p class="section-tag">06 / LIMITATIONS</p><h2>{copy["limits_title"]}</h2><ul class="limits-list">{limits}</ul><p>{copy["evidence"]}</p>{downloads()}<p class="evidence-links"><a href="data/tables_1_7.json">Tables JSON</a> · <a href="data/scene_comparison.json">Scene manifest</a> · <a href="data/fixed_views.json">Fixed views</a> · <a href="evidence/publication_manifest.json">Publication manifest</a> · <a href="evidence/SHA256SUMS">SHA256</a></p></section>
+<section id="citation"><p class="section-tag">07 / CITATION</p><h2>{"Cite this work" if en else "引用本文"}</h2><p>{"If this controlled study, its frozen assets, or its evaluation protocol supports your work, please cite the interactive article below." if en else "如果本研究的受控实验、冻结资产或评测协议对你的工作有帮助，请引用下面的交互式研究文章。"}</p><pre class="citation-block"><code>{citation}</code></pre></section>
+<section id="references"><p class="section-tag">08 / REFERENCES</p><h2>{"References" if en else "参考文献"}</h2><ol class="references-list">{reference_list}</ol></section><noscript>JavaScript is required for tables and interactive figures.</noscript></main></div><footer>Agentic World · frozen evidence, editable outputs · <a href="{other}">{"中文" if en else "English"}</a></footer></body></html>'''
+
+
+    return editorial_page(page, en)
+
+
+README='''# Agentic World Simulation and Mapping
+
+Bilingual static research journal rebuilt from frozen local sources. Figure 3 uses one local Three.js renderer/camera/full viewport for split comparison of M1–M4 with GT. Figure 4 switches among 25 hash-verified source images. All viewer dependencies are local, with no CDN.
+
+The build copies the unchanged `scene.glb`/`scene.blend` pairs, downloads and verifies `GT.glb`, copies five local Three.js modules, applies registrations only at runtime, and records provenance in `data/scene_comparison.json`.
+
+```bash
+python scripts/build.py --source ../world_lobby_four_trajectory_20260929
+python scripts/validate.py
+python -m http.server 8765
+python scripts/browser_qa.py --url http://127.0.0.1:8765/
+```
+'''
+
+def main():
+    ap=argparse.ArgumentParser(); ap.add_argument("--source",type=Path,default=DEFAULT_SOURCE)
+    source=ap.parse_args().source.resolve(); clean()
+    table_src=source/TABLE_SOURCE; tables=json.loads(table_src.read_text())
+    if list(tables["tables"]) != [f"table{i}" for i in range(1,8)]: raise ValueError("exactly table1..table7 required")
+    for key,rows in tables["tables"].items():
+        if [r["method"] for r in rows] != list(METHODS): raise ValueError(f"{key}: exactly M1..M4 required")
+    assets={}
+    for name in ("figure14_abc.png","fixed_five_view_comparison_m1_m4.jpg","model_depth_error_five_views_m1_m4.png"):
+        rel=f"astra_blender2/report/space/assets/{name}"; assets[f"assets/{name}"]=copy_checked(source/rel,ROOT/f"assets/{name}",source_label=rel)
+    teaser_rel="teaser_originals.png"
+    assets["assets/teaser_originals.png"]=copy_checked(source/teaser_rel,ROOT/"assets/teaser_originals.png",source_label=teaser_rel)
+    models={}
+    for method in METHODS:
+        models[method]={}
+        for ext in ("glb","blend"):
+            rel=f"astra_blender/models/{method}/scene.{ext}"; models[method][ext]=copy_checked(source/rel,ROOT/f"models/{method}/scene.{ext}",source_label=rel)
+        dump(ROOT/f"evidence/{method}/manifest.json",{"schema_version":2,"method":method,"files":models[method],"table_rows":{k:next(r for r in v if r["method"]==method) for k,v in tables["tables"].items()}})
+    gt=download_gt()
+    vendor={}
+    for rel in ("three.module.js","three.core.js","loaders/GLTFLoader.js","controls/OrbitControls.js","environments/RoomEnvironment.js","utils/BufferGeometryUtils.js"):
+        label=(REFERENCE/"vendor/three"/rel).relative_to(ROOT.parent).as_posix()
+        vendor[rel]=copy_checked(REFERENCE/"vendor/three"/rel,ROOT/"vendor/three"/rel,source_label=label)
+    c=[[1,0,0,0],[0,0,1,0],[0,-1,0,0],[0,0,0,1]]; ci=[[1,0,0,0],[0,0,-1,0],[0,1,0,0],[0,0,0,1]]
+    scene_models={"GT":{"url":"assets/comparison/GT.glb","sha256":GT_SHA,"bytes":GT_BYTES,"registration":"identity"}}
+    registrations={}
+    for method in METHODS:
+        if method=="M4":
+            blender=[[1 if i==j else 0 for j in range(4)] for i in range(4)]
+            src={"path":None,"sha256":None,"transform_direction":"GT_from_model","rule":"identity_world; source GLB retains Blender Z-up axes"}
+            gltf=c
+            formula="C (axis conversion only; GT_from_model is identity)"
+        else:
+            rel=f"{REG_DIR}/registration_{method}.json"; p=source/rel; reg=json.loads(p.read_text())
+            if reg.get("transform_direction")!="GT_from_model": raise ValueError(f"{method}: wrong direction")
+            blender=reg["transform"]; src={"path":rel,"sha256":sha(p),"transform_direction":"GT_from_model","status":reg.get("status"),"scope":reg.get("scope"),"limitations":reg.get("limitations")}
+            gltf=mm(mm(c,blender),ci)
+            formula="C * GT_from_model * C^-1"
+        registrations[method]={"source":src,"GT_from_model_blender_z_up":blender,"runtime_matrix_gltf_y_up":gltf,"formula":formula,"diagnostic_caveat":("M1 only: GT-assisted Sim(3) from five manual camera associations; diagnostic only; not native metric recovery." if method=="M1" else None)}
+        scene_models[method]={"url":f"models/{method}/scene.glb","sha256":models[method]["glb"]["sha256"],"bytes":models[method]["glb"]["bytes"],"runtime_matrix_gltf_y_up":gltf,"source_registration":src["path"]}
+    dump(ROOT/"data/scene_comparison.json",{"schema_version":2,"models":scene_models,"camera":{"target":[18,1.2,-21],"position":[24,23,-5],"fov":42,"near":.05,"far":500},"coordinate_conversion":{"reason":"glTF is Y-up while registration is Blender Z-up","C_mapping":"[x,y,z] -> [x,z,-y]","C":c,"C_inverse":ci,"runtime_formula":"C * T * C^-1"},"registrations":registrations,"display_cutaway":{"scope":"runtime display only; frozen GLB/Blend bytes remain unchanged","hide_name_pattern":"ceiling|roof|wall|facade","keep_name_pattern":"emblem|art|grass|metal","applies_to":list(METHODS),"gt_asset":"already exported as a cutaway display model"},"immutability":"Display copies and runtime registration do not alter frozen source GLB bytes.","m1_diagnostic_caveat":"M1 registration is diagnostic only and is not native metric recovery.","gt_asset":gt})
+    view_src=source/VIEW_SOURCE; inputs=json.loads(view_src.read_text())["rgb_figure"]["inputs"]
+    if len(inputs)!=25: raise ValueError("fixed-view input count is not 25")
+    fixed=[]
+    for item in inputs:
+        if item["method"] not in (*METHODS,"GT") or item["frame"] not in FRAMES: raise ValueError(item)
+        source_path=Path(item["path"])
+        source_label=source_path.relative_to(ROOT.parent).as_posix()
+        copied=copy_checked(source_path,ROOT/f"assets/fixed_views/{item['method']}/{item['frame']:03d}.png",item["sha256"],source_label=source_label)
+        fixed.append({"method":item["method"],"frame":item["frame"],"source":source_label,"published":copied["published"],"sha256":item["sha256"],"bytes":copied["bytes"]})
+    if {(x["method"],x["frame"]) for x in fixed}!={(m,f) for m in (*METHODS,"GT") for f in FRAMES}: raise ValueError("fixed-view matrix incomplete")
+    dump(ROOT/"data/fixed_views.json",{"schema_version":1,"source_manifest":VIEW_SOURCE,"source_manifest_sha256":sha(view_src),"frames":list(FRAMES),"methods":[*METHODS,"GT"],"images":fixed})
+    (ROOT/"data").mkdir(parents=True,exist_ok=True); shutil.copy2(table_src,ROOT/"data/tables_1_7.json")
+    dump(ROOT/"data/source_contract.json",{"schema_version":2,"old_site_baseline":"ae69dbe","source_root_name":source.name,"table_source":{"path":TABLE_SOURCE,"sha256":sha(table_src),"published_url":"data/tables_1_7.json"},"figures":assets,"models":models,"gt_asset":gt,"fixed_views_manifest":"data/fixed_views.json","scene_comparison_manifest":"data/scene_comparison.json","office_cafe_case_study":{"office_cafe":{"source_url":OFFICE_URL,"presentation":"local_models_and_synchronized_original_videos","included_assets":["original_model","legacy_model","input_video","procedural_model_video","scan_video","strong_shake"],"video_manifest":"data/office_video_media.json"}},"vendor":vendor,"tables":[{"number":i,"key":f"table{i}","methods":list(METHODS),"columns":list(tables["tables"][f"table{i}"][0]),"url":f"index.html#table-{i}","english_url":f"en.html#table-{i}"} for i in range(1,8)],"presentation":{"table2_hidden_columns":["alignment"],"table3_hidden_columns":["model_sha256","alignment"],"figure1_layout":"single supplied teaser image","metric_best_values":"bold, raw-value comparison"},"protocols":tables["protocols"],"m1_notice":tables["m1_notice"]})
+    academic_style=r'''#figure-2,#figure-5{width:100%;margin:30px 0}#figure-2>img,#figure-5>img{width:100%;background:var(--paper)}main h3{font-family:Georgia,"Noto Serif CJK SC",serif;font-size:23px;font-weight:500;line-height:1.4;margin:32px 0 10px}main h4{font:500 18px/1.45 Georgia,"Noto Serif CJK SC",serif;margin:28px 0 8px}.tldr{margin:0 0 34px;padding:20px 24px;border-left:4px solid var(--green);background:#edf3ee}.tldr strong{font-size:11px;letter-spacing:.16em;color:var(--green)}.tldr p{margin:5px 0 0;font:500 19px/1.65 Georgia,"Noto Serif CJK SC",serif}.contributions{padding-left:24px}.contributions li{margin:10px 0;padding-left:5px}.inline-sources{display:block;margin-top:12px;font-size:13px}.workflow-strip{display:grid;grid-template-columns:repeat(5,1fr);gap:1px;margin:22px 0 34px;border:1px solid var(--line);background:var(--line)}.workflow-strip div{display:flex;min-height:180px;padding:18px 15px;background:var(--white);flex-direction:column}.workflow-strip strong{margin-bottom:14px;font-size:12px;letter-spacing:.08em;color:var(--green)}.workflow-strip span{font-size:13px;line-height:1.6;color:var(--muted)}.insight-strip{display:grid;grid-template-columns:repeat(3,1fr);gap:1px;margin:34px 0 12px;background:var(--line);border:1px solid var(--line)}.insight-strip div{display:flex;min-height:125px;padding:22px 18px;background:var(--white);flex-direction:column;justify-content:space-between}.insight-strip strong{font:500 25px/1.15 Georgia,serif;color:var(--green);letter-spacing:-.02em}.insight-strip span{font-size:11px;line-height:1.45;color:var(--muted);text-transform:uppercase;letter-spacing:.07em}.metric-caveat{margin:8px 0 0;color:var(--muted);font-size:12px}.claim{margin:30px 0;padding:20px 24px;border:1px solid var(--line);background:var(--white)}.claim strong,.finding-kicker{font-size:11px;letter-spacing:.13em;text-transform:uppercase;color:var(--green)}.claim p{margin:8px 0 0;font:500 21px/1.55 Georgia,"Noto Serif CJK SC",serif}.finding{margin:42px 0 18px;padding:0 0 22px;border-bottom:1px solid var(--line)}.finding h3{margin:5px 0 12px}.finding p:last-child{margin-bottom:0}.analysis-block{margin-top:58px;padding-top:12px;border-top:1px solid var(--line)}.external-case{margin-bottom:10px}.limits-list{padding-left:22px}.limits-list li{margin:10px 0}.evidence-links{font-size:12px;margin-top:25px}.citation-block{overflow:auto;margin:22px 0;padding:22px;border:1px solid var(--line);background:var(--white);font:13px/1.65 ui-monospace,SFMono-Regular,Consolas,monospace}.references-list{padding-left:24px}.references-list li{margin:13px 0;padding-left:6px}.references-list a{color:var(--ink)}@media(max-width:900px){.workflow-strip{grid-template-columns:1fr 1fr}.workflow-strip div{min-height:130px}}@media(max-width:700px){.insight-strip,.workflow-strip{grid-template-columns:1fr}.insight-strip div{min-height:98px}.workflow-strip div{min-height:auto}.claim p{font-size:18px}.tldr{padding:17px 19px}.tldr p{font-size:17px}.citation-block{padding:15px;font-size:11px}}'''
+    eager_preload='const fixedPreloads=[];for(const m of ["M1","M2","M3","M4","GT"])for(const f of [0,36,72,108,144]){const image=new Image();image.src=`assets/fixed_views/${m}/${String(f).padStart(3,"0")}.png`;fixedPreloads.push(image)}'
+    controls='const method=document.querySelector("#compare-method"),frame=document.querySelector("#compare-frame");'
+    if eager_preload not in APP or controls not in APP: raise ValueError("Figure 4 app template changed")
+    app_source=APP.replace(eager_preload,"").replace(controls,controls+"method.disabled=false;frame.disabled=false;")
+    write(ROOT/"style.css",STYLE+academic_style+"\n"); write(ROOT/"app.js",app_source+"\n"); write(ROOT/"scene-compare.js",SCENE_JS+"\n"); write(ROOT/"index.html",academic_page()+"\n"); write(ROOT/"en.html",academic_page(True)+"\n"); write(ROOT/"README.md",README); write(ROOT/"requirements.txt","playwright>=1.40,<2\n"); write(ROOT/".gitignore","__pycache__/\nqa/\n.venv/\n.publication.json\n"); (ROOT/".nojekyll").touch()
+    files=[]
+    for p in sorted(ROOT.rglob("*")):
+        if not p.is_file(): continue
+        rel=p.relative_to(ROOT)
+        if rel.parts[0] in (".git","scripts") or rel.name in ("SHA256SUMS","publication_manifest.json") or (len(rel.parts)==1 and rel.name.startswith("office-") and rel.suffix==".mp4") or rel.as_posix() in ("README.md","requirements.txt",".gitignore"): continue
+        files.append({"path":rel.as_posix(),"sha256":sha(p),"bytes":p.stat().st_size})
+    dump(ROOT/"evidence/publication_manifest.json",{"schema_version":2,"source_contract":"data/source_contract.json","files":files})
+    sums=[f'{sha(ROOT/r["path"])}  {r["path"]}' for r in files]; sums.append(f'{sha(ROOT/"evidence/publication_manifest.json")}  evidence/publication_manifest.json'); write(ROOT/"evidence/SHA256SUMS","\n".join(sums)+"\n")
+    print(json.dumps({"status":"BUILT","tables":7,"models":4,"fixed_views":25,"files":len(files)}))
+if __name__=="__main__": main()

scripts/site_validator.py

@@ -0,0 +1,6 @@
+#!/usr/bin/env python3
+"""Validate the interactive publication and frozen-source provenance."""
+import runpy
+from pathlib import Path
+
+runpy.run_path(str(Path(__file__).with_name("site_validator_interactive.py")), run_name="__main__")

scripts/site_validator_interactive.py

@@ -0,0 +1,175 @@
+#!/usr/bin/env python3
+"""Validate interactive publication structure, provenance, and hashes."""
+from pathlib import Path
+from html.parser import HTMLParser
+from urllib.parse import urlsplit, unquote
+import hashlib, json, re
+
+ROOT=Path(__file__).resolve().parents[1]
+SOURCE=ROOT.parent/"world_lobby_four_trajectory_20260929"
+METHODS=("M1","M2","M3","M4"); FRAMES=(0,36,72,108,144)
+GT_SHA="6c28ab067c63c5b1fce19c6a972f3d66c7ac4d9bd9ce3b5a57e2e18f5641fe8f"
+GT_BYTES=24816108
+def sha(path): return hashlib.sha256(path.read_bytes()).hexdigest()
+def mm(a,b): return [[sum(a[i][k]*b[k][j] for k in range(4)) for j in range(4)] for i in range(4)]
+
+class Doc(HTMLParser):
+    def __init__(self,path):
+        super().__init__(); self.path=path; self.ids=set(); self.links=[]; self.tags=[]; self.meta=[]; self.canon=[]
+        self.feed(path.read_text(encoding="utf-8"))
+    def handle_starttag(self,tag,attrs):
+        a=dict(attrs); self.tags.append(tag)
+        if "id" in a:
+            assert a["id"] not in self.ids, f"duplicate id {a['id']}"; self.ids.add(a["id"])
+        for key in ("href","src"):
+            if key in a: self.links.append(a[key])
+        if tag=="meta" and a.get("property","").startswith("og:"): self.meta.append(a["property"])
+        if tag=="link" and a.get("rel")=="canonical": self.canon.append(a.get("href"))
+
+tables=json.loads((ROOT/"data/tables_1_7.json").read_text())
+assert list(tables["tables"])==[f"table{i}" for i in range(1,8)]
+for key,rows in tables["tables"].items(): assert [r["method"] for r in rows]==list(METHODS),key
+m4_table2=next(row for row in tables["tables"]["table2"] if row["method"]=="M4")
+assert m4_table2["ate_m"]=="-" and m4_table2["rotation_deg"]=="-"
+docs={name:Doc(ROOT/name) for name in ("index.html","en.html")}
+required={"results","motivation","workflow","more-results","related-work","limitations","citation","references","office-cafe",
+          "model-select","scene-mode","scene-reset","scene-split","scene-retry","compare-method","compare-frame",
+          *{f"table-{i}" for i in range(1,8)},*{f"figure-{i}" for i in range(1,7)}}
+for name,doc in docs.items():
+    assert required<=doc.ids,f"{name}: missing {required-doc.ids}"
+    assert "model-viewer" not in doc.tags and "model-viewer" not in (ROOT/name).read_text().lower()
+    assert len(doc.canon)==1 and doc.canon[0].startswith("https://")
+    assert {"og:title","og:description","og:image","og:url"}<=set(doc.meta)
+    for link in doc.links:
+        u=urlsplit(link)
+        if u.scheme or u.netloc: continue
+        target=(ROOT/unquote(u.path)).resolve() if u.path else ROOT/name
+        assert target.is_relative_to(ROOT) and target.exists(),f"broken link {name}: {link}"
+        if u.fragment and target.suffix==".html":
+            target_doc=docs.get(target.name) or Doc(target)
+            assert unquote(u.fragment) in target_doc.ids,f"broken anchor {link}"
+    html=(ROOT/name).read_text()
+    assert all(f"Figure {i}." in html for i in range(1,7))
+    assert html.index('id="table-2"') < html.index('id="figure-2"')
+    assert html.index('id="figure-4"') < html.index('id="figure-5"')
+    assert 'assets/fixed_views/GT/000.png' in html and 'assets/fixed_views/M1/000.png' in html
+    assert 'src="https://office-cafe-vipe.hiwtishere.chatgpt.site/"' in html
+    assert 'href="https://office-cafe-vipe.hiwtishere.chatgpt.site/shake.html"' in html
+    assert 'type="importmap"' in html and 'scene-compare.js' in html
+    assert 'name="twitter:card" content="summary_large_image"' in html
+    assert 'name="twitter:image" content="https://wentingw.github.io/agentic-world-blog/assets/teaser_originals.png"' in html
+    assert '<option value="M4" selected>M4</option>' in html
+    section_order=[html.index(f'<section id="{section}"') for section in (
+        "results","motivation","workflow","more-results","related-work","limitations","citation","references")]
+    assert section_order==sorted(section_order)
+    assert 'class="citation-block"' in html and 'class="references-list"' in html
+
+assert not (ROOT/"vendor/model-viewer-4.1.0.min.js").exists()
+published_runtime="\n".join((ROOT/name).read_text(errors="ignore")
+    for name in ("index.html","en.html","app.js","scene-compare.js","style.css"))
+assert "model-viewer" not in published_runtime.lower()
+zh=(ROOT/"index.html").read_text()
+en=(ROOT/"en.html").read_text()
+assert "M4 revision 4 尚未完成完整独立视觉复审" in zh
+assert "运行时配准与显示选择不会修改可下载的冻结 GLB 或 Blend 文件" in zh
+assert "runtime registration and display choices do not modify the downloadable frozen glb or blend files" in en.lower()
+assert "100 视角检查衡量一致性，而不是未见视角泛化" in zh
+assert "the 100-view check measures consistency, not unseen-view generalization" in en.lower()
+assert "一个场景可以看起来合理，却在空间上是错的" in zh
+assert "a scene can look plausible and still be spatially wrong" in en.lower()
+for html in (zh,en):
+    assert 'class="tldr"' in html
+    assert 'class="contributions"' in html
+    assert html.count('class="workflow-strip"')==1
+    assert html.count('<li><a href=')>=7
+    assert all(url in html for url in (
+        "https://research.nvidia.com/labs/toronto-ai/vipe/",
+        "https://arxiv.org/abs/2007.11898",
+        "https://arxiv.org/abs/2511.10647",
+        "https://kevinxu02.github.io/real2sim-indoor-site/"))
+assert all(stage in zh for stage in ("观察","建模","验证","冻结","评测"))
+assert all(stage in en for stage in ("Observe","Build","Verify","Freeze","Evaluate"))
+
+contract=json.loads((ROOT/"data/source_contract.json").read_text())
+assert contract["presentation"]["table3_hidden_columns"]==["model_sha256","alignment"]
+assert contract["presentation"]["table2_hidden_columns"]==["alignment"]
+assert contract["presentation"]["figure1_layout"]=="single supplied teaser image"
+assert contract["external_embed"]["office_cafe"]["retained_features"]==[
+    "model","scan_to_model","source_video_to_model","source_camera_trajectory_playback",
+    "shake_control","shake_gentle","shake_strong"]
+src_table=SOURCE/contract["table_source"]["path"]
+assert sha(src_table)==contract["table_source"]["sha256"]==sha(ROOT/"data/tables_1_7.json")
+for published,item in contract["figures"].items():
+    assert sha(ROOT/published)==item["sha256"]==sha(SOURCE/item["source"])
+for method,files in contract["models"].items():
+    for item in files.values():
+        assert sha(ROOT/item["published"])==item["sha256"]==sha(SOURCE/item["source"])
+assert len(list((ROOT/"models").glob("M*/scene.glb")))==4
+assert len(list((ROOT/"models").glob("M*/scene.blend")))==4
+
+gt=ROOT/"assets/comparison/GT.glb"
+assert gt.stat().st_size==GT_BYTES and sha(gt)==GT_SHA
+scene=json.loads((ROOT/"data/scene_comparison.json").read_text())
+assert scene["gt_asset"]["source_url"]=="https://wentingw.github.io/astra-world-model-blog/assets/comparison/GT.glb"
+assert scene["gt_asset"]["sha256"]==GT_SHA and scene["gt_asset"]["bytes"]==GT_BYTES
+assert scene["camera"]=={"target":[18,1.2,-21],"position":[18,10,-3],"fov":42,"near":0.05,"far":500}
+c=[[1,0,0,0],[0,0,1,0],[0,-1,0,0],[0,0,0,1]]; ci=[[1,0,0,0],[0,0,-1,0],[0,1,0,0],[0,0,0,1]]
+assert scene["coordinate_conversion"]["C_mapping"]=="[x,y,z] -> [x,z,-y]"
+assert scene["display_cutaway"]["hide_name_pattern"]=="ceiling|roof|wall|facade"
+assert scene["display_cutaway"]["keep_name_pattern"]=="mirror|emblem|art|grass|metal"
+for method in METHODS:
+    reg=scene["registrations"][method]
+    if method=="M4":
+        assert reg["source"]["rule"]=="identity_world; source GLB retains Blender Z-up axes"
+        assert reg["source"]["path"] is None
+        assert reg["formula"]=="C (axis conversion only; GT_from_model is identity)"
+        assert reg["runtime_matrix_gltf_y_up"]==c
+    else:
+        assert reg["formula"]=="C * GT_from_model * C^-1"
+        expected=f"astra_blender2/evaluation/reference_style_figures/registration_{method}.json"
+        source=SOURCE/expected; raw=json.loads(source.read_text())
+        assert reg["source"]["path"]==expected and reg["source"]["sha256"]==sha(source)
+        assert reg["source"]["transform_direction"]=="GT_from_model"
+        assert reg["GT_from_model_blender_z_up"]==raw["transform"]
+        assert reg["runtime_matrix_gltf_y_up"]==mm(mm(c,reg["GT_from_model_blender_z_up"]),ci)
+assert "diagnostic only" in scene["m1_diagnostic_caveat"].lower()
+
+fixed=json.loads((ROOT/"data/fixed_views.json").read_text())
+source_views=SOURCE/fixed["source_manifest"]
+assert fixed["source_manifest_sha256"]==sha(source_views)
+source_data=json.loads(source_views.read_text())["rgb_figure"]["inputs"]
+expected={(x["method"],x["frame"]):x["sha256"] for x in source_data}
+assert len(fixed["images"])==25 and len(expected)==25
+assert {(x["method"],x["frame"]) for x in fixed["images"]}=={(m,f) for m in (*METHODS,"GT") for f in FRAMES}
+for item in fixed["images"]:
+    assert item["sha256"]==expected[(item["method"],item["frame"])]==sha(ROOT/item["published"])
+    assert not Path(item["source"]).is_absolute(), item["source"]
+
+assert all(not Path(item["source"]).is_absolute() for item in contract["vendor"].values())
+
+for rel in ("three.module.js","three.core.js","loaders/GLTFLoader.js","controls/OrbitControls.js",
+            "environments/RoomEnvironment.js","utils/BufferGeometryUtils.js"):
+    assert (ROOT/"vendor/three"/rel).is_file(),rel
+js=(ROOT/"scene-compare.js").read_text()
+assert "setScissor(" in js and "singleCamera:true" in js and "singleViewport:true" in js
+assert "runtime_matrix_gltf_y_up" in js and "applyMatrix4" in js
+assert "hiddenCutaway" in js and "ceiling|roof|wall|facade" in js
+app=(ROOT/"app.js").read_text()
+assert '2:new Set(["alignment"])' in app
+assert '3:new Set(["model_sha256","alignment"])' in app
+assert 'document.createElement("strong")' in app
+assert "fixedPreloads" not in app
+assert 'method.disabled=false;frame.disabled=false;' in app
+style=(ROOT/"style.css").read_text()
+assert "#figure-5,#figure-6{width:100%;margin:30px 0}" in style
+for name in ("index.html","en.html"):
+    html=(ROOT/name).read_text()
+    assert '<figure class="hero" id="figure-1"><img src="assets/teaser_originals.png"' in html
+
+manifest=json.loads((ROOT/"evidence/publication_manifest.json").read_text())
+assert not any(row["path"].startswith("office-") and row["path"].endswith(".mp4") for row in manifest["files"])
+for row in manifest["files"]: assert sha(ROOT/row["path"])==row["sha256"],row["path"]
+for line in (ROOT/"evidence/SHA256SUMS").read_text().splitlines():
+    expected_sha,rel=line.split("  ",1); assert sha(ROOT/rel)==expected_sha,rel
+print(json.dumps({"status":"PASS","tables":7,"models":4,"fixed_views":25,"figures":6,
+                  "checks":["links/anchors","GT SHA256","fixed-view SHA256","registration provenance","original GLB identity","no model-viewer"]},ensure_ascii=False))

scripts/validate.py

@@ -0,0 +1,6 @@
+#!/usr/bin/env python3
+"""Public entry point for publication validation."""
+import runpy
+from pathlib import Path
+
+runpy.run_path(str(Path(__file__).with_name("validate_publication.py")), run_name="__main__")

scripts/validate_publication.py

@@ -0,0 +1,423 @@
+#!/usr/bin/env python3
+"""Validate a publication checkout without access to its original sources."""
+
+from __future__ import annotations
+
+import hashlib
+import json
+import re
+import unittest
+from copy import deepcopy
+from collections import Counter
+from html.parser import HTMLParser
+from html import unescape
+from pathlib import Path
+from urllib.parse import unquote, urlsplit
+from unittest.mock import patch
+
+from editorial import narrative_copy, reconstruction_reductions
+from site_builder_interactive import academic_page
+
+
+ROOT = Path(__file__).resolve().parents[1]
+PAGES = ("index.html", "en.html", "zh.html")
+TITLE = "AWSM: Agentic World Simulation and Mapping — Phygital AI"
+CANONICAL_ROOT = "https://phygital-ai.github.io/agentic-world-simulation-and-mapping/"
+VIDEO = "assets/embodied-demo/four-robots-214924-156s-phone.mp4"
+POSTER = "assets/embodied-demo/navigation-214924-poster.png"
+HEX_SHA256 = re.compile(r"^[0-9a-f]{64}$")
+
+
+def load_json(relative_path: str) -> dict:
+    return json.loads((ROOT / relative_path).read_text(encoding="utf-8"))
+
+
+def sha256(path: Path) -> str:
+    digest = hashlib.sha256()
+    with path.open("rb") as stream:
+        for chunk in iter(lambda: stream.read(1024 * 1024), b""):
+            digest.update(chunk)
+    return digest.hexdigest()
+
+
+class Document(HTMLParser):
+    def __init__(self, path: Path) -> None:
+        super().__init__(convert_charrefs=True)
+        self.path = path
+        self.ids: list[str] = []
+        self.references: list[tuple[str, str, str]] = []
+        self.canonicals: list[str] = []
+        self.titles: list[str] = []
+        self.videos: list[dict] = []
+        self._video: dict | None = None
+        self.section_text: dict[str, list[str]] = {}
+        self._title_depth = 0
+        self._sections: list[str] = []
+        self.feed(path.read_text(encoding="utf-8"))
+
+    def handle_starttag(self, tag: str, attrs: list[tuple[str, str | None]]) -> None:
+        attributes = dict(attrs)
+        element_id = attributes.get("id")
+        if element_id is not None:
+            self.ids.append(element_id)
+        if tag == "section":
+            self._sections.append(element_id or "")
+            if element_id:
+                self.section_text.setdefault(element_id, [])
+        if tag == "title":
+            self._title_depth += 1
+        for attribute in ("href", "src"):
+            value = attributes.get(attribute)
+            if value:
+                self.references.append((tag, attribute, value))
+        rel = (attributes.get("rel") or "").split()
+        if tag == "link" and "canonical" in rel:
+            self.canonicals.append(attributes.get("href") or "")
+        if tag == "video":
+            self._video = {
+                "attributes": attributes,
+                "section": self._sections[-1] if self._sections else "",
+                "sources": [],
+            }
+            self.videos.append(self._video)
+        elif tag == "source" and self._video is not None and attributes.get("src"):
+            self._video["sources"].append(attributes["src"])
+
+    def handle_startendtag(self, tag: str, attrs: list[tuple[str, str | None]]) -> None:
+        self.handle_starttag(tag, attrs)
+
+    def handle_endtag(self, tag: str) -> None:
+        if tag == "title":
+            self._title_depth -= 1
+        if tag == "video":
+            self._video = None
+        if tag == "section" and self._sections:
+            self._sections.pop()
+
+    def handle_data(self, data: str) -> None:
+        if self._title_depth:
+            self.titles.append(data)
+        for section_id in self._sections:
+            if section_id:
+                self.section_text[section_id].append(data)
+
+    @property
+    def title(self) -> str:
+        return " ".join("".join(self.titles).split())
+
+    def text_for(self, section_id: str) -> str:
+        return " ".join("".join(self.section_text.get(section_id, [])).split())
+
+
+class PublicationValidation(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls) -> None:
+        cls.contract = load_json("data/source_contract.json")
+        cls.manifest = load_json("evidence/publication_manifest.json")
+        cls.demo = load_json("data/embodied_demo.json")
+        cls.documents = {name: Document(ROOT / name) for name in PAGES}
+
+    def assert_hashed_file(self, relative_path: str, expected_sha: str, expected_bytes=None) -> None:
+        self.assertIsInstance(relative_path, str)
+        self.assertTrue(relative_path, "empty published path")
+        self.assertRegex(expected_sha, HEX_SHA256)
+        path = ROOT / relative_path
+        self.assertTrue(path.is_file(), f"missing file: {relative_path}")
+        self.assertEqual(sha256(path), expected_sha, relative_path)
+        if expected_bytes is not None:
+            self.assertEqual(path.stat().st_size, expected_bytes, relative_path)
+
+    def test_source_contract_published_hashes(self) -> None:
+        records = list(self.contract["figures"].values())
+        for model in self.contract["models"].values():
+            records.extend(model.values())
+        records.extend(self.contract["vendor"].values())
+        records.append(self.contract["gt_asset"])
+        for record in records:
+            self.assert_hashed_file(record["published"], record["sha256"], record.get("bytes"))
+
+    def test_tables_and_figures_exist_in_both_languages(self) -> None:
+        required = {f"table-{number}" for number in (1, 2, 3, 6, 7)}
+        required |= {f"figure-{number}" for number in range(1, 6)}
+        self.assertEqual([row["number"] for row in self.contract["tables"]], list(range(1, 8)))
+        for name, document in self.documents.items():
+            self.assertTrue(required <= set(document.ids), f"{name}: missing {sorted(required - set(document.ids))}")
+            self.assertFalse({"table-4", "table-5"} & set(document.ids), f"{name}: removed appearance tables returned")
+
+    def test_editorial_preserves_experiment_copy(self) -> None:
+        protected = (
+            "intro2", "method1", "method2", "workflow", "finding1_title", "finding1", "finding2_title", "finding2",
+            "finding3_title", "discussion1_title", "discussion1",
+            "discussion2_title", "discussion2", "discussion3_title", "discussion3",
+            "metric1", "metric1_label", "metric2", "metric2_label", "metric3", "metric3_label",
+        )
+
+        def check_copy(copy, english):
+            original = deepcopy(copy)
+            narrative_copy(copy, english)
+            for key in protected:
+                self.assertEqual(copy[key], original[key], f"{english}: {key}")
+            self.assertEqual(copy["limits"][:len(original["limits"])], original["limits"])
+            self.assertEqual(len(copy["contributions"]), 3)
+
+        with patch("site_builder_interactive.narrative_copy", side_effect=check_copy):
+            academic_page(True)
+            academic_page(False)
+
+    def test_narrative_roles_and_scope(self) -> None:
+        sections = ["results", "motivation", "workflow", "more-results", "embodied-demo", "related-work", "limitations", "citation", "references"]
+        for name, document in self.documents.items():
+            self.assertEqual([element for element in document.ids if element in sections], sections)
+            markup = (ROOT / name).read_text(encoding="utf-8")
+            toc = re.search(r'<aside class="toc">(.*?)</aside>', markup, re.S).group(1)
+            self.assertEqual(re.findall(r'href="#([^"]+)"', toc), sections)
+            chinese = name == "zh.html"
+            results = document.text_for("results")
+            self.assertTrue(results.startswith("01 / OVERVIEW"))
+            self.assertNotIn("01 / RESULTS", markup)
+            self.assertIn('<a href="#results">' + ("概览" if chinese else "Overview") + '</a>', toc)
+            self.assertIn('<a href="#more-results">' + ("结果与分析" if chinese else "Results & analysis") + '</a>', toc)
+            self.assertTrue(document.text_for("more-results").startswith("04 / RESULTS AND ANALYSIS"))
+            self.assertIn("大模型智能体" if chinese else "large-model agents", results)
+            self.assertIn("agentic", results.lower())
+            self.assertIn("NVIDIA Isaac Sim", results)
+            self.assertNotIn("No universal winner", results)
+            self.assertNotIn("PSNR, SSIM, and LPIPS disagree", results)
+            self.assertLess(markup.index('id="figure-2"'), markup.index('id="figure-3"'))
+            self.assertLess(markup.index('id="figure-3"'), markup.index('id="figure-4"'))
+            for suffix in ("?lang=en", "shake?lang=en"):
+                self.assertIn(f'href="https://office-cafe-vipe.hiwtishere.chatgpt.site/{suffix}"', markup)
+            self.assertIn("可复用的仿真就绪" if chinese else "reusable simulation-ready", results)
+            self.assertIn("控制器承担另一种角色" if chinese else "controllers have a separate role", document.text_for("workflow"))
+            self.assertIn("而非学习得到的动力学预测模型" if chinese else "not a learned dynamics predictor", document.text_for("motivation"))
+            self.assertIn("特定场景的仿真集成" if chinese else "scene-specific simulation integration", document.text_for("related-work"))
+
+    def test_ids_links_and_anchors(self) -> None:
+        for name, document in self.documents.items():
+            duplicates = sorted(item for item, count in Counter(document.ids).items() if count > 1)
+            self.assertFalse(duplicates, f"{name}: duplicate ids {duplicates}")
+            for _tag, _attribute, reference in document.references:
+                parsed = urlsplit(reference)
+                if parsed.scheme or parsed.netloc or reference.startswith("//"):
+                    continue
+                if any(marker in reference for marker in ("{{", "}}", "${")):
+                    continue
+                if parsed.path:
+                    relative = Path(unquote(parsed.path.lstrip("/")))
+                    target = (ROOT / relative) if parsed.path.startswith("/") else (document.path.parent / relative)
+                    target = target.resolve()
+                    self.assertTrue(target.is_relative_to(ROOT), f"{name}: path escapes publication: {reference}")
+                    self.assertTrue(target.exists(), f"{name}: broken local reference: {reference}")
+                else:
+                    target = document.path
+                if parsed.fragment:
+                    self.assertEqual(target.suffix.lower(), ".html", f"{name}: fragment targets non-HTML: {reference}")
+                    target_document = self.documents.get(target.name) or Document(target)
+                    self.assertIn(unquote(parsed.fragment), target_document.ids, f"{name}: broken anchor: {reference}")
+
+    def test_manifest_files_match_bytes_and_hashes(self) -> None:
+        seen: set[str] = set()
+        for record in self.manifest["files"]:
+            relative_path = record["path"]
+            self.assertNotIn(relative_path, seen, f"duplicate manifest path: {relative_path}")
+            seen.add(relative_path)
+            self.assert_hashed_file(relative_path, record["sha256"], record["bytes"])
+
+    def test_sha256sums_matches_manifest_and_files(self) -> None:
+        sums: dict[str, str] = {}
+        for line_number, line in enumerate((ROOT / "evidence/SHA256SUMS").read_text(encoding="utf-8").splitlines(), 1):
+            match = re.fullmatch(r"([0-9a-f]{64})  (.+)", line)
+            self.assertIsNotNone(match, f"SHA256SUMS:{line_number}: invalid line")
+            expected_sha, relative_path = match.groups()
+            self.assertNotIn(relative_path, sums, f"SHA256SUMS duplicate: {relative_path}")
+            sums[relative_path] = expected_sha
+            self.assert_hashed_file(relative_path, expected_sha)
+        manifest_sums = {record["path"]: record["sha256"] for record in self.manifest["files"]}
+        manifest_path = "evidence/publication_manifest.json"
+        expected_sums = dict(manifest_sums)
+        expected_sums[manifest_path] = sha256(ROOT / manifest_path)
+        self.assertEqual(sums, expected_sums)
+
+    def test_embodied_demo_metadata_and_media(self) -> None:
+        media = self.demo["media"]
+        self.assertEqual(media["video"], VIDEO)
+        self.assertEqual(media["poster"], POSTER)
+        self.assertIsInstance(media["duration_s"], (int, float))
+        self.assertGreater(media["duration_s"], 0)
+        self.assertIsInstance(self.demo["run_id"], str)
+        self.assertTrue(self.demo["run_id"].strip())
+        self.assertEqual(self.demo["audit"]["status"], "FAILED")
+        self.assertEqual(self.demo["navigation"]["localization"], "simulator_pose")
+        self.assertEqual(self.demo["navigation"]["route"], "predefined")
+        self.assert_hashed_file(media["video"], media["sha256"], media["bytes"])
+        self.assertEqual(media["duration_s"], 156.0)
+        self.assertEqual(media["frame_count"], 4680)
+        self.assertEqual(self.demo["run_id"], "sceneweft-walljourney-20260930-214924")
+        self.assert_hashed_file(media["poster"], media["poster_sha256"])
+        self.assert_hashed_file(media["hd"]["video"], media["hd"]["sha256"], media["hd"]["bytes"])
+        self.assert_hashed_file(media["map"]["image"], media["map"]["sha256"], media["map"]["bytes"])
+        self.assertEqual(self.demo["audit"]["measured_terminal_hold_s"], 7.994999821297824)
+        audit = load_json("evidence/embodied-demo-audit.json")
+        self.assertEqual(audit["run_id"], self.demo["run_id"])
+        self.assertEqual([check["check"] for check in audit["failed_checks"]], ["measured_terminal_health_and_formation_hold"])
+        delivery = load_json("evidence/embodied-demo-delivery.json")
+        self.assertEqual(delivery["run_id"], self.demo["run_id"])
+        for binding in delivery["asset_bindings"]:
+            self.assert_hashed_file(binding["published"], binding["sha256"])
+
+    def test_embodied_demo_markup_and_bilingual_copy(self) -> None:
+        navigation_terms = {
+            "index.html": ("navigation", "simulation"),
+            "en.html": ("navigation", "simulation"),
+            "zh.html": ("导航", "仿真"),
+        }
+        for name, document in self.documents.items():
+            self.assertIn("embodied-demo", document.ids, f"{name}: missing embodied-demo section")
+            matching_videos = [
+                video for video in document.videos
+                if video["section"] == "embodied-demo" and video["sources"] == [VIDEO]
+            ]
+            self.assertEqual(len(matching_videos), 1, f"{name}: expected one embodied demo video")
+            self.assertEqual(
+                sum(VIDEO in video["sources"] for video in document.videos),
+                1,
+                f"{name}: demo source must belong to exactly one video",
+            )
+            video = matching_videos[0]["attributes"]
+            self.assertNotIn("src", video)
+            self.assertEqual(video.get("poster"), POSTER)
+            self.assertIn(video.get("preload", "metadata").lower(), {"none", "metadata"})
+            self.assertIn("controls", video)
+            self.assertIn("playsinline", video)
+            self.assertNotIn("autoplay", video)
+            section_text = document.text_for("embodied-demo").lower()
+            for term in navigation_terms[name]:
+                self.assertIn(term, section_text, f"{name}: embodied demo must state {term!r}")
+            self.assertIn("20260930-214924", section_text)
+            self.assertIn("2:36", section_text)
+            self.assertNotIn("235.2", section_text)
+            self.assertIn("demo-route-map", document.ids)
+
+    def test_title_and_language_specific_canonical(self) -> None:
+        expected = {
+            "index.html": CANONICAL_ROOT,
+            "en.html": CANONICAL_ROOT,
+            "zh.html": CANONICAL_ROOT + "zh.html",
+        }
+        for name, document in self.documents.items():
+            self.assertEqual(document.title, "AWSM：智能体世界仿真与建图 — Phygital AI" if name == "zh.html" else TITLE)
+            self.assertEqual(document.canonicals, [expected[name]])
+            markup = (ROOT / name).read_text(encoding="utf-8")
+            self.assertIn('<html lang="zh-CN">' if name == "zh.html" else '<html lang="en">', markup)
+            self.assertIn(f'hreflang="zh-CN" href="{CANONICAL_ROOT}zh.html"', markup)
+            self.assertIn(f'hreflang="en" href="{CANONICAL_ROOT}"', markup)
+            self.assertIn(f'hreflang="x-default" href="{CANONICAL_ROOT}"', markup)
+            self.assertIn('<a href="zh.html"', markup)
+            self.assertIn('<a href="index.html"', markup)
+        self.assertEqual((ROOT / "index.html").read_bytes(), (ROOT / "en.html").read_bytes())
+
+    def test_awsm_header(self) -> None:
+        for name in PAGES:
+            markup = (ROOT / name).read_text(encoding="utf-8")
+            header = markup.split('<header>', 1)[1].split('</header>', 1)[0]
+            chinese = name == "zh.html"
+            title_name = '智能体世界仿真与建图' if chinese else 'Agentic World Simulation and Mapping'
+            self.assertIn(f'<p class="title-name">{title_name}</p>', header)
+            self.assertIn(f'<h1 aria-label="AWSM: {title_name}">AWSM</h1>', header)
+            self.assertLess(header.index('class="title-name"'), header.index('class="title-lockup"'))
+            subtitle = '把真实空间，变成虚实融合智能体可以使用的世界。' if chinese else 'From real spaces to worlds phygital agents can use.'
+            lead = re.search(r'<p class="lead">(.*?)</p>', header).group(1)
+            self.assertEqual(re.sub(r'<[^>]+>', '', lead), subtitle)
+            emphasis = '虚实融合智能体可以使用的世界。' if chinese else 'worlds phygital agents can use.'
+            self.assertIn(f'<strong class="tagline-emphasis">{emphasis}</strong>', lead)
+            self.assertIn(f'<meta name="description" content="{subtitle}">', markup)
+            self.assertNotIn('class="brand-tagline"', header)
+            pronunciation = '读作 “awesome”' if chinese else 'pronounced “awesome”'
+            self.assertIn(f'<p class="pronunciation">{pronunciation}</p>', header)
+            brand_note = 'Phygital = physical（物理）+ digital（数字），即虚实融合。' if chinese else 'Phygital = physical + digital.'
+            self.assertIn(f'<p class="brand-note">{brand_note}</p>', header)
+            self.assertNotIn(')Phygital', header)
+            self.assertNotIn('class="eyebrow"', header)
+            self.assertNotIn('class="meta"', header)
+            self.assertIn('@misc{awsm_2026,', markup)
+            self.assertIn('title        = {{AWSM}: Agentic World Simulation and Mapping}', markup)
+            self.assertNotIn('@misc{agentic_world_2026,', markup)
+            self.assertIn('<time datetime="2026-10-01">', header)
+            self.assertIn('2026年10月1日' if chinese else 'October 1, 2026', header)
+
+    def test_viewer_sources_and_concurrent_work(self) -> None:
+        from site_builder_interactive import SCENE_JS
+        self.assertEqual((ROOT / "scene-compare.js").read_text().strip(), SCENE_JS.strip())
+        manifest = json.loads((ROOT / "data/scene_comparison.json").read_text())
+        self.assertGreater(manifest["camera"]["position"][1], 20)
+        self.assertNotIn("mirror", manifest["display_cutaway"]["keep_name_pattern"])
+        for name in PAGES:
+            markup = (ROOT / name).read_text(encoding="utf-8")
+            for asset in ("app.js", "scene-compare.js", "office-models.js"):
+                self.assertIn(f'src="{asset}?v={sha256(ROOT / asset)[:12]}"', markup)
+            self.assertIn('同期工作 AHa-3D' if name == 'zh.html' else 'Concurrent work AHa-3D', markup)
+            self.assertIn('剖切视图' if name == 'zh.html' else 'Toggle Cutaway', markup)
+
+    def test_editorial_stylesheet_cache_version(self) -> None:
+        expected = f'editorial.css?v={sha256(ROOT / "editorial.css")[:12]}'
+        for name in PAGES:
+            markup = (ROOT / name).read_text(encoding="utf-8")
+            self.assertIn(f'<link rel="stylesheet" href="{expected}">', markup)
+            self.assertNotIn('href="editorial.css"', markup)
+
+    def test_downloadable_citation(self) -> None:
+        citation = (ROOT / "data/awsm.bib").read_text(encoding="utf-8")
+        self.assertIn('author       = {{Phygital AI}}', citation)
+        self.assertIn(f'url          = {{{CANONICAL_ROOT}}}', citation)
+        for name in PAGES:
+            markup = (ROOT / name).read_text(encoding="utf-8")
+            rendered = re.search(r'<pre class="citation-block"><code>(.*?)</code></pre>', markup, re.S).group(1)
+            self.assertEqual(unescape(rendered), citation.rstrip())
+            self.assertIn('href="data/awsm.bib" download="awsm.bib"', markup)
+
+    def test_revision_presentation_and_review_scope(self) -> None:
+        for name in PAGES:
+            markup = (ROOT / name).read_text()
+            nav = re.search(r'<nav>(.*?)</nav>', markup, re.S).group(1)
+            self.assertIn('href="https://github.com/wentingw/AWSM"', nav)
+            self.assertIn('GitHub ↗', nav)
+            self.assertNotIn('DATA ↗', nav)
+            self.assertEqual(re.findall(r'<figure[^>]*id="figure-(\d+)"', markup), list('12345'))
+            for number in range(1, 6):
+                self.assertEqual(markup.count(f'Figure {number}.'), 1)
+            self.assertIn('<figure id="figure-2"><img loading="lazy" src="assets/fixed_five_view_comparison_m1_m4.jpg"', markup)
+            self.assertIn('class="scene-comparison" id="figure-3"', markup)
+            self.assertIn('class="fixed-comparison" id="figure-4"', markup)
+            self.assertIn('class="hero" id="overview-video"', markup)
+            self.assertNotIn('figure-1-caption', markup)
+            self.assertNotIn('AWSM: from real spaces to worlds phygital agents can use.', markup)
+            for phrase in ('7.994999821', 'unresolved audit failures', 'did not pass every task-audit', 'We discuss it as a parallel effort', '前置工作', '未通过的审计项', '尚未通过全部任务审计', '源运行报告飞行任务'):
+                self.assertNotIn(phrase, markup)
+            self.assertIn('让机器人前往前台，并在那里排成一列。' if name=='zh.html' else 'Send the robots to the reception desk and have them line up there.', markup)
+            self.assertIn('已完成人工审阅' if name=='zh.html' else 'manually reviewed', markup)
+            self.assertIn('不改变其原始判定' if name=='zh.html' else 'does not revise their recorded verdicts', markup)
+            self.assertEqual(markup.count('data-office-external='), 2)
+            self.assertIn('并非外站页面截图' if name=='zh.html' else 'not screenshots of the external website', markup)
+        self.assertEqual(self.demo['audit']['status'], 'FAILED')
+
+    def test_tldr_reductions_and_task_scope(self) -> None:
+        surface, depth = reconstruction_reductions()
+        self.assertAlmostEqual(surface, 80.95096, places=3)
+        self.assertAlmostEqual(depth, 53.77076, places=3)
+        for name, document in self.documents.items():
+            markup = (ROOT / name).read_text()
+            chinese = name == 'zh.html'
+            summary = re.search(r'<aside class="tldr"><strong>TL;DR</strong><p>(.*?)</p>', markup).group(1)
+            for term in ('IMU', 'M1', 'M4', '81%', '54%'):
+                self.assertIn(term, summary)
+            self.assertIn('真值位姿条件' if chinese else 'GT-pose-conditioned', summary)
+            self.assertIn('预设路线' if chinese else 'predefined routes', summary)
+            self.assertIn('180 个评估视角' if chinese else '180 evaluation views', document.text_for('results'))
+            self.assertEqual(markup.count('class="metric-reduction"'), 2)
+            self.assertNotIn('180 度', document.text_for('results'))
+            self.assertIn('相对降低约 81%' if chinese else '≈81% relative reduction', markup)
+            self.assertIn('相对降低约 54%' if chinese else '≈54% relative reduction', markup)
+
+
+if __name__ == "__main__":
+    unittest.main(verbosity=2)

style.css

@@ -0,0 +1 @@
+:root{--paper:#f7f6f1;--white:#fffefa;--ink:#182c29;--muted:#60706a;--green:#196451;--line:#d9dfd7;--gold:#b87e40}*{box-sizing:border-box}html{scroll-behavior:smooth}body{margin:0;background:var(--paper);color:var(--ink);font:17px/1.8 Inter,"Noto Sans CJK SC",system-ui,sans-serif}a{color:var(--green);text-underline-offset:4px}img{display:block;max-width:100%;height:auto}.skip{position:absolute;left:-9999px}.skip:focus{left:12px;top:12px;background:white;padding:8px;z-index:9}nav{max-width:1420px;margin:auto;padding:22px 4vw;display:flex;justify-content:space-between;border-bottom:1px solid var(--line);font-size:12px;letter-spacing:.09em}nav div{display:flex;gap:20px}.brand{font-weight:750;text-decoration:none}header{max-width:1120px;margin:56px auto 36px;padding:0 32px}.eyebrow,.section-tag{font-size:11px;letter-spacing:.16em;color:var(--green);font-weight:700}h1,h2{font-family:Georgia,"Noto Serif CJK SC",serif;font-weight:500}h1{font-size:clamp(46px,7vw,88px);line-height:1.12;letter-spacing:-.035em;margin:18px 0}.lead{max-width:800px;color:var(--muted);font-size:20px}.meta{display:flex;gap:24px;flex-wrap:wrap;font-size:12px;color:var(--muted)}.hero{max-width:1320px;margin:35px auto 68px;padding:0 28px}.hero img,figure>img{width:100%;background:#e8ebe5}.hero-pair-grid{display:grid;grid-template-columns:1fr 1fr;gap:3px;background:var(--ink)}.hero-pair-grid img{width:100%;aspect-ratio:4/3;object-fit:cover}.hero-pair-grid p{margin:0;padding:7px 12px;background:var(--ink);color:#f7f6f1;font-size:10px;letter-spacing:.11em}figcaption{font-size:12px;color:var(--muted);margin-top:10px}figcaption span{font-weight:700;color:var(--green)}.layout{max-width:1300px;margin:auto;padding:0 30px;display:grid;grid-template-columns:190px minmax(0,920px);gap:60px;justify-content:center}.toc{position:sticky;top:24px;align-self:start;display:flex;flex-direction:column;gap:9px;font-size:12px}.toc a{text-decoration:none;color:var(--muted)}main{min-width:0}section{padding:28px 0 38px;border-top:1px solid var(--line)}section:first-child{border-top:0}.standfirst{font-size:23px;line-height:1.55;color:var(--green)}h2{font-size:32px;margin:7px 0 22px}.notice{border-left:3px solid var(--gold);padding:10px 18px;margin:24px 0;color:var(--muted);background:#fff9}.table-figure{margin:30px 0}.table-scroll{overflow:auto;border:1px solid var(--line);background:var(--white);margin-top:9px}table{width:100%;border-collapse:collapse;white-space:nowrap;font-size:12px;font-variant-numeric:tabular-nums}th,td{padding:11px 12px;border-bottom:1px solid var(--line);text-align:left;vertical-align:top}th{background:#edf0e9;color:var(--muted)}td.wrap{white-space:normal;min-width:240px}.scene-comparison,.fixed-comparison{margin:34px 0;padding:20px;border:1px solid var(--line);border-radius:6px;background:var(--white)}.scene-comparison h3{margin:0 0 18px}.scene-toolbar,.controlrow{display:flex;gap:14px 20px;align-items:end;flex-wrap:wrap}.scene-toolbar label,.controlrow label{display:flex;flex-direction:column;font-size:12px;color:var(--muted)}.method-control{flex:1 1 280px}select,button{font:inherit}.scene-toolbar button,.scene-retry{font-size:13px;padding:9px 13px;color:var(--green);border:1px solid var(--line);border-radius:3px;background:var(--paper);cursor:pointer}.scene-stage{position:relative;height:540px;width:100%;background:#e9ede5;margin-top:18px;overflow:hidden;border-radius:4px;--split:50%}.scene-stage canvas{display:block;width:100%;height:100%;touch-action:none}.scene-label{position:absolute;top:12px;z-index:3;padding:4px 8px;background:#fffde;color:var(--ink);font-size:11px;pointer-events:none}.scene-label-left{left:12px}.scene-label-right{right:12px}.scene-divider{position:absolute;z-index:4;left:var(--split);top:0;width:3px;height:100%;padding:0;border:0;background:white;box-shadow:0 0 0 1px #182c2955;transform:translateX(-50%);cursor:ew-resize;touch-action:none}.scene-divider:after{content:"↔";position:absolute;top:50%;left:50%;transform:translate(-50%,-50%);width:30px;height:30px;line-height:27px;border-radius:50%;background:white;color:var(--green);border:1px solid var(--line)}.scene-split{position:absolute;width:1px;height:1px;opacity:0}.scene-status{position:absolute;left:50%;top:50%;transform:translate(-50%,-50%);padding:9px 12px;background:#fffef0;z-index:5;font-size:13px}.scene-status.error{color:#8d251d}.scene-retry{display:none;margin-top:10px}.scene-comparison[data-state=error] .scene-retry{display:inline-block}.scene-comparison:not(.is-comparing) .scene-divider{display:none}.compare-pair{display:grid;grid-template-columns:1fr 1fr;gap:14px;margin-top:16px}.compare-pair img{width:100%;aspect-ratio:4/3;object-fit:contain;background:#e9ede5}.compare-subcaption{margin:6px 0 0;font-size:11px;color:var(--muted);letter-spacing:.04em}.asset-downloads{display:grid;grid-template-columns:1fr 1fr;gap:0 20px;font-size:12px}.asset-downloads p{margin:4px 0}.office-cafe-embed{margin-top:24px;border:1px solid var(--line);border-radius:6px;overflow:hidden;background:#111}.office-cafe-embed iframe{display:block;width:100%;height:min(880px,85vh);min-height:680px;border:0;background:#111}.office-links{display:flex;gap:18px;flex-wrap:wrap;font-size:12px}footer{max-width:1120px;margin:70px auto 0;padding:30px;border-top:1px solid var(--line);font-size:12px;color:var(--muted)}@media(max-width:950px){.layout{grid-template-columns:1fr}.toc{position:static;display:grid;grid-template-columns:1fr 1fr}}@media(max-width:600px){body{font-size:16px}nav{padding:17px 18px}.brand{font-size:10px}nav div{gap:12px}header{padding:0 20px;margin-top:35px}h1{font-size:43px}.lead{font-size:17px}.hero{padding:0 12px;margin-bottom:40px}.layout{padding:0 20px}.toc{grid-template-columns:1fr}.standfirst{font-size:20px}h2{font-size:27px}.scene-comparison,.fixed-comparison{padding:12px}.scene-stage{height:360px}.compare-pair,.asset-downloads{grid-template-columns:1fr}.office-cafe-embed iframe{height:760px;min-height:620px}}#figure-2,#figure-5{width:100%;margin:30px 0}#figure-2>img,#figure-5>img{width:100%;background:var(--paper)}main h3{font-family:Georgia,"Noto Serif CJK SC",serif;font-size:23px;font-weight:500;line-height:1.4;margin:32px 0 10px}main h4{font:500 18px/1.45 Georgia,"Noto Serif CJK SC",serif;margin:28px 0 8px}.tldr{margin:0 0 34px;padding:20px 24px;border-left:4px solid var(--green);background:#edf3ee}.tldr strong{font-size:11px;letter-spacing:.16em;color:var(--green)}.tldr p{margin:5px 0 0;font:500 19px/1.65 Georgia,"Noto Serif CJK SC",serif}.contributions{padding-left:24px}.contributions li{margin:10px 0;padding-left:5px}.inline-sources{display:block;margin-top:12px;font-size:13px}.workflow-strip{display:grid;grid-template-columns:repeat(5,1fr);gap:1px;margin:22px 0 34px;border:1px solid var(--line);background:var(--line)}.workflow-strip div{display:flex;min-height:180px;padding:18px 15px;background:var(--white);flex-direction:column}.workflow-strip strong{margin-bottom:14px;font-size:12px;letter-spacing:.08em;color:var(--green)}.workflow-strip span{font-size:13px;line-height:1.6;color:var(--muted)}.insight-strip{display:grid;grid-template-columns:repeat(3,1fr);gap:1px;margin:34px 0 12px;background:var(--line);border:1px solid var(--line)}.insight-strip div{display:flex;min-height:125px;padding:22px 18px;background:var(--white);flex-direction:column;justify-content:space-between}.insight-strip strong{font:500 25px/1.15 Georgia,serif;color:var(--green);letter-spacing:-.02em}.insight-strip span{font-size:11px;line-height:1.45;color:var(--muted);text-transform:uppercase;letter-spacing:.07em}.metric-caveat{margin:8px 0 0;color:var(--muted);font-size:12px}.claim{margin:30px 0;padding:20px 24px;border:1px solid var(--line);background:var(--white)}.claim strong,.finding-kicker{font-size:11px;letter-spacing:.13em;text-transform:uppercase;color:var(--green)}.claim p{margin:8px 0 0;font:500 21px/1.55 Georgia,"Noto Serif CJK SC",serif}.finding{margin:42px 0 18px;padding:0 0 22px;border-bottom:1px solid var(--line)}.finding h3{margin:5px 0 12px}.finding p:last-child{margin-bottom:0}.analysis-block{margin-top:58px;padding-top:12px;border-top:1px solid var(--line)}.external-case{margin-bottom:10px}.limits-list{padding-left:22px}.limits-list li{margin:10px 0}.evidence-links{font-size:12px;margin-top:25px}.citation-block{overflow:auto;margin:22px 0;padding:22px;border:1px solid var(--line);background:var(--white);font:13px/1.65 ui-monospace,SFMono-Regular,Consolas,monospace}.references-list{padding-left:24px}.references-list li{margin:13px 0;padding-left:6px}.references-list a{color:var(--ink)}@media(max-width:900px){.workflow-strip{grid-template-columns:1fr 1fr}.workflow-strip div{min-height:130px}}@media(max-width:700px){.insight-strip,.workflow-strip{grid-template-columns:1fr}.insight-strip div{min-height:98px}.workflow-strip div{min-height:auto}.claim p{font-size:18px}.tldr{padding:17px 19px}.tldr p{font-size:17px}.citation-block{padding:15px;font-size:11px}}

vendor/three/controls/OrbitControls.js

@@ -0,0 +1,1860 @@
+import {
+	Controls,
+	MOUSE,
+	Quaternion,
+	Spherical,
+	TOUCH,
+	Vector2,
+	Vector3,
+	Plane,
+	Ray,
+	MathUtils
+} from 'three';
+
+/**
+ * Fires when the camera has been transformed by the controls.
+ *
+ * @event OrbitControls#change
+ * @type {Object}
+ */
+const _changeEvent = { type: 'change' };
+
+/**
+ * Fires when an interaction was initiated.
+ *
+ * @event OrbitControls#start
+ * @type {Object}
+ */
+const _startEvent = { type: 'start' };
+
+/**
+ * Fires when an interaction has finished.
+ *
+ * @event OrbitControls#end
+ * @type {Object}
+ */
+const _endEvent = { type: 'end' };
+
+const _ray = new Ray();
+const _plane = new Plane();
+const _TILT_LIMIT = Math.cos( 70 * MathUtils.DEG2RAD );
+
+const _v = new Vector3();
+const _twoPI = 2 * Math.PI;
+
+const _STATE = {
+	NONE: - 1,
+	ROTATE: 0,
+	DOLLY: 1,
+	PAN: 2,
+	TOUCH_ROTATE: 3,
+	TOUCH_PAN: 4,
+	TOUCH_DOLLY_PAN: 5,
+	TOUCH_DOLLY_ROTATE: 6
+};
+const _EPS = 0.000001;
+
+
+/**
+ * Orbit controls allow the camera to orbit around a target.
+ *
+ * OrbitControls performs orbiting, dollying (zooming), and panning. Unlike {@link TrackballControls},
+ * it maintains the "up" direction `object.up` (+Y by default).
+ *
+ * - Orbit: Left mouse / touch: one-finger move.
+ * - Zoom: Middle mouse, or mousewheel / touch: two-finger spread or squish.
+ * - Pan: Right mouse, or left mouse + ctrl/meta/shiftKey, or arrow keys / touch: two-finger move.
+ *
+ * ```js
+ * const controls = new OrbitControls( camera, renderer.domElement );
+ *
+ * // controls.update() must be called after any manual changes to the camera's transform
+ * camera.position.set( 0, 20, 100 );
+ * controls.update();
+ *
+ * function animate() {
+ *
+ * 	// required if controls.enableDamping or controls.autoRotate are set to true
+ * 	controls.update();
+ *
+ * 	renderer.render( scene, camera );
+ *
+ * }
+ * ```
+ *
+ * @augments Controls
+ * @three_import import { OrbitControls } from 'three/addons/controls/OrbitControls.js';
+ */
+class OrbitControls extends Controls {
+
+	/**
+	 * Constructs a new controls instance.
+	 *
+	 * @param {Object3D} object - The object that is managed by the controls.
+	 * @param {?HTMLDOMElement} domElement - The HTML element used for event listeners.
+	 */
+	constructor( object, domElement = null ) {
+
+		super( object, domElement );
+
+		this.state = _STATE.NONE;
+
+		/**
+		 * The focus point of the controls, the `object` orbits around this.
+		 * It can be updated manually at any point to change the focus of the controls.
+		 *
+		 * @type {Vector3}
+		 */
+		this.target = new Vector3();
+
+		/**
+		 * The focus point of the `minTargetRadius` and `maxTargetRadius` limits.
+		 * It can be updated manually at any point to change the center of interest
+		 * for the `target`.
+		 *
+		 * @type {Vector3}
+		 */
+		this.cursor = new Vector3();
+
+		/**
+		 * How far you can dolly in (perspective camera only).
+		 *
+		 * @type {number}
+		 * @default 0
+		 */
+		this.minDistance = 0;
+
+		/**
+		 * How far you can dolly out (perspective camera only).
+		 *
+		 * @type {number}
+		 * @default Infinity
+		 */
+		this.maxDistance = Infinity;
+
+		/**
+		 * How far you can zoom in (orthographic camera only).
+		 *
+		 * @type {number}
+		 * @default 0
+		 */
+		this.minZoom = 0;
+
+		/**
+		 * How far you can zoom out (orthographic camera only).
+		 *
+		 * @type {number}
+		 * @default Infinity
+		 */
+		this.maxZoom = Infinity;
+
+		/**
+		 * How close you can get the target to the 3D `cursor`.
+		 *
+		 * @type {number}
+		 * @default 0
+		 */
+		this.minTargetRadius = 0;
+
+		/**
+		 * How far you can move the target from the 3D `cursor`.
+		 *
+		 * @type {number}
+		 * @default Infinity
+		 */
+		this.maxTargetRadius = Infinity;
+
+		/**
+		 * How far you can orbit vertically, lower limit. Range is `[0, Math.PI]` radians.
+		 *
+		 * @type {number}
+		 * @default 0
+		 */
+		this.minPolarAngle = 0;
+
+		/**
+		 * How far you can orbit vertically, upper limit. Range is `[0, Math.PI]` radians.
+		 *
+		 * @type {number}
+		 * @default Math.PI
+		 */
+		this.maxPolarAngle = Math.PI;
+
+		/**
+		 * How far you can orbit horizontally, lower limit. If set, the interval `[ min, max ]`
+		 * must be a sub-interval of `[ - 2 PI, 2 PI ]`, with `( max - min < 2 PI )`.
+		 *
+		 * @type {number}
+		 * @default -Infinity
+		 */
+		this.minAzimuthAngle = - Infinity;
+
+		/**
+		 * How far you can orbit horizontally, upper limit. If set, the interval `[ min, max ]`
+		 * must be a sub-interval of `[ - 2 PI, 2 PI ]`, with `( max - min < 2 PI )`.
+		 *
+		 * @type {number}
+		 * @default -Infinity
+		 */
+		this.maxAzimuthAngle = Infinity;
+
+		/**
+		 * Set to `true` to enable damping (inertia), which can be used to give a sense of weight
+		 * to the controls. Note that if this is enabled, you must call `update()` in your animation
+		 * loop.
+		 *
+		 * @type {boolean}
+		 * @default false
+		 */
+		this.enableDamping = false;
+
+		/**
+		 * The damping inertia used if `enableDamping` is set to `true`.
+		 *
+		 * Note that for this to work, you must call `update()` in your animation loop.
+		 *
+		 * @type {number}
+		 * @default 0.05
+		 */
+		this.dampingFactor = 0.05;
+
+		/**
+		 * Enable or disable zooming (dollying) of the camera.
+		 *
+		 * @type {boolean}
+		 * @default true
+		 */
+		this.enableZoom = true;
+
+		/**
+		 * Speed of zooming / dollying.
+		 *
+		 * @type {number}
+		 * @default 1
+		 */
+		this.zoomSpeed = 1.0;
+
+		/**
+		 * Enable or disable horizontal and vertical rotation of the camera.
+		 *
+		 * Note that it is possible to disable a single axis by setting the min and max of the
+		 * `minPolarAngle` or `minAzimuthAngle` to the same value, which will cause the vertical
+		 * or horizontal rotation to be fixed at that value.
+		 *
+		 * @type {boolean}
+		 * @default true
+		 */
+		this.enableRotate = true;
+
+		/**
+		 * Speed of rotation.
+		 *
+		 * @type {number}
+		 * @default 1
+		 */
+		this.rotateSpeed = 1.0;
+
+		/**
+		 * How fast to rotate the camera when the keyboard is used.
+		 *
+		 * @type {number}
+		 * @default 1
+		 */
+		this.keyRotateSpeed = 1.0;
+
+		/**
+		 * Enable or disable camera panning.
+		 *
+		 * @type {boolean}
+		 * @default true
+		 */
+		this.enablePan = true;
+
+		/**
+		 * Speed of panning.
+		 *
+		 * @type {number}
+		 * @default 1
+		 */
+		this.panSpeed = 1.0;
+
+		/**
+		 * Defines how the camera's position is translated when panning. If `true`, the camera pans
+		 * in screen space. Otherwise, the camera pans in the plane orthogonal to the camera's up
+		 * direction.
+		 *
+		 * @type {boolean}
+		 * @default true
+		 */
+		this.screenSpacePanning = true;
+
+		/**
+		 * How fast to pan the camera when the keyboard is used in
+		 * pixels per keypress.
+		 *
+		 * @type {number}
+		 * @default 7
+		 */
+		this.keyPanSpeed = 7.0;
+
+		/**
+		 * Setting this property to `true` allows to zoom to the cursor's position.
+		 *
+		 * @type {boolean}
+		 * @default false
+		 */
+		this.zoomToCursor = false;
+
+		/**
+		 * Set to true to automatically rotate around the target
+		 *
+		 * Note that if this is enabled, you must call `update()` in your animation loop.
+		 * If you want the auto-rotate speed to be independent of the frame rate (the refresh
+		 * rate of the display), you must pass the time `deltaTime`, in seconds, to `update()`.
+		 *
+		 * @type {boolean}
+		 * @default false
+		 */
+		this.autoRotate = false;
+
+		/**
+		 * How fast to rotate around the target if `autoRotate` is `true`. The default  equates to 30 seconds
+		 * per orbit at 60fps.
+		 *
+		 * Note that if `autoRotate` is enabled, you must call `update()` in your animation loop.
+		 *
+		 * @type {number}
+		 * @default 2
+		 */
+		this.autoRotateSpeed = 2.0;
+
+		/**
+		 * This object contains references to the keycodes for controlling camera panning.
+		 *
+		 * ```js
+		 * controls.keys = {
+		 * 	LEFT: 'ArrowLeft', //left arrow
+		 * 	UP: 'ArrowUp', // up arrow
+		 * 	RIGHT: 'ArrowRight', // right arrow
+		 * 	BOTTOM: 'ArrowDown' // down arrow
+		 * }
+		 * ```
+		 * @type {Object}
+		 */
+		this.keys = { LEFT: 'ArrowLeft', UP: 'ArrowUp', RIGHT: 'ArrowRight', BOTTOM: 'ArrowDown' };
+
+		/**
+		 * This object contains references to the mouse actions used by the controls.
+		 *
+		 * ```js
+		 * controls.mouseButtons = {
+		 * 	LEFT: THREE.MOUSE.ROTATE,
+		 * 	MIDDLE: THREE.MOUSE.DOLLY,
+		 * 	RIGHT: THREE.MOUSE.PAN
+		 * }
+		 * ```
+		 * @type {Object}
+		 */
+		this.mouseButtons = { LEFT: MOUSE.ROTATE, MIDDLE: MOUSE.DOLLY, RIGHT: MOUSE.PAN };
+
+		/**
+		 * This object contains references to the touch actions used by the controls.
+		 *
+		 * ```js
+		 * controls.mouseButtons = {
+		 * 	ONE: THREE.TOUCH.ROTATE,
+		 * 	TWO: THREE.TOUCH.DOLLY_PAN
+		 * }
+		 * ```
+		 * @type {Object}
+		 */
+		this.touches = { ONE: TOUCH.ROTATE, TWO: TOUCH.DOLLY_PAN };
+
+		/**
+		 * Used internally by `saveState()` and `reset()`.
+		 *
+		 * @type {Vector3}
+		 */
+		this.target0 = this.target.clone();
+
+		/**
+		 * Used internally by `saveState()` and `reset()`.
+		 *
+		 * @type {Vector3}
+		 */
+		this.position0 = this.object.position.clone();
+
+		/**
+		 * Used internally by `saveState()` and `reset()`.
+		 *
+		 * @type {number}
+		 */
+		this.zoom0 = this.object.zoom;
+
+		// the target DOM element for key events
+		this._domElementKeyEvents = null;
+
+		// internals
+
+		this._lastPosition = new Vector3();
+		this._lastQuaternion = new Quaternion();
+		this._lastTargetPosition = new Vector3();
+
+		// so camera.up is the orbit axis
+		this._quat = new Quaternion().setFromUnitVectors( object.up, new Vector3( 0, 1, 0 ) );
+		this._quatInverse = this._quat.clone().invert();
+
+		// current position in spherical coordinates
+		this._spherical = new Spherical();
+		this._sphericalDelta = new Spherical();
+
+		this._scale = 1;
+		this._panOffset = new Vector3();
+
+		this._rotateStart = new Vector2();
+		this._rotateEnd = new Vector2();
+		this._rotateDelta = new Vector2();
+
+		this._panStart = new Vector2();
+		this._panEnd = new Vector2();
+		this._panDelta = new Vector2();
+
+		this._dollyStart = new Vector2();
+		this._dollyEnd = new Vector2();
+		this._dollyDelta = new Vector2();
+
+		this._dollyDirection = new Vector3();
+		this._mouse = new Vector2();
+		this._performCursorZoom = false;
+
+		this._pointers = [];
+		this._pointerPositions = {};
+
+		this._controlActive = false;
+
+		// event listeners
+
+		this._onPointerMove = onPointerMove.bind( this );
+		this._onPointerDown = onPointerDown.bind( this );
+		this._onPointerUp = onPointerUp.bind( this );
+		this._onContextMenu = onContextMenu.bind( this );
+		this._onMouseWheel = onMouseWheel.bind( this );
+		this._onKeyDown = onKeyDown.bind( this );
+
+		this._onTouchStart = onTouchStart.bind( this );
+		this._onTouchMove = onTouchMove.bind( this );
+
+		this._onMouseDown = onMouseDown.bind( this );
+		this._onMouseMove = onMouseMove.bind( this );
+
+		this._interceptControlDown = interceptControlDown.bind( this );
+		this._interceptControlUp = interceptControlUp.bind( this );
+
+		//
+
+		if ( this.domElement !== null ) {
+
+			this.connect( this.domElement );
+
+		}
+
+		this.update();
+
+	}
+
+	connect( element ) {
+
+		super.connect( element );
+
+		this.domElement.addEventListener( 'pointerdown', this._onPointerDown );
+		this.domElement.addEventListener( 'pointercancel', this._onPointerUp );
+
+		this.domElement.addEventListener( 'contextmenu', this._onContextMenu );
+		this.domElement.addEventListener( 'wheel', this._onMouseWheel, { passive: false } );
+
+		const document = this.domElement.getRootNode(); // offscreen canvas compatibility
+		document.addEventListener( 'keydown', this._interceptControlDown, { passive: true, capture: true } );
+
+		this.domElement.style.touchAction = 'none'; // disable touch scroll
+
+	}
+
+	disconnect() {
+
+		this.domElement.removeEventListener( 'pointerdown', this._onPointerDown );
+		this.domElement.removeEventListener( 'pointermove', this._onPointerMove );
+		this.domElement.removeEventListener( 'pointerup', this._onPointerUp );
+		this.domElement.removeEventListener( 'pointercancel', this._onPointerUp );
+
+		this.domElement.removeEventListener( 'wheel', this._onMouseWheel );
+		this.domElement.removeEventListener( 'contextmenu', this._onContextMenu );
+
+		this.stopListenToKeyEvents();
+
+		const document = this.domElement.getRootNode(); // offscreen canvas compatibility
+		document.removeEventListener( 'keydown', this._interceptControlDown, { capture: true } );
+
+		this.domElement.style.touchAction = 'auto';
+
+	}
+
+	dispose() {
+
+		this.disconnect();
+
+	}
+
+	/**
+	 * Get the current vertical rotation, in radians.
+	 *
+	 * @return {number} The current vertical rotation, in radians.
+	 */
+	getPolarAngle() {
+
+		return this._spherical.phi;
+
+	}
+
+	/**
+	 * Get the current horizontal rotation, in radians.
+	 *
+	 * @return {number} The current horizontal rotation, in radians.
+	 */
+	getAzimuthalAngle() {
+
+		return this._spherical.theta;
+
+	}
+
+	/**
+	 * Returns the distance from the camera to the target.
+	 *
+	 * @return {number} The distance from the camera to the target.
+	 */
+	getDistance() {
+
+		return this.object.position.distanceTo( this.target );
+
+	}
+
+	/**
+	 * Adds key event listeners to the given DOM element.
+	 * `window` is a recommended argument for using this method.
+	 *
+	 * @param {HTMLDOMElement} domElement - The DOM element
+	 */
+	listenToKeyEvents( domElement ) {
+
+		domElement.addEventListener( 'keydown', this._onKeyDown );
+		this._domElementKeyEvents = domElement;
+
+	}
+
+	/**
+	 * Removes the key event listener previously defined with `listenToKeyEvents()`.
+	 */
+	stopListenToKeyEvents() {
+
+		if ( this._domElementKeyEvents !== null ) {
+
+			this._domElementKeyEvents.removeEventListener( 'keydown', this._onKeyDown );
+			this._domElementKeyEvents = null;
+
+		}
+
+	}
+
+	/**
+	 * Save the current state of the controls. This can later be recovered with `reset()`.
+	 */
+	saveState() {
+
+		this.target0.copy( this.target );
+		this.position0.copy( this.object.position );
+		this.zoom0 = this.object.zoom;
+
+	}
+
+	/**
+	 * Reset the controls to their state from either the last time the `saveState()`
+	 * was called, or the initial state.
+	 */
+	reset() {
+
+		this.target.copy( this.target0 );
+		this.object.position.copy( this.position0 );
+		this.object.zoom = this.zoom0;
+
+		this.object.updateProjectionMatrix();
+		this.dispatchEvent( _changeEvent );
+
+		this.update();
+
+		this.state = _STATE.NONE;
+
+	}
+
+	update( deltaTime = null ) {
+
+		const position = this.object.position;
+
+		_v.copy( position ).sub( this.target );
+
+		// rotate offset to "y-axis-is-up" space
+		_v.applyQuaternion( this._quat );
+
+		// angle from z-axis around y-axis
+		this._spherical.setFromVector3( _v );
+
+		if ( this.autoRotate && this.state === _STATE.NONE ) {
+
+			this._rotateLeft( this._getAutoRotationAngle( deltaTime ) );
+
+		}
+
+		if ( this.enableDamping ) {
+
+			this._spherical.theta += this._sphericalDelta.theta * this.dampingFactor;
+			this._spherical.phi += this._sphericalDelta.phi * this.dampingFactor;
+
+		} else {
+
+			this._spherical.theta += this._sphericalDelta.theta;
+			this._spherical.phi += this._sphericalDelta.phi;
+
+		}
+
+		// restrict theta to be between desired limits
+
+		let min = this.minAzimuthAngle;
+		let max = this.maxAzimuthAngle;
+
+		if ( isFinite( min ) && isFinite( max ) ) {
+
+			if ( min < - Math.PI ) min += _twoPI; else if ( min > Math.PI ) min -= _twoPI;
+
+			if ( max < - Math.PI ) max += _twoPI; else if ( max > Math.PI ) max -= _twoPI;
+
+			if ( min <= max ) {
+
+				this._spherical.theta = Math.max( min, Math.min( max, this._spherical.theta ) );
+
+			} else {
+
+				this._spherical.theta = ( this._spherical.theta > ( min + max ) / 2 ) ?
+					Math.max( min, this._spherical.theta ) :
+					Math.min( max, this._spherical.theta );
+
+			}
+
+		}
+
+		// restrict phi to be between desired limits
+		this._spherical.phi = Math.max( this.minPolarAngle, Math.min( this.maxPolarAngle, this._spherical.phi ) );
+
+		this._spherical.makeSafe();
+
+
+		// move target to panned location
+
+		if ( this.enableDamping === true ) {
+
+			this.target.addScaledVector( this._panOffset, this.dampingFactor );
+
+		} else {
+
+			this.target.add( this._panOffset );
+
+		}
+
+		// Limit the target distance from the cursor to create a sphere around the center of interest
+		this.target.sub( this.cursor );
+		this.target.clampLength( this.minTargetRadius, this.maxTargetRadius );
+		this.target.add( this.cursor );
+
+		let zoomChanged = false;
+		// adjust the camera position based on zoom only if we're not zooming to the cursor or if it's an ortho camera
+		// we adjust zoom later in these cases
+		if ( this.zoomToCursor && this._performCursorZoom || this.object.isOrthographicCamera ) {
+
+			this._spherical.radius = this._clampDistance( this._spherical.radius );
+
+		} else {
+
+			const prevRadius = this._spherical.radius;
+			this._spherical.radius = this._clampDistance( this._spherical.radius * this._scale );
+			zoomChanged = prevRadius != this._spherical.radius;
+
+		}
+
+		_v.setFromSpherical( this._spherical );
+
+		// rotate offset back to "camera-up-vector-is-up" space
+		_v.applyQuaternion( this._quatInverse );
+
+		position.copy( this.target ).add( _v );
+
+		this.object.lookAt( this.target );
+
+		if ( this.enableDamping === true ) {
+
+			this._sphericalDelta.theta *= ( 1 - this.dampingFactor );
+			this._sphericalDelta.phi *= ( 1 - this.dampingFactor );
+
+			this._panOffset.multiplyScalar( 1 - this.dampingFactor );
+
+		} else {
+
+			this._sphericalDelta.set( 0, 0, 0 );
+
+			this._panOffset.set( 0, 0, 0 );
+
+		}
+
+		// adjust camera position
+		if ( this.zoomToCursor && this._performCursorZoom ) {
+
+			let newRadius = null;
+			if ( this.object.isPerspectiveCamera ) {
+
+				// move the camera down the pointer ray
+				// this method avoids floating point error
+				const prevRadius = _v.length();
+				newRadius = this._clampDistance( prevRadius * this._scale );
+
+				const radiusDelta = prevRadius - newRadius;
+				this.object.position.addScaledVector( this._dollyDirection, radiusDelta );
+				this.object.updateMatrixWorld();
+
+				zoomChanged = !! radiusDelta;
+
+			} else if ( this.object.isOrthographicCamera ) {
+
+				// adjust the ortho camera position based on zoom changes
+				const mouseBefore = new Vector3( this._mouse.x, this._mouse.y, 0 );
+				mouseBefore.unproject( this.object );
+
+				const prevZoom = this.object.zoom;
+				this.object.zoom = Math.max( this.minZoom, Math.min( this.maxZoom, this.object.zoom / this._scale ) );
+				this.object.updateProjectionMatrix();
+
+				zoomChanged = prevZoom !== this.object.zoom;
+
+				const mouseAfter = new Vector3( this._mouse.x, this._mouse.y, 0 );
+				mouseAfter.unproject( this.object );
+
+				this.object.position.sub( mouseAfter ).add( mouseBefore );
+				this.object.updateMatrixWorld();
+
+				newRadius = _v.length();
+
+			} else {
+
+				console.warn( 'WARNING: OrbitControls.js encountered an unknown camera type - zoom to cursor disabled.' );
+				this.zoomToCursor = false;
+
+			}
+
+			// handle the placement of the target
+			if ( newRadius !== null ) {
+
+				if ( this.screenSpacePanning ) {
+
+					// position the orbit target in front of the new camera position
+					this.target.set( 0, 0, - 1 )
+						.transformDirection( this.object.matrix )
+						.multiplyScalar( newRadius )
+						.add( this.object.position );
+
+				} else {
+
+					// get the ray and translation plane to compute target
+					_ray.origin.copy( this.object.position );
+					_ray.direction.set( 0, 0, - 1 ).transformDirection( this.object.matrix );
+
+					// if the camera is 20 degrees above the horizon then don't adjust the focus target to avoid
+					// extremely large values
+					if ( Math.abs( this.object.up.dot( _ray.direction ) ) < _TILT_LIMIT ) {
+
+						this.object.lookAt( this.target );
+
+					} else {
+
+						_plane.setFromNormalAndCoplanarPoint( this.object.up, this.target );
+						_ray.intersectPlane( _plane, this.target );
+
+					}
+
+				}
+
+			}
+
+		} else if ( this.object.isOrthographicCamera ) {
+
+			const prevZoom = this.object.zoom;
+			this.object.zoom = Math.max( this.minZoom, Math.min( this.maxZoom, this.object.zoom / this._scale ) );
+
+			if ( prevZoom !== this.object.zoom ) {
+
+				this.object.updateProjectionMatrix();
+				zoomChanged = true;
+
+			}
+
+		}
+
+		this._scale = 1;
+		this._performCursorZoom = false;
+
+		// update condition is:
+		// min(camera displacement, camera rotation in radians)^2 > EPS
+		// using small-angle approximation cos(x/2) = 1 - x^2 / 8
+
+		if ( zoomChanged ||
+			this._lastPosition.distanceToSquared( this.object.position ) > _EPS ||
+			8 * ( 1 - this._lastQuaternion.dot( this.object.quaternion ) ) > _EPS ||
+			this._lastTargetPosition.distanceToSquared( this.target ) > _EPS ) {
+
+			this.dispatchEvent( _changeEvent );
+
+			this._lastPosition.copy( this.object.position );
+			this._lastQuaternion.copy( this.object.quaternion );
+			this._lastTargetPosition.copy( this.target );
+
+			return true;
+
+		}
+
+		return false;
+
+	}
+
+	_getAutoRotationAngle( deltaTime ) {
+
+		if ( deltaTime !== null ) {
+
+			return ( _twoPI / 60 * this.autoRotateSpeed ) * deltaTime;
+
+		} else {
+
+			return _twoPI / 60 / 60 * this.autoRotateSpeed;
+
+		}
+
+	}
+
+	_getZoomScale( delta ) {
+
+		const normalizedDelta = Math.abs( delta * 0.01 );
+		return Math.pow( 0.95, this.zoomSpeed * normalizedDelta );
+
+	}
+
+	_rotateLeft( angle ) {
+
+		this._sphericalDelta.theta -= angle;
+
+	}
+
+	_rotateUp( angle ) {
+
+		this._sphericalDelta.phi -= angle;
+
+	}
+
+	_panLeft( distance, objectMatrix ) {
+
+		_v.setFromMatrixColumn( objectMatrix, 0 ); // get X column of objectMatrix
+		_v.multiplyScalar( - distance );
+
+		this._panOffset.add( _v );
+
+	}
+
+	_panUp( distance, objectMatrix ) {
+
+		if ( this.screenSpacePanning === true ) {
+
+			_v.setFromMatrixColumn( objectMatrix, 1 );
+
+		} else {
+
+			_v.setFromMatrixColumn( objectMatrix, 0 );
+			_v.crossVectors( this.object.up, _v );
+
+		}
+
+		_v.multiplyScalar( distance );
+
+		this._panOffset.add( _v );
+
+	}
+
+	// deltaX and deltaY are in pixels; right and down are positive
+	_pan( deltaX, deltaY ) {
+
+		const element = this.domElement;
+
+		if ( this.object.isPerspectiveCamera ) {
+
+			// perspective
+			const position = this.object.position;
+			_v.copy( position ).sub( this.target );
+			let targetDistance = _v.length();
+
+			// half of the fov is center to top of screen
+			targetDistance *= Math.tan( ( this.object.fov / 2 ) * Math.PI / 180.0 );
+
+			// we use only clientHeight here so aspect ratio does not distort speed
+			this._panLeft( 2 * deltaX * targetDistance / element.clientHeight, this.object.matrix );
+			this._panUp( 2 * deltaY * targetDistance / element.clientHeight, this.object.matrix );
+
+		} else if ( this.object.isOrthographicCamera ) {
+
+			// orthographic
+			this._panLeft( deltaX * ( this.object.right - this.object.left ) / this.object.zoom / element.clientWidth, this.object.matrix );
+			this._panUp( deltaY * ( this.object.top - this.object.bottom ) / this.object.zoom / element.clientHeight, this.object.matrix );
+
+		} else {
+
+			// camera neither orthographic nor perspective
+			console.warn( 'WARNING: OrbitControls.js encountered an unknown camera type - pan disabled.' );
+			this.enablePan = false;
+
+		}
+
+	}
+
+	_dollyOut( dollyScale ) {
+
+		if ( this.object.isPerspectiveCamera || this.object.isOrthographicCamera ) {
+
+			this._scale /= dollyScale;
+
+		} else {
+
+			console.warn( 'WARNING: OrbitControls.js encountered an unknown camera type - dolly/zoom disabled.' );
+			this.enableZoom = false;
+
+		}
+
+	}
+
+	_dollyIn( dollyScale ) {
+
+		if ( this.object.isPerspectiveCamera || this.object.isOrthographicCamera ) {
+
+			this._scale *= dollyScale;
+
+		} else {
+
+			console.warn( 'WARNING: OrbitControls.js encountered an unknown camera type - dolly/zoom disabled.' );
+			this.enableZoom = false;
+
+		}
+
+	}
+
+	_updateZoomParameters( x, y ) {
+
+		if ( ! this.zoomToCursor ) {
+
+			return;
+
+		}
+
+		this._performCursorZoom = true;
+
+		const rect = this.domElement.getBoundingClientRect();
+		const dx = x - rect.left;
+		const dy = y - rect.top;
+		const w = rect.width;
+		const h = rect.height;
+
+		this._mouse.x = ( dx / w ) * 2 - 1;
+		this._mouse.y = - ( dy / h ) * 2 + 1;
+
+		this._dollyDirection.set( this._mouse.x, this._mouse.y, 1 ).unproject( this.object ).sub( this.object.position ).normalize();
+
+	}
+
+	_clampDistance( dist ) {
+
+		return Math.max( this.minDistance, Math.min( this.maxDistance, dist ) );
+
+	}
+
+	//
+	// event callbacks - update the object state
+	//
+
+	_handleMouseDownRotate( event ) {
+
+		this._rotateStart.set( event.clientX, event.clientY );
+
+	}
+
+	_handleMouseDownDolly( event ) {
+
+		this._updateZoomParameters( event.clientX, event.clientX );
+		this._dollyStart.set( event.clientX, event.clientY );
+
+	}
+
+	_handleMouseDownPan( event ) {
+
+		this._panStart.set( event.clientX, event.clientY );
+
+	}
+
+	_handleMouseMoveRotate( event ) {
+
+		this._rotateEnd.set( event.clientX, event.clientY );
+
+		this._rotateDelta.subVectors( this._rotateEnd, this._rotateStart ).multiplyScalar( this.rotateSpeed );
+
+		const element = this.domElement;
+
+		this._rotateLeft( _twoPI * this._rotateDelta.x / element.clientHeight ); // yes, height
+
+		this._rotateUp( _twoPI * this._rotateDelta.y / element.clientHeight );
+
+		this._rotateStart.copy( this._rotateEnd );
+
+		this.update();
+
+	}
+
+	_handleMouseMoveDolly( event ) {
+
+		this._dollyEnd.set( event.clientX, event.clientY );
+
+		this._dollyDelta.subVectors( this._dollyEnd, this._dollyStart );
+
+		if ( this._dollyDelta.y > 0 ) {
+
+			this._dollyOut( this._getZoomScale( this._dollyDelta.y ) );
+
+		} else if ( this._dollyDelta.y < 0 ) {
+
+			this._dollyIn( this._getZoomScale( this._dollyDelta.y ) );
+
+		}
+
+		this._dollyStart.copy( this._dollyEnd );
+
+		this.update();
+
+	}
+
+	_handleMouseMovePan( event ) {
+
+		this._panEnd.set( event.clientX, event.clientY );
+
+		this._panDelta.subVectors( this._panEnd, this._panStart ).multiplyScalar( this.panSpeed );
+
+		this._pan( this._panDelta.x, this._panDelta.y );
+
+		this._panStart.copy( this._panEnd );
+
+		this.update();
+
+	}
+
+	_handleMouseWheel( event ) {
+
+		this._updateZoomParameters( event.clientX, event.clientY );
+
+		if ( event.deltaY < 0 ) {
+
+			this._dollyIn( this._getZoomScale( event.deltaY ) );
+
+		} else if ( event.deltaY > 0 ) {
+
+			this._dollyOut( this._getZoomScale( event.deltaY ) );
+
+		}
+
+		this.update();
+
+	}
+
+	_handleKeyDown( event ) {
+
+		let needsUpdate = false;
+
+		switch ( event.code ) {
+
+			case this.keys.UP:
+
+				if ( event.ctrlKey || event.metaKey || event.shiftKey ) {
+
+					if ( this.enableRotate ) {
+
+						this._rotateUp( _twoPI * this.keyRotateSpeed / this.domElement.clientHeight );
+
+					}
+
+				} else {
+
+					if ( this.enablePan ) {
+
+						this._pan( 0, this.keyPanSpeed );
+
+					}
+
+				}
+
+				needsUpdate = true;
+				break;
+
+			case this.keys.BOTTOM:
+
+				if ( event.ctrlKey || event.metaKey || event.shiftKey ) {
+
+					if ( this.enableRotate ) {
+
+						this._rotateUp( - _twoPI * this.keyRotateSpeed / this.domElement.clientHeight );
+
+					}
+
+				} else {
+
+					if ( this.enablePan ) {
+
+						this._pan( 0, - this.keyPanSpeed );
+
+					}
+
+				}
+
+				needsUpdate = true;
+				break;
+
+			case this.keys.LEFT:
+
+				if ( event.ctrlKey || event.metaKey || event.shiftKey ) {
+
+					if ( this.enableRotate ) {
+
+						this._rotateLeft( _twoPI * this.keyRotateSpeed / this.domElement.clientHeight );
+
+					}
+
+				} else {
+
+					if ( this.enablePan ) {
+
+						this._pan( this.keyPanSpeed, 0 );
+
+					}
+
+				}
+
+				needsUpdate = true;
+				break;
+
+			case this.keys.RIGHT:
+
+				if ( event.ctrlKey || event.metaKey || event.shiftKey ) {
+
+					if ( this.enableRotate ) {
+
+						this._rotateLeft( - _twoPI * this.keyRotateSpeed / this.domElement.clientHeight );
+
+					}
+
+				} else {
+
+					if ( this.enablePan ) {
+
+						this._pan( - this.keyPanSpeed, 0 );
+
+					}
+
+				}
+
+				needsUpdate = true;
+				break;
+
+		}
+
+		if ( needsUpdate ) {
+
+			// prevent the browser from scrolling on cursor keys
+			event.preventDefault();
+
+			this.update();
+
+		}
+
+
+	}
+
+	_handleTouchStartRotate( event ) {
+
+		if ( this._pointers.length === 1 ) {
+
+			this._rotateStart.set( event.pageX, event.pageY );
+
+		} else {
+
+			const position = this._getSecondPointerPosition( event );
+
+			const x = 0.5 * ( event.pageX + position.x );
+			const y = 0.5 * ( event.pageY + position.y );
+
+			this._rotateStart.set( x, y );
+
+		}
+
+	}
+
+	_handleTouchStartPan( event ) {
+
+		if ( this._pointers.length === 1 ) {
+
+			this._panStart.set( event.pageX, event.pageY );
+
+		} else {
+
+			const position = this._getSecondPointerPosition( event );
+
+			const x = 0.5 * ( event.pageX + position.x );
+			const y = 0.5 * ( event.pageY + position.y );
+
+			this._panStart.set( x, y );
+
+		}
+
+	}
+
+	_handleTouchStartDolly( event ) {
+
+		const position = this._getSecondPointerPosition( event );
+
+		const dx = event.pageX - position.x;
+		const dy = event.pageY - position.y;
+
+		const distance = Math.sqrt( dx * dx + dy * dy );
+
+		this._dollyStart.set( 0, distance );
+
+	}
+
+	_handleTouchStartDollyPan( event ) {
+
+		if ( this.enableZoom ) this._handleTouchStartDolly( event );
+
+		if ( this.enablePan ) this._handleTouchStartPan( event );
+
+	}
+
+	_handleTouchStartDollyRotate( event ) {
+
+		if ( this.enableZoom ) this._handleTouchStartDolly( event );
+
+		if ( this.enableRotate ) this._handleTouchStartRotate( event );
+
+	}
+
+	_handleTouchMoveRotate( event ) {
+
+		if ( this._pointers.length == 1 ) {
+
+			this._rotateEnd.set( event.pageX, event.pageY );
+
+		} else {
+
+			const position = this._getSecondPointerPosition( event );
+
+			const x = 0.5 * ( event.pageX + position.x );
+			const y = 0.5 * ( event.pageY + position.y );
+
+			this._rotateEnd.set( x, y );
+
+		}
+
+		this._rotateDelta.subVectors( this._rotateEnd, this._rotateStart ).multiplyScalar( this.rotateSpeed );
+
+		const element = this.domElement;
+
+		this._rotateLeft( _twoPI * this._rotateDelta.x / element.clientHeight ); // yes, height
+
+		this._rotateUp( _twoPI * this._rotateDelta.y / element.clientHeight );
+
+		this._rotateStart.copy( this._rotateEnd );
+
+	}
+
+	_handleTouchMovePan( event ) {
+
+		if ( this._pointers.length === 1 ) {
+
+			this._panEnd.set( event.pageX, event.pageY );
+
+		} else {
+
+			const position = this._getSecondPointerPosition( event );
+
+			const x = 0.5 * ( event.pageX + position.x );
+			const y = 0.5 * ( event.pageY + position.y );
+
+			this._panEnd.set( x, y );
+
+		}
+
+		this._panDelta.subVectors( this._panEnd, this._panStart ).multiplyScalar( this.panSpeed );
+
+		this._pan( this._panDelta.x, this._panDelta.y );
+
+		this._panStart.copy( this._panEnd );
+
+	}
+
+	_handleTouchMoveDolly( event ) {
+
+		const position = this._getSecondPointerPosition( event );
+
+		const dx = event.pageX - position.x;
+		const dy = event.pageY - position.y;
+
+		const distance = Math.sqrt( dx * dx + dy * dy );
+
+		this._dollyEnd.set( 0, distance );
+
+		this._dollyDelta.set( 0, Math.pow( this._dollyEnd.y / this._dollyStart.y, this.zoomSpeed ) );
+
+		this._dollyOut( this._dollyDelta.y );
+
+		this._dollyStart.copy( this._dollyEnd );
+
+		const centerX = ( event.pageX + position.x ) * 0.5;
+		const centerY = ( event.pageY + position.y ) * 0.5;
+
+		this._updateZoomParameters( centerX, centerY );
+
+	}
+
+	_handleTouchMoveDollyPan( event ) {
+
+		if ( this.enableZoom ) this._handleTouchMoveDolly( event );
+
+		if ( this.enablePan ) this._handleTouchMovePan( event );
+
+	}
+
+	_handleTouchMoveDollyRotate( event ) {
+
+		if ( this.enableZoom ) this._handleTouchMoveDolly( event );
+
+		if ( this.enableRotate ) this._handleTouchMoveRotate( event );
+
+	}
+
+	// pointers
+
+	_addPointer( event ) {
+
+		this._pointers.push( event.pointerId );
+
+	}
+
+	_removePointer( event ) {
+
+		delete this._pointerPositions[ event.pointerId ];
+
+		for ( let i = 0; i < this._pointers.length; i ++ ) {
+
+			if ( this._pointers[ i ] == event.pointerId ) {
+
+				this._pointers.splice( i, 1 );
+				return;
+
+			}
+
+		}
+
+	}
+
+	_isTrackingPointer( event ) {
+
+		for ( let i = 0; i < this._pointers.length; i ++ ) {
+
+			if ( this._pointers[ i ] == event.pointerId ) return true;
+
+		}
+
+		return false;
+
+	}
+
+	_trackPointer( event ) {
+
+		let position = this._pointerPositions[ event.pointerId ];
+
+		if ( position === undefined ) {
+
+			position = new Vector2();
+			this._pointerPositions[ event.pointerId ] = position;
+
+		}
+
+		position.set( event.pageX, event.pageY );
+
+	}
+
+	_getSecondPointerPosition( event ) {
+
+		const pointerId = ( event.pointerId === this._pointers[ 0 ] ) ? this._pointers[ 1 ] : this._pointers[ 0 ];
+
+		return this._pointerPositions[ pointerId ];
+
+	}
+
+	//
+
+	_customWheelEvent( event ) {
+
+		const mode = event.deltaMode;
+
+		// minimal wheel event altered to meet delta-zoom demand
+		const newEvent = {
+			clientX: event.clientX,
+			clientY: event.clientY,
+			deltaY: event.deltaY,
+		};
+
+		switch ( mode ) {
+
+			case 1: // LINE_MODE
+				newEvent.deltaY *= 16;
+				break;
+
+			case 2: // PAGE_MODE
+				newEvent.deltaY *= 100;
+				break;
+
+		}
+
+		// detect if event was triggered by pinching
+		if ( event.ctrlKey && ! this._controlActive ) {
+
+			newEvent.deltaY *= 10;
+
+		}
+
+		return newEvent;
+
+	}
+
+}
+
+function onPointerDown( event ) {
+
+	if ( this.enabled === false ) return;
+
+	if ( this._pointers.length === 0 ) {
+
+		this.domElement.setPointerCapture( event.pointerId );
+
+		this.domElement.addEventListener( 'pointermove', this._onPointerMove );
+		this.domElement.addEventListener( 'pointerup', this._onPointerUp );
+
+	}
+
+	//
+
+	if ( this._isTrackingPointer( event ) ) return;
+
+	//
+
+	this._addPointer( event );
+
+	if ( event.pointerType === 'touch' ) {
+
+		this._onTouchStart( event );
+
+	} else {
+
+		this._onMouseDown( event );
+
+	}
+
+}
+
+function onPointerMove( event ) {
+
+	if ( this.enabled === false ) return;
+
+	if ( event.pointerType === 'touch' ) {
+
+		this._onTouchMove( event );
+
+	} else {
+
+		this._onMouseMove( event );
+
+	}
+
+}
+
+function onPointerUp( event ) {
+
+	this._removePointer( event );
+
+	switch ( this._pointers.length ) {
+
+		case 0:
+
+			this.domElement.releasePointerCapture( event.pointerId );
+
+			this.domElement.removeEventListener( 'pointermove', this._onPointerMove );
+			this.domElement.removeEventListener( 'pointerup', this._onPointerUp );
+
+			this.dispatchEvent( _endEvent );
+
+			this.state = _STATE.NONE;
+
+			break;
+
+		case 1:
+
+			const pointerId = this._pointers[ 0 ];
+			const position = this._pointerPositions[ pointerId ];
+
+			// minimal placeholder event - allows state correction on pointer-up
+			this._onTouchStart( { pointerId: pointerId, pageX: position.x, pageY: position.y } );
+
+			break;
+
+	}
+
+}
+
+function onMouseDown( event ) {
+
+	let mouseAction;
+
+	switch ( event.button ) {
+
+		case 0:
+
+			mouseAction = this.mouseButtons.LEFT;
+			break;
+
+		case 1:
+
+			mouseAction = this.mouseButtons.MIDDLE;
+			break;
+
+		case 2:
+
+			mouseAction = this.mouseButtons.RIGHT;
+			break;
+
+		default:
+
+			mouseAction = - 1;
+
+	}
+
+	switch ( mouseAction ) {
+
+		case MOUSE.DOLLY:
+
+			if ( this.enableZoom === false ) return;
+
+			this._handleMouseDownDolly( event );
+
+			this.state = _STATE.DOLLY;
+
+			break;
+
+		case MOUSE.ROTATE:
+
+			if ( event.ctrlKey || event.metaKey || event.shiftKey ) {
+
+				if ( this.enablePan === false ) return;
+
+				this._handleMouseDownPan( event );
+
+				this.state = _STATE.PAN;
+
+			} else {
+
+				if ( this.enableRotate === false ) return;
+
+				this._handleMouseDownRotate( event );
+
+				this.state = _STATE.ROTATE;
+
+			}
+
+			break;
+
+		case MOUSE.PAN:
+
+			if ( event.ctrlKey || event.metaKey || event.shiftKey ) {
+
+				if ( this.enableRotate === false ) return;
+
+				this._handleMouseDownRotate( event );
+
+				this.state = _STATE.ROTATE;
+
+			} else {
+
+				if ( this.enablePan === false ) return;
+
+				this._handleMouseDownPan( event );
+
+				this.state = _STATE.PAN;
+
+			}
+
+			break;
+
+		default:
+
+			this.state = _STATE.NONE;
+
+	}
+
+	if ( this.state !== _STATE.NONE ) {
+
+		this.dispatchEvent( _startEvent );
+
+	}
+
+}
+
+function onMouseMove( event ) {
+
+	switch ( this.state ) {
+
+		case _STATE.ROTATE:
+
+			if ( this.enableRotate === false ) return;
+
+			this._handleMouseMoveRotate( event );
+
+			break;
+
+		case _STATE.DOLLY:
+
+			if ( this.enableZoom === false ) return;
+
+			this._handleMouseMoveDolly( event );
+
+			break;
+
+		case _STATE.PAN:
+
+			if ( this.enablePan === false ) return;
+
+			this._handleMouseMovePan( event );
+
+			break;
+
+	}
+
+}
+
+function onMouseWheel( event ) {
+
+	if ( this.enabled === false || this.enableZoom === false || this.state !== _STATE.NONE ) return;
+
+	event.preventDefault();
+
+	this.dispatchEvent( _startEvent );
+
+	this._handleMouseWheel( this._customWheelEvent( event ) );
+
+	this.dispatchEvent( _endEvent );
+
+}
+
+function onKeyDown( event ) {
+
+	if ( this.enabled === false ) return;
+
+	this._handleKeyDown( event );
+
+}
+
+function onTouchStart( event ) {
+
+	this._trackPointer( event );
+
+	switch ( this._pointers.length ) {
+
+		case 1:
+
+			switch ( this.touches.ONE ) {
+
+				case TOUCH.ROTATE:
+
+					if ( this.enableRotate === false ) return;
+
+					this._handleTouchStartRotate( event );
+
+					this.state = _STATE.TOUCH_ROTATE;
+
+					break;
+
+				case TOUCH.PAN:
+
+					if ( this.enablePan === false ) return;
+
+					this._handleTouchStartPan( event );
+
+					this.state = _STATE.TOUCH_PAN;
+
+					break;
+
+				default:
+
+					this.state = _STATE.NONE;
+
+			}
+
+			break;
+
+		case 2:
+
+			switch ( this.touches.TWO ) {
+
+				case TOUCH.DOLLY_PAN:
+
+					if ( this.enableZoom === false && this.enablePan === false ) return;
+
+					this._handleTouchStartDollyPan( event );
+
+					this.state = _STATE.TOUCH_DOLLY_PAN;
+
+					break;
+
+				case TOUCH.DOLLY_ROTATE:
+
+					if ( this.enableZoom === false && this.enableRotate === false ) return;
+
+					this._handleTouchStartDollyRotate( event );
+
+					this.state = _STATE.TOUCH_DOLLY_ROTATE;
+
+					break;
+
+				default:
+
+					this.state = _STATE.NONE;
+
+			}
+
+			break;
+
+		default:
+
+			this.state = _STATE.NONE;
+
+	}
+
+	if ( this.state !== _STATE.NONE ) {
+
+		this.dispatchEvent( _startEvent );
+
+	}
+
+}
+
+function onTouchMove( event ) {
+
+	this._trackPointer( event );
+
+	switch ( this.state ) {
+
+		case _STATE.TOUCH_ROTATE:
+
+			if ( this.enableRotate === false ) return;
+
+			this._handleTouchMoveRotate( event );
+
+			this.update();
+
+			break;
+
+		case _STATE.TOUCH_PAN:
+
+			if ( this.enablePan === false ) return;
+
+			this._handleTouchMovePan( event );
+
+			this.update();
+
+			break;
+
+		case _STATE.TOUCH_DOLLY_PAN:
+
+			if ( this.enableZoom === false && this.enablePan === false ) return;
+
+			this._handleTouchMoveDollyPan( event );
+
+			this.update();
+
+			break;
+
+		case _STATE.TOUCH_DOLLY_ROTATE:
+
+			if ( this.enableZoom === false && this.enableRotate === false ) return;
+
+			this._handleTouchMoveDollyRotate( event );
+
+			this.update();
+
+			break;
+
+		default:
+
+			this.state = _STATE.NONE;
+
+	}
+
+}
+
+function onContextMenu( event ) {
+
+	if ( this.enabled === false ) return;
+
+	event.preventDefault();
+
+}
+
+function interceptControlDown( event ) {
+
+	if ( event.key === 'Control' ) {
+
+		this._controlActive = true;
+
+		const document = this.domElement.getRootNode(); // offscreen canvas compatibility
+
+		document.addEventListener( 'keyup', this._interceptControlUp, { passive: true, capture: true } );
+
+	}
+
+}
+
+function interceptControlUp( event ) {
+
+	if ( event.key === 'Control' ) {
+
+		this._controlActive = false;
+
+		const document = this.domElement.getRootNode(); // offscreen canvas compatibility
+
+		document.removeEventListener( 'keyup', this._interceptControlUp, { passive: true, capture: true } );
+
+	}
+
+}
+
+export { OrbitControls };

vendor/three/environments/RoomEnvironment.js

@@ -0,0 +1,182 @@
+import {
+ 	BackSide,
+ 	BoxGeometry,
+ 	InstancedMesh,
+ 	Mesh,
+	MeshLambertMaterial,
+ 	MeshStandardMaterial,
+ 	PointLight,
+ 	Scene,
+ 	Object3D,
+} from 'three';
+
+/**
+ * This class represents a scene with a basic room setup that can be used as
+ * input for {@link PMREMGenerator#fromScene}. The resulting PMREM represents the room's
+ * lighting and can be used for Image Based Lighting by assigning it to {@link Scene#environment}
+ * or directly as an environment map to PBR materials.
+ *
+ * The implementation is based on the [EnvironmentScene](https://github.com/google/model-viewer/blob/master/packages/model-viewer/src/three-components/EnvironmentScene.ts)
+ * component from the `model-viewer` project.
+ *
+ * ```js
+ * const environment = new RoomEnvironment();
+ * const pmremGenerator = new THREE.PMREMGenerator( renderer );
+ *
+ * const envMap = pmremGenerator.fromScene( environment ).texture;
+ * scene.environment = envMap;
+ * ```
+ *
+ * @augments Scene
+ * @three_import import { RoomEnvironment } from 'three/addons/environments/RoomEnvironment.js';
+ */
+class RoomEnvironment extends Scene {
+
+	constructor() {
+
+		super();
+
+		const geometry = new BoxGeometry();
+		geometry.deleteAttribute( 'uv' );
+
+		const roomMaterial = new MeshStandardMaterial( { side: BackSide } );
+		const boxMaterial = new MeshStandardMaterial();
+
+		const mainLight = new PointLight( 0xffffff, 900, 28, 2 );
+		mainLight.position.set( 0.418, 16.199, 0.300 );
+		this.add( mainLight );
+
+		const room = new Mesh( geometry, roomMaterial );
+		room.position.set( - 0.757, 13.219, 0.717 );
+		room.scale.set( 31.713, 28.305, 28.591 );
+		this.add( room );
+
+		const boxes = new InstancedMesh( geometry, boxMaterial, 6 );
+		const transform = new Object3D();
+
+		// box1
+		transform.position.set( - 10.906, 2.009, 1.846 );
+		transform.rotation.set( 0, - 0.195, 0 );
+		transform.scale.set( 2.328, 7.905, 4.651 );
+		transform.updateMatrix();
+		boxes.setMatrixAt( 0, transform.matrix );
+
+		// box2
+		transform.position.set( - 5.607, - 0.754, - 0.758 );
+		transform.rotation.set( 0, 0.994, 0 );
+		transform.scale.set( 1.970, 1.534, 3.955 );
+		transform.updateMatrix();
+		boxes.setMatrixAt( 1, transform.matrix );
+
+		// box3
+		transform.position.set( 6.167, 0.857, 7.803 );
+		transform.rotation.set( 0, 0.561, 0 );
+		transform.scale.set( 3.927, 6.285, 3.687 );
+		transform.updateMatrix();
+		boxes.setMatrixAt( 2, transform.matrix );
+
+		// box4
+		transform.position.set( - 2.017, 0.018, 6.124 );
+		transform.rotation.set( 0, 0.333, 0 );
+		transform.scale.set( 2.002, 4.566, 2.064 );
+		transform.updateMatrix();
+		boxes.setMatrixAt( 3, transform.matrix );
+
+		// box5
+		transform.position.set( 2.291, - 0.756, - 2.621 );
+		transform.rotation.set( 0, - 0.286, 0 );
+		transform.scale.set( 1.546, 1.552, 1.496 );
+		transform.updateMatrix();
+		boxes.setMatrixAt( 4, transform.matrix );
+
+		// box6
+		transform.position.set( - 2.193, - 0.369, - 5.547 );
+		transform.rotation.set( 0, 0.516, 0 );
+		transform.scale.set( 3.875, 3.487, 2.986 );
+		transform.updateMatrix();
+		boxes.setMatrixAt( 5, transform.matrix );
+
+		this.add( boxes );
+
+
+		// -x right
+		const light1 = new Mesh( geometry, createAreaLightMaterial( 50 ) );
+		light1.position.set( - 16.116, 14.37, 8.208 );
+		light1.scale.set( 0.1, 2.428, 2.739 );
+		this.add( light1 );
+
+		// -x left
+		const light2 = new Mesh( geometry, createAreaLightMaterial( 50 ) );
+		light2.position.set( - 16.109, 18.021, - 8.207 );
+		light2.scale.set( 0.1, 2.425, 2.751 );
+		this.add( light2 );
+
+		// +x
+		const light3 = new Mesh( geometry, createAreaLightMaterial( 17 ) );
+		light3.position.set( 14.904, 12.198, - 1.832 );
+		light3.scale.set( 0.15, 4.265, 6.331 );
+		this.add( light3 );
+
+		// +z
+		const light4 = new Mesh( geometry, createAreaLightMaterial( 43 ) );
+		light4.position.set( - 0.462, 8.89, 14.520 );
+		light4.scale.set( 4.38, 5.441, 0.088 );
+		this.add( light4 );
+
+		// -z
+		const light5 = new Mesh( geometry, createAreaLightMaterial( 20 ) );
+		light5.position.set( 3.235, 11.486, - 12.541 );
+		light5.scale.set( 2.5, 2.0, 0.1 );
+		this.add( light5 );
+
+		// +y
+		const light6 = new Mesh( geometry, createAreaLightMaterial( 100 ) );
+		light6.position.set( 0.0, 20.0, 0.0 );
+		light6.scale.set( 1.0, 0.1, 1.0 );
+		this.add( light6 );
+
+	}
+
+	/**
+	 * Frees internal resources. This method should be called
+	 * when the environment is no longer required.
+	 */
+	dispose() {
+
+		const resources = new Set();
+
+		this.traverse( ( object ) => {
+
+			if ( object.isMesh ) {
+
+				resources.add( object.geometry );
+				resources.add( object.material );
+
+			}
+
+		} );
+
+		for ( const resource of resources ) {
+
+			resource.dispose();
+
+		}
+
+	}
+
+}
+
+function createAreaLightMaterial( intensity ) {
+
+	// create an emissive-only material. see #31348
+	const material = new MeshLambertMaterial( {
+		color: 0x000000,
+		emissive: 0xffffff,
+		emissiveIntensity: intensity
+	} );
+
+	return material;
+
+}
+
+export { RoomEnvironment };

vendor/three/utils/BufferGeometryUtils.js

@@ -0,0 +1,1435 @@
+import {
+	BufferAttribute,
+	BufferGeometry,
+	Float32BufferAttribute,
+	InstancedBufferAttribute,
+	InterleavedBuffer,
+	InterleavedBufferAttribute,
+	TriangleFanDrawMode,
+	TriangleStripDrawMode,
+	TrianglesDrawMode,
+	Vector3,
+} from 'three';
+
+/**
+ * @module BufferGeometryUtils
+ * @three_import import * as BufferGeometryUtils from 'three/addons/utils/BufferGeometryUtils.js';
+ */
+
+/**
+ * Computes vertex tangents using the MikkTSpace algorithm. MikkTSpace generates the same tangents consistently,
+ * and is used in most modelling tools and normal map bakers. Use MikkTSpace for materials with normal maps,
+ * because inconsistent tangents may lead to subtle visual issues in the normal map, particularly around mirrored
+ * UV seams.
+ *
+ * In comparison to this method, {@link BufferGeometry#computeTangents} (a custom algorithm) generates tangents that
+ * probably will not match the tangents in other software. The custom algorithm is sufficient for general use with a
+ * custom material, and may be faster than MikkTSpace.
+ *
+ * Returns the original BufferGeometry. Indexed geometries will be de-indexed. Requires position, normal, and uv attributes.
+ *
+ * @param {BufferGeometry} geometry - The geometry to compute tangents for.
+ * @param {Object} MikkTSpace - Instance of `examples/jsm/libs/mikktspace.module.js`, or `mikktspace` npm package.
+ * Await `MikkTSpace.ready` before use.
+ * @param {boolean} [negateSign=true] - Whether to negate the sign component (.w) of each tangent.
+ * Required for normal map conventions in some formats, including glTF.
+ * @return {BufferGeometry} The updated geometry.
+ */
+function computeMikkTSpaceTangents( geometry, MikkTSpace, negateSign = true ) {
+
+	if ( ! MikkTSpace || ! MikkTSpace.isReady ) {
+
+		throw new Error( 'BufferGeometryUtils: Initialized MikkTSpace library required.' );
+
+	}
+
+	if ( ! geometry.hasAttribute( 'position' ) || ! geometry.hasAttribute( 'normal' ) || ! geometry.hasAttribute( 'uv' ) ) {
+
+		throw new Error( 'BufferGeometryUtils: Tangents require "position", "normal", and "uv" attributes.' );
+
+	}
+
+	function getAttributeArray( attribute ) {
+
+		if ( attribute.normalized || attribute.isInterleavedBufferAttribute ) {
+
+			const dstArray = new Float32Array( attribute.count * attribute.itemSize );
+
+			for ( let i = 0, j = 0; i < attribute.count; i ++ ) {
+
+				dstArray[ j ++ ] = attribute.getX( i );
+				dstArray[ j ++ ] = attribute.getY( i );
+
+				if ( attribute.itemSize > 2 ) {
+
+					dstArray[ j ++ ] = attribute.getZ( i );
+
+				}
+
+			}
+
+			return dstArray;
+
+		}
+
+		if ( attribute.array instanceof Float32Array ) {
+
+			return attribute.array;
+
+		}
+
+		return new Float32Array( attribute.array );
+
+	}
+
+	// MikkTSpace algorithm requires non-indexed input.
+
+	const _geometry = geometry.index ? geometry.toNonIndexed() : geometry;
+
+	// Compute vertex tangents.
+
+	const tangents = MikkTSpace.generateTangents(
+
+		getAttributeArray( _geometry.attributes.position ),
+		getAttributeArray( _geometry.attributes.normal ),
+		getAttributeArray( _geometry.attributes.uv )
+
+	);
+
+	// Texture coordinate convention of glTF differs from the apparent
+	// default of the MikkTSpace library; .w component must be flipped.
+
+	if ( negateSign ) {
+
+		for ( let i = 3; i < tangents.length; i += 4 ) {
+
+			tangents[ i ] *= - 1;
+
+		}
+
+	}
+
+	//
+
+	_geometry.setAttribute( 'tangent', new BufferAttribute( tangents, 4 ) );
+
+	if ( geometry !== _geometry ) {
+
+		geometry.copy( _geometry );
+
+	}
+
+	return geometry;
+
+}
+
+/**
+ * Merges a set of geometries into a single instance. All geometries must have compatible attributes.
+ *
+ * @param {Array<BufferGeometry>} geometries - The geometries to merge.
+ * @param {boolean} [useGroups=false] - Whether to use groups or not.
+ * @return {?BufferGeometry} The merged geometry. Returns `null` if the merge does not succeed.
+ */
+function mergeGeometries( geometries, useGroups = false ) {
+
+	const isIndexed = geometries[ 0 ].index !== null;
+
+	const attributesUsed = new Set( Object.keys( geometries[ 0 ].attributes ) );
+	const morphAttributesUsed = new Set( Object.keys( geometries[ 0 ].morphAttributes ) );
+
+	const attributes = {};
+	const morphAttributes = {};
+
+	const morphTargetsRelative = geometries[ 0 ].morphTargetsRelative;
+
+	const mergedGeometry = new BufferGeometry();
+
+	let offset = 0;
+
+	for ( let i = 0; i < geometries.length; ++ i ) {
+
+		const geometry = geometries[ i ];
+		let attributesCount = 0;
+
+		// ensure that all geometries are indexed, or none
+
+		if ( isIndexed !== ( geometry.index !== null ) ) {
+
+			console.error( 'THREE.BufferGeometryUtils: .mergeGeometries() failed with geometry at index ' + i + '. All geometries must have compatible attributes; make sure index attribute exists among all geometries, or in none of them.' );
+			return null;
+
+		}
+
+		// gather attributes, exit early if they're different
+
+		for ( const name in geometry.attributes ) {
+
+			if ( ! attributesUsed.has( name ) ) {
+
+				console.error( 'THREE.BufferGeometryUtils: .mergeGeometries() failed with geometry at index ' + i + '. All geometries must have compatible attributes; make sure "' + name + '" attribute exists among all geometries, or in none of them.' );
+				return null;
+
+			}
+
+			if ( attributes[ name ] === undefined ) attributes[ name ] = [];
+
+			attributes[ name ].push( geometry.attributes[ name ] );
+
+			attributesCount ++;
+
+		}
+
+		// ensure geometries have the same number of attributes
+
+		if ( attributesCount !== attributesUsed.size ) {
+
+			console.error( 'THREE.BufferGeometryUtils: .mergeGeometries() failed with geometry at index ' + i + '. Make sure all geometries have the same number of attributes.' );
+			return null;
+
+		}
+
+		// gather morph attributes, exit early if they're different
+
+		if ( morphTargetsRelative !== geometry.morphTargetsRelative ) {
+
+			console.error( 'THREE.BufferGeometryUtils: .mergeGeometries() failed with geometry at index ' + i + '. .morphTargetsRelative must be consistent throughout all geometries.' );
+			return null;
+
+		}
+
+		for ( const name in geometry.morphAttributes ) {
+
+			if ( ! morphAttributesUsed.has( name ) ) {
+
+				console.error( 'THREE.BufferGeometryUtils: .mergeGeometries() failed with geometry at index ' + i + '.  .morphAttributes must be consistent throughout all geometries.' );
+				return null;
+
+			}
+
+			if ( morphAttributes[ name ] === undefined ) morphAttributes[ name ] = [];
+
+			morphAttributes[ name ].push( geometry.morphAttributes[ name ] );
+
+		}
+
+		if ( useGroups ) {
+
+			let count;
+
+			if ( isIndexed ) {
+
+				count = geometry.index.count;
+
+			} else if ( geometry.attributes.position !== undefined ) {
+
+				count = geometry.attributes.position.count;
+
+			} else {
+
+				console.error( 'THREE.BufferGeometryUtils: .mergeGeometries() failed with geometry at index ' + i + '. The geometry must have either an index or a position attribute' );
+				return null;
+
+			}
+
+			mergedGeometry.addGroup( offset, count, i );
+
+			offset += count;
+
+		}
+
+	}
+
+	// merge indices
+
+	if ( isIndexed ) {
+
+		let indexOffset = 0;
+		const mergedIndex = [];
+
+		for ( let i = 0; i < geometries.length; ++ i ) {
+
+			const index = geometries[ i ].index;
+
+			for ( let j = 0; j < index.count; ++ j ) {
+
+				mergedIndex.push( index.getX( j ) + indexOffset );
+
+			}
+
+			indexOffset += geometries[ i ].attributes.position.count;
+
+		}
+
+		mergedGeometry.setIndex( mergedIndex );
+
+	}
+
+	// merge attributes
+
+	for ( const name in attributes ) {
+
+		const mergedAttribute = mergeAttributes( attributes[ name ] );
+
+		if ( ! mergedAttribute ) {
+
+			console.error( 'THREE.BufferGeometryUtils: .mergeGeometries() failed while trying to merge the ' + name + ' attribute.' );
+			return null;
+
+		}
+
+		mergedGeometry.setAttribute( name, mergedAttribute );
+
+	}
+
+	// merge morph attributes
+
+	for ( const name in morphAttributes ) {
+
+		const numMorphTargets = morphAttributes[ name ][ 0 ].length;
+
+		if ( numMorphTargets === 0 ) break;
+
+		mergedGeometry.morphAttributes = mergedGeometry.morphAttributes || {};
+		mergedGeometry.morphAttributes[ name ] = [];
+
+		for ( let i = 0; i < numMorphTargets; ++ i ) {
+
+			const morphAttributesToMerge = [];
+
+			for ( let j = 0; j < morphAttributes[ name ].length; ++ j ) {
+
+				morphAttributesToMerge.push( morphAttributes[ name ][ j ][ i ] );
+
+			}
+
+			const mergedMorphAttribute = mergeAttributes( morphAttributesToMerge );
+
+			if ( ! mergedMorphAttribute ) {
+
+				console.error( 'THREE.BufferGeometryUtils: .mergeGeometries() failed while trying to merge the ' + name + ' morphAttribute.' );
+				return null;
+
+			}
+
+			mergedGeometry.morphAttributes[ name ].push( mergedMorphAttribute );
+
+		}
+
+	}
+
+	return mergedGeometry;
+
+}
+
+/**
+ * Merges a set of attributes into a single instance. All attributes must have compatible properties and types.
+ * Instances of {@link InterleavedBufferAttribute} are not supported.
+ *
+ * @param {Array<BufferAttribute>} attributes - The attributes to merge.
+ * @return {?BufferAttribute} The merged attribute. Returns `null` if the merge does not succeed.
+ */
+function mergeAttributes( attributes ) {
+
+	let TypedArray;
+	let itemSize;
+	let normalized;
+	let gpuType = - 1;
+	let arrayLength = 0;
+
+	for ( let i = 0; i < attributes.length; ++ i ) {
+
+		const attribute = attributes[ i ];
+
+		if ( TypedArray === undefined ) TypedArray = attribute.array.constructor;
+		if ( TypedArray !== attribute.array.constructor ) {
+
+			console.error( 'THREE.BufferGeometryUtils: .mergeAttributes() failed. BufferAttribute.array must be of consistent array types across matching attributes.' );
+			return null;
+
+		}
+
+		if ( itemSize === undefined ) itemSize = attribute.itemSize;
+		if ( itemSize !== attribute.itemSize ) {
+
+			console.error( 'THREE.BufferGeometryUtils: .mergeAttributes() failed. BufferAttribute.itemSize must be consistent across matching attributes.' );
+			return null;
+
+		}
+
+		if ( normalized === undefined ) normalized = attribute.normalized;
+		if ( normalized !== attribute.normalized ) {
+
+			console.error( 'THREE.BufferGeometryUtils: .mergeAttributes() failed. BufferAttribute.normalized must be consistent across matching attributes.' );
+			return null;
+
+		}
+
+		if ( gpuType === - 1 ) gpuType = attribute.gpuType;
+		if ( gpuType !== attribute.gpuType ) {
+
+			console.error( 'THREE.BufferGeometryUtils: .mergeAttributes() failed. BufferAttribute.gpuType must be consistent across matching attributes.' );
+			return null;
+
+		}
+
+		arrayLength += attribute.count * itemSize;
+
+	}
+
+	const array = new TypedArray( arrayLength );
+	const result = new BufferAttribute( array, itemSize, normalized );
+	let offset = 0;
+
+	for ( let i = 0; i < attributes.length; ++ i ) {
+
+		const attribute = attributes[ i ];
+		if ( attribute.isInterleavedBufferAttribute ) {
+
+			const tupleOffset = offset / itemSize;
+			for ( let j = 0, l = attribute.count; j < l; j ++ ) {
+
+				for ( let c = 0; c < itemSize; c ++ ) {
+
+					const value = attribute.getComponent( j, c );
+					result.setComponent( j + tupleOffset, c, value );
+
+				}
+
+			}
+
+		} else {
+
+			array.set( attribute.array, offset );
+
+		}
+
+		offset += attribute.count * itemSize;
+
+	}
+
+	if ( gpuType !== undefined ) {
+
+		result.gpuType = gpuType;
+
+	}
+
+	return result;
+
+}
+
+/**
+ * Performs a deep clone of the given buffer attribute.
+ *
+ * @param {BufferAttribute} attribute - The attribute to clone.
+ * @return {BufferAttribute} The cloned attribute.
+ */
+function deepCloneAttribute( attribute ) {
+
+	if ( attribute.isInstancedInterleavedBufferAttribute || attribute.isInterleavedBufferAttribute ) {
+
+		return deinterleaveAttribute( attribute );
+
+	}
+
+	if ( attribute.isInstancedBufferAttribute ) {
+
+		return new InstancedBufferAttribute().copy( attribute );
+
+	}
+
+	return new BufferAttribute().copy( attribute );
+
+}
+
+/**
+ * Interleaves a set of attributes and returns a new array of corresponding attributes that share a
+ * single {@link InterleavedBuffer} instance. All attributes must have compatible types.
+ *
+ * @param {Array<BufferAttribute>} attributes - The attributes to interleave.
+ * @return {?Array<InterleavedBufferAttribute>} An array of interleaved attributes. If interleave does not succeed, the method returns `null`.
+ */
+function interleaveAttributes( attributes ) {
+
+	// Interleaves the provided attributes into an InterleavedBuffer and returns
+	// a set of InterleavedBufferAttributes for each attribute
+	let TypedArray;
+	let arrayLength = 0;
+	let stride = 0;
+
+	// calculate the length and type of the interleavedBuffer
+	for ( let i = 0, l = attributes.length; i < l; ++ i ) {
+
+		const attribute = attributes[ i ];
+
+		if ( TypedArray === undefined ) TypedArray = attribute.array.constructor;
+		if ( TypedArray !== attribute.array.constructor ) {
+
+			console.error( 'AttributeBuffers of different types cannot be interleaved' );
+			return null;
+
+		}
+
+		arrayLength += attribute.array.length;
+		stride += attribute.itemSize;
+
+	}
+
+	// Create the set of buffer attributes
+	const interleavedBuffer = new InterleavedBuffer( new TypedArray( arrayLength ), stride );
+	let offset = 0;
+	const res = [];
+	const getters = [ 'getX', 'getY', 'getZ', 'getW' ];
+	const setters = [ 'setX', 'setY', 'setZ', 'setW' ];
+
+	for ( let j = 0, l = attributes.length; j < l; j ++ ) {
+
+		const attribute = attributes[ j ];
+		const itemSize = attribute.itemSize;
+		const count = attribute.count;
+		const iba = new InterleavedBufferAttribute( interleavedBuffer, itemSize, offset, attribute.normalized );
+		res.push( iba );
+
+		offset += itemSize;
+
+		// Move the data for each attribute into the new interleavedBuffer
+		// at the appropriate offset
+		for ( let c = 0; c < count; c ++ ) {
+
+			for ( let k = 0; k < itemSize; k ++ ) {
+
+				iba[ setters[ k ] ]( c, attribute[ getters[ k ] ]( c ) );
+
+			}
+
+		}
+
+	}
+
+	return res;
+
+}
+
+/**
+ * Returns a new, non-interleaved version of the given attribute.
+ *
+ * @param {InterleavedBufferAttribute} attribute - The interleaved attribute.
+ * @return {BufferAttribute} The non-interleaved attribute.
+ */
+function deinterleaveAttribute( attribute ) {
+
+	const cons = attribute.data.array.constructor;
+	const count = attribute.count;
+	const itemSize = attribute.itemSize;
+	const normalized = attribute.normalized;
+
+	const array = new cons( count * itemSize );
+	let newAttribute;
+	if ( attribute.isInstancedInterleavedBufferAttribute ) {
+
+		newAttribute = new InstancedBufferAttribute( array, itemSize, normalized, attribute.meshPerAttribute );
+
+	} else {
+
+		newAttribute = new BufferAttribute( array, itemSize, normalized );
+
+	}
+
+	for ( let i = 0; i < count; i ++ ) {
+
+		newAttribute.setX( i, attribute.getX( i ) );
+
+		if ( itemSize >= 2 ) {
+
+			newAttribute.setY( i, attribute.getY( i ) );
+
+		}
+
+		if ( itemSize >= 3 ) {
+
+			newAttribute.setZ( i, attribute.getZ( i ) );
+
+		}
+
+		if ( itemSize >= 4 ) {
+
+			newAttribute.setW( i, attribute.getW( i ) );
+
+		}
+
+	}
+
+	return newAttribute;
+
+}
+
+/**
+ * Deinterleaves all attributes on the given geometry.
+ *
+ * @param {BufferGeometry} geometry - The geometry to deinterleave.
+ */
+function deinterleaveGeometry( geometry ) {
+
+	const attributes = geometry.attributes;
+	const morphTargets = geometry.morphTargets;
+	const attrMap = new Map();
+
+	for ( const key in attributes ) {
+
+		const attr = attributes[ key ];
+		if ( attr.isInterleavedBufferAttribute ) {
+
+			if ( ! attrMap.has( attr ) ) {
+
+				attrMap.set( attr, deinterleaveAttribute( attr ) );
+
+			}
+
+			attributes[ key ] = attrMap.get( attr );
+
+		}
+
+	}
+
+	for ( const key in morphTargets ) {
+
+		const attr = morphTargets[ key ];
+		if ( attr.isInterleavedBufferAttribute ) {
+
+			if ( ! attrMap.has( attr ) ) {
+
+				attrMap.set( attr, deinterleaveAttribute( attr ) );
+
+			}
+
+			morphTargets[ key ] = attrMap.get( attr );
+
+		}
+
+	}
+
+}
+
+/**
+ * Returns the amount of bytes used by all attributes to represent the geometry.
+ *
+ * @param {BufferGeometry} geometry - The geometry.
+ * @return {number} The estimate bytes used.
+ */
+function estimateBytesUsed( geometry ) {
+
+	// Return the estimated memory used by this geometry in bytes
+	// Calculate using itemSize, count, and BYTES_PER_ELEMENT to account
+	// for InterleavedBufferAttributes.
+	let mem = 0;
+	for ( const name in geometry.attributes ) {
+
+		const attr = geometry.getAttribute( name );
+		mem += attr.count * attr.itemSize * attr.array.BYTES_PER_ELEMENT;
+
+	}
+
+	const indices = geometry.getIndex();
+	mem += indices ? indices.count * indices.itemSize * indices.array.BYTES_PER_ELEMENT : 0;
+	return mem;
+
+}
+
+/**
+ * Returns a new geometry with vertices for which all similar vertex attributes (within tolerance) are merged.
+ *
+ * @param {BufferGeometry} geometry - The geometry to merge vertices for.
+ * @param {number} [tolerance=1e-4] - The tolerance value.
+ * @return {BufferGeometry} - The new geometry with merged vertices.
+ */
+function mergeVertices( geometry, tolerance = 1e-4 ) {
+
+	tolerance = Math.max( tolerance, Number.EPSILON );
+
+	// Generate an index buffer if the geometry doesn't have one, or optimize it
+	// if it's already available.
+	const hashToIndex = {};
+	const indices = geometry.getIndex();
+	const positions = geometry.getAttribute( 'position' );
+	const vertexCount = indices ? indices.count : positions.count;
+
+	// next value for triangle indices
+	let nextIndex = 0;
+
+	// attributes and new attribute arrays
+	const attributeNames = Object.keys( geometry.attributes );
+	const tmpAttributes = {};
+	const tmpMorphAttributes = {};
+	const newIndices = [];
+	const getters = [ 'getX', 'getY', 'getZ', 'getW' ];
+	const setters = [ 'setX', 'setY', 'setZ', 'setW' ];
+
+	// Initialize the arrays, allocating space conservatively. Extra
+	// space will be trimmed in the last step.
+	for ( let i = 0, l = attributeNames.length; i < l; i ++ ) {
+
+		const name = attributeNames[ i ];
+		const attr = geometry.attributes[ name ];
+
+		tmpAttributes[ name ] = new attr.constructor(
+			new attr.array.constructor( attr.count * attr.itemSize ),
+			attr.itemSize,
+			attr.normalized
+		);
+
+		const morphAttributes = geometry.morphAttributes[ name ];
+		if ( morphAttributes ) {
+
+			if ( ! tmpMorphAttributes[ name ] ) tmpMorphAttributes[ name ] = [];
+			morphAttributes.forEach( ( morphAttr, i ) => {
+
+				const array = new morphAttr.array.constructor( morphAttr.count * morphAttr.itemSize );
+				tmpMorphAttributes[ name ][ i ] = new morphAttr.constructor( array, morphAttr.itemSize, morphAttr.normalized );
+
+			} );
+
+		}
+
+	}
+
+	// convert the error tolerance to an amount of decimal places to truncate to
+	const halfTolerance = tolerance * 0.5;
+	const exponent = Math.log10( 1 / tolerance );
+	const hashMultiplier = Math.pow( 10, exponent );
+	const hashAdditive = halfTolerance * hashMultiplier;
+	for ( let i = 0; i < vertexCount; i ++ ) {
+
+		const index = indices ? indices.getX( i ) : i;
+
+		// Generate a hash for the vertex attributes at the current index 'i'
+		let hash = '';
+		for ( let j = 0, l = attributeNames.length; j < l; j ++ ) {
+
+			const name = attributeNames[ j ];
+			const attribute = geometry.getAttribute( name );
+			const itemSize = attribute.itemSize;
+
+			for ( let k = 0; k < itemSize; k ++ ) {
+
+				// double tilde truncates the decimal value
+				hash += `${ ~ ~ ( attribute[ getters[ k ] ]( index ) * hashMultiplier + hashAdditive ) },`;
+
+			}
+
+		}
+
+		// Add another reference to the vertex if it's already
+		// used by another index
+		if ( hash in hashToIndex ) {
+
+			newIndices.push( hashToIndex[ hash ] );
+
+		} else {
+
+			// copy data to the new index in the temporary attributes
+			for ( let j = 0, l = attributeNames.length; j < l; j ++ ) {
+
+				const name = attributeNames[ j ];
+				const attribute = geometry.getAttribute( name );
+				const morphAttributes = geometry.morphAttributes[ name ];
+				const itemSize = attribute.itemSize;
+				const newArray = tmpAttributes[ name ];
+				const newMorphArrays = tmpMorphAttributes[ name ];
+
+				for ( let k = 0; k < itemSize; k ++ ) {
+
+					const getterFunc = getters[ k ];
+					const setterFunc = setters[ k ];
+					newArray[ setterFunc ]( nextIndex, attribute[ getterFunc ]( index ) );
+
+					if ( morphAttributes ) {
+
+						for ( let m = 0, ml = morphAttributes.length; m < ml; m ++ ) {
+
+							newMorphArrays[ m ][ setterFunc ]( nextIndex, morphAttributes[ m ][ getterFunc ]( index ) );
+
+						}
+
+					}
+
+				}
+
+			}
+
+			hashToIndex[ hash ] = nextIndex;
+			newIndices.push( nextIndex );
+			nextIndex ++;
+
+		}
+
+	}
+
+	// generate result BufferGeometry
+	const result = geometry.clone();
+	for ( const name in geometry.attributes ) {
+
+		const tmpAttribute = tmpAttributes[ name ];
+
+		result.setAttribute( name, new tmpAttribute.constructor(
+			tmpAttribute.array.slice( 0, nextIndex * tmpAttribute.itemSize ),
+			tmpAttribute.itemSize,
+			tmpAttribute.normalized,
+		) );
+
+		if ( ! ( name in tmpMorphAttributes ) ) continue;
+
+		for ( let j = 0; j < tmpMorphAttributes[ name ].length; j ++ ) {
+
+			const tmpMorphAttribute = tmpMorphAttributes[ name ][ j ];
+
+			result.morphAttributes[ name ][ j ] = new tmpMorphAttribute.constructor(
+				tmpMorphAttribute.array.slice( 0, nextIndex * tmpMorphAttribute.itemSize ),
+				tmpMorphAttribute.itemSize,
+				tmpMorphAttribute.normalized,
+			);
+
+		}
+
+	}
+
+	// indices
+
+	result.setIndex( newIndices );
+
+	return result;
+
+}
+
+/**
+ * Returns a new indexed geometry based on `TrianglesDrawMode` draw mode.
+ * This mode corresponds to the `gl.TRIANGLES` primitive in WebGL.
+ *
+ * @param {BufferGeometry} geometry - The geometry to convert.
+ * @param {number} drawMode - The current draw mode.
+ * @return {BufferGeometry} The new geometry using `TrianglesDrawMode`.
+ */
+function toTrianglesDrawMode( geometry, drawMode ) {
+
+	if ( drawMode === TrianglesDrawMode ) {
+
+		console.warn( 'THREE.BufferGeometryUtils.toTrianglesDrawMode(): Geometry already defined as triangles.' );
+		return geometry;
+
+	}
+
+	if ( drawMode === TriangleFanDrawMode || drawMode === TriangleStripDrawMode ) {
+
+		let index = geometry.getIndex();
+
+		// generate index if not present
+
+		if ( index === null ) {
+
+			const indices = [];
+
+			const position = geometry.getAttribute( 'position' );
+
+			if ( position !== undefined ) {
+
+				for ( let i = 0; i < position.count; i ++ ) {
+
+					indices.push( i );
+
+				}
+
+				geometry.setIndex( indices );
+				index = geometry.getIndex();
+
+			} else {
+
+				console.error( 'THREE.BufferGeometryUtils.toTrianglesDrawMode(): Undefined position attribute. Processing not possible.' );
+				return geometry;
+
+			}
+
+		}
+
+		//
+
+		const numberOfTriangles = index.count - 2;
+		const newIndices = [];
+
+		if ( drawMode === TriangleFanDrawMode ) {
+
+			// gl.TRIANGLE_FAN
+
+			for ( let i = 1; i <= numberOfTriangles; i ++ ) {
+
+				newIndices.push( index.getX( 0 ) );
+				newIndices.push( index.getX( i ) );
+				newIndices.push( index.getX( i + 1 ) );
+
+			}
+
+		} else {
+
+			// gl.TRIANGLE_STRIP
+
+			for ( let i = 0; i < numberOfTriangles; i ++ ) {
+
+				if ( i % 2 === 0 ) {
+
+					newIndices.push( index.getX( i ) );
+					newIndices.push( index.getX( i + 1 ) );
+					newIndices.push( index.getX( i + 2 ) );
+
+				} else {
+
+					newIndices.push( index.getX( i + 2 ) );
+					newIndices.push( index.getX( i + 1 ) );
+					newIndices.push( index.getX( i ) );
+
+				}
+
+			}
+
+		}
+
+		if ( ( newIndices.length / 3 ) !== numberOfTriangles ) {
+
+			console.error( 'THREE.BufferGeometryUtils.toTrianglesDrawMode(): Unable to generate correct amount of triangles.' );
+
+		}
+
+		// build final geometry
+
+		const newGeometry = geometry.clone();
+		newGeometry.setIndex( newIndices );
+		newGeometry.clearGroups();
+
+		return newGeometry;
+
+	} else {
+
+		console.error( 'THREE.BufferGeometryUtils.toTrianglesDrawMode(): Unknown draw mode:', drawMode );
+		return geometry;
+
+	}
+
+}
+
+/**
+ * Calculates the morphed attributes of a morphed/skinned BufferGeometry.
+ *
+ * Helpful for Raytracing or Decals (i.e. a `DecalGeometry` applied to a morphed Object with a `BufferGeometry`
+ * will use the original `BufferGeometry`, not the morphed/skinned one, generating an incorrect result.
+ * Using this function to create a shadow `Object3`D the `DecalGeometry` can be correctly generated).
+ *
+ * @param {Mesh|Line|Points} object - The 3D object to compute morph attributes for.
+ * @return {Object} An object with original position/normal attributes and morphed ones.
+ */
+function computeMorphedAttributes( object ) {
+
+	const _vA = new Vector3();
+	const _vB = new Vector3();
+	const _vC = new Vector3();
+
+	const _tempA = new Vector3();
+	const _tempB = new Vector3();
+	const _tempC = new Vector3();
+
+	const _morphA = new Vector3();
+	const _morphB = new Vector3();
+	const _morphC = new Vector3();
+
+	function _calculateMorphedAttributeData(
+		object,
+		attribute,
+		morphAttribute,
+		morphTargetsRelative,
+		a,
+		b,
+		c,
+		modifiedAttributeArray
+	) {
+
+		_vA.fromBufferAttribute( attribute, a );
+		_vB.fromBufferAttribute( attribute, b );
+		_vC.fromBufferAttribute( attribute, c );
+
+		const morphInfluences = object.morphTargetInfluences;
+
+		if ( morphAttribute && morphInfluences ) {
+
+			_morphA.set( 0, 0, 0 );
+			_morphB.set( 0, 0, 0 );
+			_morphC.set( 0, 0, 0 );
+
+			for ( let i = 0, il = morphAttribute.length; i < il; i ++ ) {
+
+				const influence = morphInfluences[ i ];
+				const morph = morphAttribute[ i ];
+
+				if ( influence === 0 ) continue;
+
+				_tempA.fromBufferAttribute( morph, a );
+				_tempB.fromBufferAttribute( morph, b );
+				_tempC.fromBufferAttribute( morph, c );
+
+				if ( morphTargetsRelative ) {
+
+					_morphA.addScaledVector( _tempA, influence );
+					_morphB.addScaledVector( _tempB, influence );
+					_morphC.addScaledVector( _tempC, influence );
+
+				} else {
+
+					_morphA.addScaledVector( _tempA.sub( _vA ), influence );
+					_morphB.addScaledVector( _tempB.sub( _vB ), influence );
+					_morphC.addScaledVector( _tempC.sub( _vC ), influence );
+
+				}
+
+			}
+
+			_vA.add( _morphA );
+			_vB.add( _morphB );
+			_vC.add( _morphC );
+
+		}
+
+		if ( object.isSkinnedMesh ) {
+
+			object.applyBoneTransform( a, _vA );
+			object.applyBoneTransform( b, _vB );
+			object.applyBoneTransform( c, _vC );
+
+		}
+
+		modifiedAttributeArray[ a * 3 + 0 ] = _vA.x;
+		modifiedAttributeArray[ a * 3 + 1 ] = _vA.y;
+		modifiedAttributeArray[ a * 3 + 2 ] = _vA.z;
+		modifiedAttributeArray[ b * 3 + 0 ] = _vB.x;
+		modifiedAttributeArray[ b * 3 + 1 ] = _vB.y;
+		modifiedAttributeArray[ b * 3 + 2 ] = _vB.z;
+		modifiedAttributeArray[ c * 3 + 0 ] = _vC.x;
+		modifiedAttributeArray[ c * 3 + 1 ] = _vC.y;
+		modifiedAttributeArray[ c * 3 + 2 ] = _vC.z;
+
+	}
+
+	const geometry = object.geometry;
+	const material = object.material;
+
+	let a, b, c;
+	const index = geometry.index;
+	const positionAttribute = geometry.attributes.position;
+	const morphPosition = geometry.morphAttributes.position;
+	const morphTargetsRelative = geometry.morphTargetsRelative;
+	const normalAttribute = geometry.attributes.normal;
+	const morphNormal = geometry.morphAttributes.position;
+
+	const groups = geometry.groups;
+	const drawRange = geometry.drawRange;
+	let i, j, il, jl;
+	let group;
+	let start, end;
+
+	const modifiedPosition = new Float32Array( positionAttribute.count * positionAttribute.itemSize );
+	const modifiedNormal = new Float32Array( normalAttribute.count * normalAttribute.itemSize );
+
+	if ( index !== null ) {
+
+		// indexed buffer geometry
+
+		if ( Array.isArray( material ) ) {
+
+			for ( i = 0, il = groups.length; i < il; i ++ ) {
+
+				group = groups[ i ];
+
+				start = Math.max( group.start, drawRange.start );
+				end = Math.min( ( group.start + group.count ), ( drawRange.start + drawRange.count ) );
+
+				for ( j = start, jl = end; j < jl; j += 3 ) {
+
+					a = index.getX( j );
+					b = index.getX( j + 1 );
+					c = index.getX( j + 2 );
+
+					_calculateMorphedAttributeData(
+						object,
+						positionAttribute,
+						morphPosition,
+						morphTargetsRelative,
+						a, b, c,
+						modifiedPosition
+					);
+
+					_calculateMorphedAttributeData(
+						object,
+						normalAttribute,
+						morphNormal,
+						morphTargetsRelative,
+						a, b, c,
+						modifiedNormal
+					);
+
+				}
+
+			}
+
+		} else {
+
+			start = Math.max( 0, drawRange.start );
+			end = Math.min( index.count, ( drawRange.start + drawRange.count ) );
+
+			for ( i = start, il = end; i < il; i += 3 ) {
+
+				a = index.getX( i );
+				b = index.getX( i + 1 );
+				c = index.getX( i + 2 );
+
+				_calculateMorphedAttributeData(
+					object,
+					positionAttribute,
+					morphPosition,
+					morphTargetsRelative,
+					a, b, c,
+					modifiedPosition
+				);
+
+				_calculateMorphedAttributeData(
+					object,
+					normalAttribute,
+					morphNormal,
+					morphTargetsRelative,
+					a, b, c,
+					modifiedNormal
+				);
+
+			}
+
+		}
+
+	} else {
+
+		// non-indexed buffer geometry
+
+		if ( Array.isArray( material ) ) {
+
+			for ( i = 0, il = groups.length; i < il; i ++ ) {
+
+				group = groups[ i ];
+
+				start = Math.max( group.start, drawRange.start );
+				end = Math.min( ( group.start + group.count ), ( drawRange.start + drawRange.count ) );
+
+				for ( j = start, jl = end; j < jl; j += 3 ) {
+
+					a = j;
+					b = j + 1;
+					c = j + 2;
+
+					_calculateMorphedAttributeData(
+						object,
+						positionAttribute,
+						morphPosition,
+						morphTargetsRelative,
+						a, b, c,
+						modifiedPosition
+					);
+
+					_calculateMorphedAttributeData(
+						object,
+						normalAttribute,
+						morphNormal,
+						morphTargetsRelative,
+						a, b, c,
+						modifiedNormal
+					);
+
+				}
+
+			}
+
+		} else {
+
+			start = Math.max( 0, drawRange.start );
+			end = Math.min( positionAttribute.count, ( drawRange.start + drawRange.count ) );
+
+			for ( i = start, il = end; i < il; i += 3 ) {
+
+				a = i;
+				b = i + 1;
+				c = i + 2;
+
+				_calculateMorphedAttributeData(
+					object,
+					positionAttribute,
+					morphPosition,
+					morphTargetsRelative,
+					a, b, c,
+					modifiedPosition
+				);
+
+				_calculateMorphedAttributeData(
+					object,
+					normalAttribute,
+					morphNormal,
+					morphTargetsRelative,
+					a, b, c,
+					modifiedNormal
+				);
+
+			}
+
+		}
+
+	}
+
+	const morphedPositionAttribute = new Float32BufferAttribute( modifiedPosition, 3 );
+	const morphedNormalAttribute = new Float32BufferAttribute( modifiedNormal, 3 );
+
+	return {
+
+		positionAttribute: positionAttribute,
+		normalAttribute: normalAttribute,
+		morphedPositionAttribute: morphedPositionAttribute,
+		morphedNormalAttribute: morphedNormalAttribute
+
+	};
+
+}
+
+/**
+ * Merges the {@link BufferGeometry#groups} for the given geometry.
+ *
+ * @param {BufferGeometry} geometry - The geometry to modify.
+ * @return {BufferGeometry} - The updated geometry
+ */
+function mergeGroups( geometry ) {
+
+	if ( geometry.groups.length === 0 ) {
+
+		console.warn( 'THREE.BufferGeometryUtils.mergeGroups(): No groups are defined. Nothing to merge.' );
+		return geometry;
+
+	}
+
+	let groups = geometry.groups;
+
+	// sort groups by material index
+
+	groups = groups.sort( ( a, b ) => {
+
+		if ( a.materialIndex !== b.materialIndex ) return a.materialIndex - b.materialIndex;
+
+		return a.start - b.start;
+
+	} );
+
+	// create index for non-indexed geometries
+
+	if ( geometry.getIndex() === null ) {
+
+		const positionAttribute = geometry.getAttribute( 'position' );
+		const indices = [];
+
+		for ( let i = 0; i < positionAttribute.count; i += 3 ) {
+
+			indices.push( i, i + 1, i + 2 );
+
+		}
+
+		geometry.setIndex( indices );
+
+	}
+
+	// sort index
+
+	const index = geometry.getIndex();
+
+	const newIndices = [];
+
+	for ( let i = 0; i < groups.length; i ++ ) {
+
+		const group = groups[ i ];
+
+		const groupStart = group.start;
+		const groupLength = groupStart + group.count;
+
+		for ( let j = groupStart; j < groupLength; j ++ ) {
+
+			newIndices.push( index.getX( j ) );
+
+		}
+
+	}
+
+	geometry.dispose(); // Required to force buffer recreation
+	geometry.setIndex( newIndices );
+
+	// update groups indices
+
+	let start = 0;
+
+	for ( let i = 0; i < groups.length; i ++ ) {
+
+		const group = groups[ i ];
+
+		group.start = start;
+		start += group.count;
+
+	}
+
+	// merge groups
+
+	let currentGroup = groups[ 0 ];
+
+	geometry.groups = [ currentGroup ];
+
+	for ( let i = 1; i < groups.length; i ++ ) {
+
+		const group = groups[ i ];
+
+		if ( currentGroup.materialIndex === group.materialIndex ) {
+
+			currentGroup.count += group.count;
+
+		} else {
+
+			currentGroup = group;
+			geometry.groups.push( currentGroup );
+
+		}
+
+	}
+
+	return geometry;
+
+}
+
+/**
+ * Modifies the supplied geometry if it is non-indexed, otherwise creates a new,
+ * non-indexed geometry. Returns the geometry with smooth normals everywhere except
+ * faces that meet at an angle greater than the crease angle.
+ *
+ * @param {BufferGeometry} geometry - The geometry to modify.
+ * @param {number} [creaseAngle=Math.PI/3] - The crease angle in radians.
+ * @return {BufferGeometry} - The updated geometry
+ */
+function toCreasedNormals( geometry, creaseAngle = Math.PI / 3 /* 60 degrees */ ) {
+
+	const creaseDot = Math.cos( creaseAngle );
+	const hashMultiplier = ( 1 + 1e-10 ) * 1e2;
+
+	// reusable vectors
+	const verts = [ new Vector3(), new Vector3(), new Vector3() ];
+	const tempVec1 = new Vector3();
+	const tempVec2 = new Vector3();
+	const tempNorm = new Vector3();
+	const tempNorm2 = new Vector3();
+
+	// hashes a vector
+	function hashVertex( v ) {
+
+		const x = ~ ~ ( v.x * hashMultiplier );
+		const y = ~ ~ ( v.y * hashMultiplier );
+		const z = ~ ~ ( v.z * hashMultiplier );
+		return `${x},${y},${z}`;
+
+	}
+
+	// BufferGeometry.toNonIndexed() warns if the geometry is non-indexed
+	// and returns the original geometry
+	const resultGeometry = geometry.index ? geometry.toNonIndexed() : geometry;
+	const posAttr = resultGeometry.attributes.position;
+	const vertexMap = {};
+
+	// find all the normals shared by commonly located vertices
+	for ( let i = 0, l = posAttr.count / 3; i < l; i ++ ) {
+
+		const i3 = 3 * i;
+		const a = verts[ 0 ].fromBufferAttribute( posAttr, i3 + 0 );
+		const b = verts[ 1 ].fromBufferAttribute( posAttr, i3 + 1 );
+		const c = verts[ 2 ].fromBufferAttribute( posAttr, i3 + 2 );
+
+		tempVec1.subVectors( c, b );
+		tempVec2.subVectors( a, b );
+
+		// add the normal to the map for all vertices
+		const normal = new Vector3().crossVectors( tempVec1, tempVec2 ).normalize();
+		for ( let n = 0; n < 3; n ++ ) {
+
+			const vert = verts[ n ];
+			const hash = hashVertex( vert );
+			if ( ! ( hash in vertexMap ) ) {
+
+				vertexMap[ hash ] = [];
+
+			}
+
+			vertexMap[ hash ].push( normal );
+
+		}
+
+	}
+
+	// average normals from all vertices that share a common location if they are within the
+	// provided crease threshold
+	const normalArray = new Float32Array( posAttr.count * 3 );
+	const normAttr = new BufferAttribute( normalArray, 3, false );
+	for ( let i = 0, l = posAttr.count / 3; i < l; i ++ ) {
+
+		// get the face normal for this vertex
+		const i3 = 3 * i;
+		const a = verts[ 0 ].fromBufferAttribute( posAttr, i3 + 0 );
+		const b = verts[ 1 ].fromBufferAttribute( posAttr, i3 + 1 );
+		const c = verts[ 2 ].fromBufferAttribute( posAttr, i3 + 2 );
+
+		tempVec1.subVectors( c, b );
+		tempVec2.subVectors( a, b );
+
+		tempNorm.crossVectors( tempVec1, tempVec2 ).normalize();
+
+		// average all normals that meet the threshold and set the normal value
+		for ( let n = 0; n < 3; n ++ ) {
+
+			const vert = verts[ n ];
+			const hash = hashVertex( vert );
+			const otherNormals = vertexMap[ hash ];
+			tempNorm2.set( 0, 0, 0 );
+
+			for ( let k = 0, lk = otherNormals.length; k < lk; k ++ ) {
+
+				const otherNorm = otherNormals[ k ];
+				if ( tempNorm.dot( otherNorm ) > creaseDot ) {
+
+					tempNorm2.add( otherNorm );
+
+				}
+
+			}
+
+			tempNorm2.normalize();
+			normAttr.setXYZ( i3 + n, tempNorm2.x, tempNorm2.y, tempNorm2.z );
+
+		}
+
+	}
+
+	resultGeometry.setAttribute( 'normal', normAttr );
+	return resultGeometry;
+
+}
+
+export {
+	computeMikkTSpaceTangents,
+	mergeGeometries,
+	mergeAttributes,
+	deepCloneAttribute,
+	deinterleaveAttribute,
+	deinterleaveGeometry,
+	interleaveAttributes,
+	estimateBytesUsed,
+	mergeVertices,
+	toTrianglesDrawMode,
+	computeMorphedAttributes,
+	mergeGroups,
+	toCreasedNormals
+};

zh.html

@@ -0,0 +1,34 @@
+<!doctype html><html lang="zh-CN"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"><title>AWSM：智能体世界仿真与建图 — Phygital AI</title><meta name="description" content="把真实空间，变成虚实融合智能体可以使用的世界。"><meta property="og:type" content="article"><meta property="og:locale" content="zh_CN"><meta property="og:title" content="AWSM：智能体世界仿真与建图"><meta property="og:description" content="把真实空间，变成虚实融合智能体可以使用的世界。"><meta property="og:image" content="https://phygital-ai.github.io/agentic-world-simulation-and-mapping/assets/teaser_originals.png"><meta property="og:url" content="https://phygital-ai.github.io/agentic-world-simulation-and-mapping/zh.html"><meta name="twitter:card" content="summary_large_image"><meta name="twitter:title" content="AWSM：智能体世界仿真与建图"><meta name="twitter:description" content="把真实空间，变成虚实融合智能体可以使用的世界。"><meta name="twitter:image" content="https://phygital-ai.github.io/agentic-world-simulation-and-mapping/assets/teaser_originals.png"><link rel="canonical" href="https://phygital-ai.github.io/agentic-world-simulation-and-mapping/zh.html"><link rel="alternate" hreflang="zh-CN" href="https://phygital-ai.github.io/agentic-world-simulation-and-mapping/zh.html"><link rel="alternate" hreflang="en" href="https://phygital-ai.github.io/agentic-world-simulation-and-mapping/"><link rel="stylesheet" href="style.css?v=cbba47743ba2"><link rel="stylesheet" href="editorial.css?v=462f0ffc2b51"><script type="importmap">{"imports":{"three":"./vendor/three/three.module.js"}}</script><script src="app.js?v=d70db3a6816b" defer></script><script type="module" src="scene-compare.js?v=bd749465e05d"></script><link rel="stylesheet" href="office.css?v=2e8c4a2f0718"><link rel="alternate" hreflang="x-default" href="https://phygital-ai.github.io/agentic-world-simulation-and-mapping/"><meta property="article:published_time" content="2026-10-01"></head>
+<body data-lang="zh"><a class="skip" href="#main">跳到正文</a><nav><a class="brand" href="index.html">PHYGITAL AI / AWSM</a><div><a href="zh.html" aria-current='page'>中文</a><a href="index.html" >EN</a><a href="https://github.com/wentingw/AWSM" target="_blank" rel="noopener noreferrer">GitHub ↗</a></div></nav><header><p class="title-name">智能体世界仿真与建图</p><div class="title-lockup"><h1 aria-label="AWSM: 智能体世界仿真与建图">AWSM</h1><p class="pronunciation">读作 “awesome”</p></div><p class="lead">把真实空间，变成<strong class="tagline-emphasis">虚实融合智能体可以使用的世界。</strong></p><p class="brand-note">Phygital = physical（物理）+ digital（数字），即虚实融合。</p><p class="publication-date">发布于 <time datetime="2026-10-01">2026年10月1日</time></p></header>
+<div class="hero" id="overview-video"><video controls playsinline preload="metadata" poster="assets/awsm-promo-v5-poster.jpg" aria-label="AWSM 介绍视频"><source src="assets/awsm-promo-v5.mp4" type="video/mp4"><a href="assets/awsm-promo-v5.mp4">AWSM 介绍视频</a></video><div class="project-links" role="group" aria-label="AWSM 项目资源"><a class="project-resource" href="https://github.com/wentingw/AWSM" target="_blank" rel="noopener noreferrer" aria-label="GitHub（在新窗口打开）" title="GitHub（在新窗口打开）"><img src="assets/icons/github.svg" width="24" height="24" alt="">GitHub<svg class="project-external-arrow" viewBox="0 0 16 16" width="12" height="12" fill="none" stroke="currentColor" stroke-width="1.6" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><path d="M4 12 12 4M4 4h8v8"/></svg></a><span class="project-resource"><img src="assets/icons/huggingface.svg" width="24" height="24" alt="">Hugging Face (Coming soon)</span></div></div><div class="layout"><aside class="toc"><p>ON THIS PAGE</p><a href="#results">概览</a><a href="#motivation">研究动机</a><a href="#workflow">工作流</a><a href="#more-results">结果与分析</a><a href="#embodied-demo">具身演示</a><a href="#related-work">相关工作</a><a href="#limitations">局限性</a><a href="#citation">引用</a><a href="#references">参考文献</a></aside><main id="main">
+<section id="results"><p class="section-tag">01 / OVERVIEW</p><aside class="tldr"><strong>TL;DR</strong><p>AWSM 让大模型智能体将视觉观测转化为可编辑的 3D 场景，并以估计的相机位姿与深度进行几何锚定，其中包括融合 IMU 的视觉惯性约束，将几何与真实物理尺度绑定。在一个受控的 Isaac Sim 场景中，四条重建路线检验的不只是视觉合理性，更是空间忠实度：经配准的纯 RGB 路线 M1 与真值位姿条件路线 M4 对比，双向表面误差降低约 81%，180 个评估视角上的模型深度绝对相对误差（AbsRel）降低约 54%。多机器人 demo 进一步展示基于重建地图的目标导航与任务执行：沿预设路线前往地���上已标记的前台，并在那里排队。下一步是提升效率与精度、形成可复用的仿真就绪资产，并进一步与虚实融合智能体（phygital agents）集成，支持持久空间记忆、寻物与交互。</p></aside><h2>一个场景可以看起来合理，却在空间上是错的。</h2><p class="standfirst">Agentic 重建由大模型检查观测、调用建模工具，并迭代构建和验证可编辑场景。挑战不只是生成一个令人信服的房间，还在于保留尺度、形状与空间关系，让结果在渲染之外仍然有用。AWSM 研究如何用几何证据约束这一过程，让输出成为下游智能体更忠实的空间参照。</p><p>我们对 NVIDIA Isaac Sim 中的 World Lobby 仿真场景，沿四条完整工程路径进行重建：从纯 RGB 建模，逐步加入估计位姿、深度与真实相机位姿。每条路径交付的是冻结的 Blender 场景，而不只是点云或新视角渲染器。实验关注的不是谁生成了最漂亮的单帧，而是几何证据究竟带来了什么、没有解决什么，以及哪些误差会穿过对象化建模继续存在。</p><div class="insight-strip"><div><strong>0.376 → 0.072 m</strong><span>双向表面误差</span><small class="metric-reduction">相对降低约 81% · M1 → M4</small></div><div><strong>16.44% → 7.60%</strong><span>模型深度绝对相对误差（AbsRel）· 180 个评估视角</span><small class="metric-reduction">相对降低约 54% · M1 → M4</small></div></div><p class="metric-caveat">单场景 M1 → M4 诊断性对比：M1 使用真值辅助的 Sim(3) 配准，M4 使用真值相机位姿。相对降幅比较的是完整管线，而非 IMU 的独立贡献。</p></section>
+<section id="motivation"><p class="section-tag">02 / MOTIVATION</p><h2>从看起来合理的重建，到可以使用的空间参照</h2><p>负责重建的智能体可以搭建一个看似合理的房间，却重建错转角、距离或通道。对具身智能体而言，这些不只是视觉缺陷，还会改变目标的位置与身体可以通过的路线。几何锚定，是在信息可用时以位姿、深度与米制证据约束建模过程，而非仅依赖视觉合理性。关键问题是：这些约束如何在观测转化为可编辑对象的过程中得到保留？</p><p>前台任务让这种联系变得具体：在重建地图中标记目标、设定路线，再由机器人控制器执行。本文用“场景表示”指可编辑的对象与几何，用“地图”描述它在导航中的作用，用“仿真资产”描述它与仿真器的集成。AWSM 中的“世界”指这种可复用的空间环境，而非学习得到的动力学预测模型。连接导航、记忆、交互与仿真的持久空间基础是长期目标；当前实验和 demo 分别研究其几何基础，并展示一种下游用途。</p><aside class="claim"><strong>研究问题</strong><p>逐步增强的几何证据，如何改变一个对象化可编辑世界的忠实度，以及它仍然保留的失败模式？</p></aside><h3>本文贡献</h3><ol class="contributions"><li>对几何锚定的 agentic 重建开展四路线研究：固定场景、同一组 180 帧 RGB 输入、建模目标与可编辑输出格式，比较具有不同几何证据的完整管线。</li><li>将轨迹误差、原生深度、最终场景深度、表面几何和渲染外观分开评测，并提供冻结 Blend/GLB 资产、可复现评测子集、公开哈希与交互式对照。</li><li>通过基于地图的多机器人仿真示例，将重建场景连接到目标标记、预设路线和控制器执行；它展示下游用途，而非不同重建路线之间的受控导航对比。</li></ol><div id="scale-and-sensors" class="analysis-block grounding-block"><h3>几何锚定：真实尺度是必要约束，但不是全部答案</h3>
+<p>纯单目投影几何存在全局尺度歧义；学习型米制深度提供有用先验，却不是对每个真实尺寸的独立测量。标定 RGB-D、已知基线或长度，以及视觉惯性估计，都可以提供米制约束。IMU 融合能在运动激励充分、初始化、偏置估计、时间同步与标定可靠时恢复米制运动；IMU 不是直接测距传感器，“带 IMU”不等于“尺度一定准确”。</p>
+<p>尺度、局部形状、相机位姿和可通行连通性需要分别检查。弯曲转角被简化成直角，不能只靠缩放修复。位姿条件可以改善深度重建，但 DA3 消融和本文 M2/M3 的结果都不支持“所有深度指标必然同步改善”；下文四条路线是完整系统比较，而非只切换 IMU 的单变量实验。</p>
+<p class="inline-sources"><a href="https://arxiv.org/abs/1708.03852" target="_blank" rel="noopener">VINS-Mono</a> · <a href="https://arxiv.org/html/2511.10647v1#S7.SS2.SSS3" target="_blank" rel="noopener">DA3 位姿条件消融</a></p></div></section>
+<section id="workflow"><p class="section-tag">03 / WORKFLOW</p><h2>一个 agentic 工作流，四条重建路线</h2><p>负责重建的智能体遵循共同的“观察—构建—验证”���环：检查可用证据，编写 Blender Python 构建对象，渲染检查视图，并在冻结评测前迭代修改。这里的 agentic 指这种工具驱动的建模过程；下游机器人控制器承担另一种角色。</p><p>四条路径使用相同的 180 张 RGB 身份和相同的 Astra–Blender 建模目标。M1 检验没有米制测量时视觉先验能走多远；M2 加入 RGB-only ViPE 位姿与 pose-conditioned DA3 深度；M3 使用 ORB-SLAM3 单目惯性位姿与同类 DA3 观测；M4 提供 GT 相机位姿，用于诊断剩余建模误差，但建模时不读取 GT 网格或 GT 深度。</p><figure class="table-figure" id="table-1" data-table-key="table1"><figcaption><span>Table 1.</span> <span class="table-title"></span></figcaption><div class="table-scroll table-mount" role="region" tabindex="0" aria-label="Table 1"></div></figure><h3>Agentic 重建闭环</h3><div class="workflow-strip"><div><strong>01 · 观察</strong><span>检查全部 180 张 RGB；在允许的方法中读取位姿条件下的 DA3 几何。</span></div><div><strong>02 · 建模</strong><span>编写 Blender Python，将房间拆解为命名对象、材质、相机与碰撞代理。</span></div><div><strong>03 · 验证</strong><span>在固定修订相机上比较成对 RGB 与 optical-Z 渲染，并在固定版本预算内修改。</span></div><div><strong>04 · 冻结</strong><span>在作者看到任何 GT 评分前，导出 Blend/GLB、manifest、语义 ID 与哈希。</span></div><div><strong>05 · 评测</strong><span>在同一冻结协议下分别测量位姿、原生深度、模型深度、表面距离与外观。</span></div></div><p>所有场景冻结后才引入真值。M2/M3 只使用一次全局 SE(3) 配准，尺度固定为 1；不进行 mesh ICP 或逐视角拟合。M1 没有原生米制轨迹，其模型空间指标使用由 5 个手工相机关联得到、冻结不变的 GT-assisted Sim(3)，因此只能解释形状与显示误差，不能视为米制恢复。</p><aside class="notice"><strong>M1 diagnostic limitation.</strong> 其 Sim(3) 配准后的几何与深度数值描述 GT-assisted 显示配准后的形状，不代表原生米制恢复。</aside></section>
+<section id="more-results"><p class="section-tag">04 / RESULTS AND ANALYSIS</p><h2>比排行榜更重要的三个发现</h2><article class="finding"><p class="finding-kicker">FINDING 01</p><h3>1. 更好的位姿不会机械地带来更好的原生深度</h3><p>M3 的轨迹优于 M2：ATE 从 0.167 m 降至 0.121 m，旋转误差从 0.31° 降至 0.19°；但 M3 的原生 DA3 AbsRel 反而略差（20.38% 对 19.06%）。经过建模后，排序再次反转：M3 的模型深度 AbsRel 为 8.37%，优于 M2 的 9.29%。上游指标与最终场景忠实度相关，却不能彼此替代。</p></article><figure class="table-figure" id="table-2" data-table-key="table2"><figcaption><span>Table 2.</span> <span class="table-title"></span></figcaption><div class="table-scroll table-mount" role="region" tabindex="0" aria-label="Table 2"></div></figure><figure id="figure-1"><img loading="lazy" src="assets/figure14_abc.png" alt="Trajectory, native depth, and model depth"><figcaption><span>Figure 1.</span> 上游轨迹与原生深度的排序，并不会一一映射为冻结场景的深度排序。</figcaption></figure><article class="finding"><p class="finding-kicker">FINDING 02</p><h3>2. 几何证据把“合理布局”约束成更忠实的空间</h3><p>在声明配准且不使用 mesh ICP 的条件下，双向表面误差从纯 RGB 诊断基线的 0.376 m，依次降至 ViPE 的 0.118 m、ORB-SLAM3 的 0.082 m 与 GT 位姿的 0.072 m。M4 在本案例中拥有最强几何，而 M3 依靠估计轨迹已经缩小了大部分差距。</p></article><figure class="table-figure" id="table-3" data-table-key="table3"><figcaption><span>Table 3.</span> <span class="table-title"></span></figcaption><div class="table-scroll table-mount" role="region" tabindex="0" aria-label="Table 3"></div></figure><article class="finding finding-comparison"><p class="finding-kicker">FINDING 03</p><h3>3. 定性比较</h3></article><div class="analysis-block"><h3>不只看汇总指标，也检查重建场景</h3><p>在共享相机中将每个重建与 GT 对照：先从 M1 观察语义上合理的房间如何偏离米制布局；再比较 M2/M3，寻找几何改善但 RGB 相似度未同步提升的区域；最后检查 M4，把相机位姿不再是瓶颈后仍然存在的建模与材质误差分离出来。</p><figure id="figure-2"><img loading="lazy" src="assets/fixed_five_view_comparison_m1_m4.jpg" alt="M1 to M4 and input GT across five fixed views"><figcaption><span>Figure 2.</span> M1–M4 与输入 GT 在五个固定视角下的并排比较。</figcaption></figure><figure class="scene-comparison" id="figure-3" data-state="idle"><h3>旋转、缩放，检查冻结场景</h3>
+<div class="scene-toolbar"><label class="method-control">重建方法 / GT<select id="model-select"><option value="M1">M1</option><option value="M2">M2</option><option value="M3">M3</option><option value="M4" selected>M4</option><option value="GT">GT</option></select></label><label>显示方式<select id="scene-mode"><option value="compare">中轴线对比 GT</option><option value="single">单独查看所选场景</option></select></label><button id="scene-reset" type="button">重置视角</button></div>
+<div class="scene-stage"><span class="scene-label scene-label-left">M4</span><span class="scene-label scene-label-right">GT</span><button class="scene-divider" type="button" role="slider" aria-label="Comparison divider" aria-valuemin="2" aria-valuemax="98" aria-valuenow="50"></button><input id="scene-split" class="scene-split" type="range" min="2" max="98" value="50" aria-label="Comparison split"><p class="scene-status">滚动到此处加载三维对比。</p></div><button id="scene-retry" class="scene-retry" type="button">重试</button>
+<figcaption><span>Figure 3.</span> 所选冻结模型与 GT 共用同一相机和完整 viewport；网页运行时隐藏顶棚与四周墙体，以相同剖视方式检查，源模型文件保持不变。</figcaption></figure><figure class="fixed-comparison" id="figure-4"><div class="controlrow"><label>方法<select id="compare-method" disabled><option value="M1">M1</option><option value="M2">M2</option><option value="M3">M3</option><option value="M4">M4</option></select></label><label>固定视角<select id="compare-frame" disabled><option value="0">0</option><option value="36">36</option><option value="72">72</option><option value="108">108</option><option value="144">144</option></select></label></div><div class="compare-pair"><div><img id="compare-pred" src="assets/fixed_views/M1/000.png" alt="Selected frozen model rendering"><p id="compare-label" class="compare-subcaption">M1 / 视角 0</p></div><div><img id="compare-gt" src="assets/fixed_views/GT/000.png" alt="Input GT RGB"><p class="compare-subcaption">INPUT / GT RGB</p></div></div><figcaption><span>Figure 4.</span> 同一所选视角下的冻结模型渲染与输入 GT RGB。</figcaption></figure><p>网页只在运行时隐藏顶棚与四周墙体，以便剖视检查；运行时配准与显示选择不会修改可下载的冻结 GLB 或 Blend 文件。</p></div><div class="analysis-block"><h3>几何锚定改变了什么，又没有解决什么</h3><h4>建模像一种结构化、同时有损的正则化</h4><p>在共享的 180 视角协议下，M2–M4 的最终场景深度都显著接近 GT：相较原生 DA3，AbsRel 约下降 51–60%。一种可能解释是，对象级聚合抑制了局部不一致预测；但这仍是假说，而非受控因果结论。同一种抽象也可能删除真实表面与细节。</p><h4>测量约束生成，却不会自动解决外观</h4><p>几何最强的路径并未统治 PSNR 或 SSIM。位姿消除了一类不确定性，外观仍受材质估计、照明、物体细节与渲染器差异支配。只评 RGB 会漏掉空间进步，只评几何也会漏掉感知失败。</p><h4>100 视角检查衡量一致性，而不是未见视角泛化</h4><p>Table 7 使用一组固定随机种子的 100 个相机，它们从固定 5 帧之外的 175 张建模帧中抽取。这些视角相对固定检查是额外视角，但其 RGB 在建模阶段可见。在该集合上，M4 的深度 AbsRel 最低（7.47%），M3 的 RMSE 最低（1.090 m）。</p><figure id="figure-5"><img loading="lazy" src="assets/model_depth_error_five_views_m1_m4.png" alt="Model-depth error at five fixed evaluation views"><figcaption><span>Figure 5.</span> 固定 5 视角检查中的模型深度误差；青绿色表示缺失或无效预测。</figcaption></figure><figure class="table-figure" id="table-6" data-table-key="table6"><figcaption><span>Table 6.</span> <span class="table-title"></span></figcaption><div class="table-scroll table-mount" role="region" tabindex="0" aria-label="Table 6"></div></figure><figure class="table-figure" id="table-7" data-table-key="table7"><figcaption><span>Table 7.</span> <span class="table-title"></span></figcaption><div class="table-scroll table-mount" role="region" tabindex="0" aria-label="Table 7"></div></figure></div><div id="office-cafe" class="analysis-block external-case"><h3>走出仿真大厅：手机采集空间中还剩下什么？</h3><p>Office Café 不是 World Lobby 基准的第二次运行，也不构成同协议的定量验证。它是一个真实采集的外部案例，用于展示对象化场景、求解后的原视频相机轨迹、视频—模型对照与物理回放如何组合成一个可检查的空间产物。作者观察到的一类失败是：真实的弯曲转角被简化成方正结构。它提醒我们区分局部形状与全局尺度——尺度正确并不能自动修复形状错误；这里将其作为定性观察，而非另一组定量结果。</p><p class="office-links"><a href="https://office-cafe-vipe.hiwtishere.chatgpt.site/?lang=en" target="_blank" rel="noopener">全屏打开模型 ↗</a><a href="https://office-cafe-vipe.hiwtishere.chatgpt.site/shake?lang=en" target="_blank" rel="noopener">打开摇晃实验 ↗</a></p><div class="office-external"><h4>打开原始交互网页</h4><div class="office-external-grid"><figure class="office-external-card"><a href="https://office-cafe-vipe.hiwtishere.chatgpt.site/?lang=en" target="_blank" rel="noopener noreferrer"><img loading="lazy" src="assets/office-cafe/office-pmodel-poster.jpg" alt="交互模型: 来自已发布模型渲染视频的本地预览。"></a><figcaption><strong>交互模型</strong><p>来自已发布模型渲染视频的本地预览。</p><button type="button" data-office-external="https://office-cafe-vipe.hiwtishere.chatgpt.site/?lang=en" aria-controls="office-external-slot">加载外部网页 · 请求英文版</button></figcaption></figure><figure class="office-external-card"><a href="https://office-cafe-vipe.hiwtishere.chatgpt.site/shake?lang=en" target="_blank" rel="noopener noreferrer"><img loading="lazy" src="assets/office-cafe/office-shake-heavy-poster.jpg" alt="摇晃实验: 来自已发布较强摇晃回放的本地预览。"></a><figcaption><strong>摇晃实验</strong><p>来自已发布较强摇晃回放的本地预览。</p><button type="button" data-office-external="https://office-cafe-vipe.hiwtishere.chatgpt.site/shake?lang=en" aria-controls="office-external-slot">加载外部网页 · 请求英文版</button></figcaption></figure></div><p class="office-external-note">链接通过 lang=en 请求英文，实际语言与可用性由外站控制。这里展示的是本地内容预览，并非外站页面截图。若嵌入不可用，可使用全屏外链，或查看下方本地模型与视频。</p><div id="office-external-slot" class="office-external-slot" hidden></div></div><div class="office-model-grid"><figure class="office-model-card"><div class="office-model-stage" data-src="assets/office-cafe/scene.glb" aria-label="可旋转查看的 Office Café 原有模型"></div><figcaption><strong>原有模型</strong><span>输入为视频与估计深度；拖动旋转、右键平移、滚轮缩放。</span></figcaption></figure><figure class="office-model-card"><div class="office-model-stage" data-src="assets/office-cafe/atlas-scene.glb" aria-label="可旋转查看的 Office Café 旧模型"></div><figcaption><strong>旧模型</strong><span>输入仅为视频；该旧版布局未按 ViPE 扫描重新标定。</span></figcaption></figure></div><p class="office-model-note">两个模型默认采用俯视剖切视图，仅在显示时裁去上部遮挡，以展示室内布局；点击“剖切视图”可切换完整模型，“重置视角”可恢复总览。原始坐标、材质和模型文件保持不变。联动旋转用于观察布局，不代表几何配准。</p><div class="office-video-grid office-video-sync"><figure class="office-video-card"><video data-synced-video muted preload="metadata" playsinline poster="assets/office-cafe/office-input-poster.jpg" src="assets/office-cafe/office-input-web.mp4?v=4915b934b5f9" aria-label="输入视频"></video><figcaption><strong>输入视频</strong><span>原始手机拍摄视频。</span></figcaption></figure><figure class="office-video-card"><video data-synced-video muted preload="metadata" playsinline poster="assets/office-cafe/office-pmodel-poster.jpg" src="assets/office-cafe/office-pmodel-web.mp4?v=3b03fe58a192" aria-label="程序化模型"></video><figcaption><strong>程序化模型</strong><span>沿原视频相机轨迹渲染的模型。</span></figcaption></figure><figure class="office-video-card"><video data-synced-video muted preload="metadata" playsinline poster="assets/office-cafe/office-scan-poster.jpg" src="assets/office-cafe/office-scan-web.mp4?v=861a074637b4" aria-label="点云"></video><figcaption><strong>点云</strong><span>沿相同轨迹渲染的扫描点云。</span></figcaption></figure></div>
+<div class="office-video-controls"><button class="office-video-play" type="button" aria-label="播放全部三个视频" aria-pressed="false">▶</button><input class="office-video-progress" type="range" min="0" max="1000" value="0" aria-label="三个视频的共同播放进度"><output class="office-video-time">00:00 / 01:10</output><p class="office-video-status" role="status" aria-live="polite">同步播放 · 70.1 秒 · 30 fps</p></div><div class="office-video-grid"><figure class="office-video-card office-video-card-wide"><video controls preload="metadata" playsinline poster="assets/office-cafe/office-shake-heavy-poster.jpg" src="assets/office-cafe/office-shake-heavy.mp4"></video><figcaption><strong>较强地震效果</strong><span>场景中的物体均可交互，柜门可以打开；此段展示较强摇晃下的物理响应。</span></figcaption></figure></div><script type="module" src="office-models.js?v=8752a7cc4f64"></script></div></section>
+<section id="embodied-demo" class="embodied-demo">
+<p class="section-tag">DEMO / MAP-BASED EMBODIED EXECUTION</p><h2>一张重建地图，一个目标，一次具身执行。</h2><p>这个 demo 在 NVIDIA Isaac Sim 中将重建连接到下游使用：在导航地图中标出前台，设定路线，再让无人机、人形、四足和轮式机器人执行任务。同一个空间参照连接目标、路径约束与运动控制：���可以查看的重建场景，走向可以用于行动的地图。</p>
+<blockquote class="mission-command"><span>任务意图：</span>让机器人前往前台，并在那里排成一列。</blockquote>
+<ol class="mission-flow" aria-label="任务流程"><li><span>01</span>任务指令</li><li><span>02</span>地图上定位前台</li><li><span>03</span>设定导航路线</li><li><span>04</span>控制器执行</li></ol>
+<figure class="demo-film"><video controls playsinline preload="none" poster="assets/embodied-demo/navigation-214924-poster.png" width="1280" height="720" aria-label="四机器人地图导航演示"><source src="assets/embodied-demo/four-robots-214924-156s-phone.mp4" type="video/mp4">浏览器无法播放此视频，请使用下方下载链接。</video><figcaption>基于重建地图的仿真导航 · 预设路线 · 仿真器位姿反馈。作者确认，本演示已完成人工审阅。</figcaption></figure>
+<div class="demo-links"><a href="assets/embodied-demo/four-robots-214924-156s-phone.mp4" download>下载手机版 · 11.65 MB · 2:36</a><a href="assets/embodied-demo/four-robots-214924-156s-hd.mp4" download>下载高清版 · 49.69 MB · 2:36</a><a href="data/embodied_demo.json">运行与视频记录 ↗</a></div>
+<figure class="demo-map" id="demo-route-map"><a href="assets/embodied-demo/planned-route-156s.png" target="_blank" rel="noopener"><img loading="lazy" src="assets/embodied-demo/planned-route-156s.png" width="1920" height="1080" alt="标出前台与航点的四机器人路线设计图；静态设计，非执行轨迹"></a><figcaption><span>导航地图。</span> 在重建场景中标出前台与规划路线。这是静态设计图，而非实跑轨迹；点击查看原尺寸地图。</figcaption></figure>
+<details class="demo-protocol"><summary>演示条件与范围</summary><p>上方指令说明任务意图，不代表本次演示实现了语言到计划的自动解析。路线人工设定，并同时参考 M4 重建地图和原始几何筛选；这不是只依赖重建图的规划 benchmark。</p><p>成片为 156 秒、1× 原速：场景（0–6秒）、重建（6–12秒）、路线设计（12–18秒）、导航（18–156秒）。导航来自新运行 <code>20260930-214924</code>，略去开头等待的 12.32 秒，后续终点保持过程不在本剪辑中。地图是静态路线设计图，不是实测执行轨迹。</p><p>本 demo 不评测在线视觉定位、学习型指令理解、记忆检索或真实机器人迁移。这些是独立能力，不改变本演示所表达的“利用重建地图执行导航任务”。</p><p>人工展示审阅与下方保留的历史自动检查属于不同验证层次，不改变其原始判定。</p><p><a href="evidence/embodied-demo-audit.json">历史自动检查记录 ↗</a></p></details>
+</section>
+<section id="related-work"><p class="section-tag">05 / RELATED WORK</p><h2>从几何测量到可编辑 Real2Sim</h2><p>AWSM 连接两类角色：负责构建可编辑场景的智能体，以及将它作为空间参照使用的下游智能体。SLAM 与视觉惯性估计提供位姿和米制约束，ViPE 与 Depth Anything 3 提供几何观测。同期工作 AHa-3D 探索视频驱动、工具辅助的 Real2Sim，包括从视频估计相机与三维参照。我们的重点是几何锚定：将额外的物理测量，尤其是融合 IMU 的视觉惯性位姿与米制约束，引入智能体场景构建，而不只依赖视频推断的证据。四路线对比评估的是重建系统，并非 IMU 独立作用的消融实验；demo 展示的是地图应用，而不是完整的自主具身系统。</p><p>因此，我们的评测同时覆盖接口两端：上游系统作为测量管线接受检验，Blender 输出则作为持久化空间产物接受检验；这不同于孤立地评估位姿、深度或新视角合成。</p><span class="inline-sources"><a href="https://research.nvidia.com/labs/toronto-ai/vipe/" target="_blank" rel="noopener">ViPE</a> · <a href="https://arxiv.org/abs/2007.11898" target="_blank" rel="noopener">ORB-SLAM3</a> · <a href="https://arxiv.org/abs/2511.10647" target="_blank" rel="noopener">Depth Anything 3</a> · <a href="https://kevinxu02.github.io/real2sim-indoor-site/" target="_blank" rel="noopener">AHa-3D</a></span><div id="spatial-foundation" class="analysis-block spatial-foundation"><p class="section-tag">研究方向</p><h3>走向虚实融合智能体可持续使用的空间</h3><p>当前输出是可编辑的场景快照，demo 展示了针对特定场景的仿真集成。长期目标是把几何锚定的 agentic 重建发展为可复用流程，构建并维护虚实融合智能体能够使用的空间。保存可编辑资产是起点，不等于已经验证长期地图维护；以下方向是在当前证据基础上的延伸。</p><div class="capability-grid"><div><h4>导航</h4><p>在同一几何坐标中表达目标、自由空间与不同机器人的通行净空。更忠实的重建可以减少地图失配，可靠导航还需要定位与实时观测。</p></div><div><h4>记忆与多模态检索</h4><p>将对象身��、位置和源视图关联，让语言、图像与过去的观察指向同一个地点。检索到对象与安全到达对象，需要分别评测。</p></div><div><h4>交互与长期维护</h4><p>编辑对象属性与关系，重访变化区域，保留证据和版本。持久记忆应区分实际观察、推断补全和未知空间，而非把所有生成内容都当作事实。</p></div><div><h4>仿真与学习</h4><p>从当前特定场景的集成，走向可复用的仿真就绪导出流程，检查尺度、碰撞几何与物理参数。它可支持任务预演、场景变体和训练；是否降低策略 sim-to-real 差距，仍需单独检验。</p></div></div><p>并非每次寻物都需要完整网格：Memory Over Maps 用带位姿 RGB-D 关键帧按需定位，3D-Mem 与任务相关场景图也探索了不同记忆和表示方式。我们主张混合路线：保留视觉证据，细化任务相关几何，在有价值的地方构建可编辑仿真资产。下一步的关键问题是：什么时候证据已经足够支持行动，什么时候 agent 应重新观察并更新地图？</p><p>近期重点是提升重建效率与精度、形成可复用的仿真就绪场景生成流程，以及进一步与虚实融合智能体集成，让重建智能体的建模与修正循环连接到下游智能体的导航、记忆和交互需求。</p><p class="inline-sources"><a href="https://arxiv.org/abs/2603.20530" target="_blank" rel="noopener">Memory Over Maps</a> · <a href="https://arxiv.org/abs/2411.17735" target="_blank" rel="noopener">3D-Mem</a> · <a href="https://arxiv.org/abs/2404.13696" target="_blank" rel="noopener">Clio</a> · <a href="https://arxiv.org/abs/2404.06609" target="_blank" rel="noopener">GOAT-Bench</a></p></div></section>
+<section id="limitations"><p class="section-tag">06 / LIMITATIONS</p><h2>边界本身也是结果的一部分</h2><ul class="limits-list"><li>一个合成场景、每条路径一次工程运行，不能建立一般优越性或统计显著性。</li><li>四条路径是完整系统，而非严格的单变量消融；M2 与 M3 的差异不能只归因于 IMU。</li><li>M4 使用 GT 相机位姿，是诊断路线，不是可部署基线，也不是理论上界。</li><li>M1 模型空间指标依赖 GT-assisted Sim(3)，不能与原生米制恢复混为一谈。</li><li>100 视角外观与深度集合包含建模 RGB 输入，不是 held-out novel-view benchmark。</li><li>外观指标混合了几何、材质、照明与渲染误差；M4 revision 4 尚未完成完整独立视觉复审。</li><li>前台 demo 展示基于地图的仿真执行，不是 M1–M4 的受控导航对比。路线预先设定，定位使用仿真器位姿；具体范围见演示条件。</li><li>多模态空间记忆、长期地图维护与策略 sim-to-real 差距的降低是研究目标，不是这段四机器人视频已量化的结果。本页也没有 Office Café 拐角的同视角 GT 定量测量。</li><li>历史下游文档对同一 M4 Blend 哈希使用了不同的深度前端描述。本页冻结表格与资产保持不变；将旧任务收益归因于 DA3 之前，需要核对上游方法命名，详见来源说明。</li></ul><p>所有表格都由冻结资产生成；公开 manifest 提供模型哈希、配准、固定视角与固定随机评测索引。</p><div class="asset-downloads"><p><strong>M1</strong> <a href="models/M1/scene.blend" download>scene.blend ↓</a> · <a href="models/M1/scene.glb" download>scene.glb ↓</a> · <a href="evidence/M1/manifest.json">manifest ↗</a></p><p><strong>M2</strong> <a href="models/M2/scene.blend" download>scene.blend ↓</a> · <a href="models/M2/scene.glb" download>scene.glb ↓</a> · <a href="evidence/M2/manifest.json">manifest ↗</a></p><p><strong>M3</strong> <a href="models/M3/scene.blend" download>scene.blend ↓</a> · <a href="models/M3/scene.glb" download>scene.glb ↓</a> · <a href="evidence/M3/manifest.json">manifest ↗</a></p><p><strong>M4</strong> <a href="models/M4/scene.blend" download>scene.blend ↓</a> · <a href="models/M4/scene.glb" download>scene.glb ↓</a> · <a href="evidence/M4/manifest.json">manifest ↗</a></p></div><p class="evidence-links"><a href="data/tables_1_7.json">Tables JSON</a> · <a href="data/scene_comparison.json">Scene manifest</a> · <a href="data/fixed_views.json">Fixed views</a> · <a href="evidence/publication_manifest.json">Publication manifest</a> · <a href="evidence/SHA256SUMS">SHA256</a></p><p><a href="evidence/provenance-note.md">版本与资产来源说明 ↗</a></p></section>
+<section id="citation"><p class="section-tag">07 / CITATION</p><h2>引用本文</h2><p>如果本研究的受控实验、冻结资产或评测协议对你的工作有帮助，请引用下面的交互式研究文章。</p><pre class="citation-block"><code>@misc{awsm_2026,
+  title        = {{AWSM}: Agentic World Simulation and Mapping},
+  author       = {{Phygital AI}},
+  year         = {2026},
+  howpublished = {Interactive research article},
+  url          = {https://phygital-ai.github.io/agentic-world-simulation-and-mapping/},
+  note         = {Reconstruction benchmarks, editable scene assets, and map-based embodied demonstration}
+}</code></pre><a class="citation-download" href="data/awsm.bib" download="awsm.bib">下载 BibTeX 引用 <span aria-hidden="true">↓</span></a></section>
+<section id="references"><p class="section-tag">08 / REFERENCES</p><h2>参考文献</h2><ol class="references-list"><li><a href="https://kevinxu02.github.io/real2sim-indoor-site/" target="_blank" rel="noopener">AHa-3D: Agentic Tool Use for Real2Sim with GPT-6 Astra.</a></li><li><a href="https://arxiv.org/abs/2508.10934" target="_blank" rel="noopener">ViPE: Video Pose Engine for camera motion and near-metric geometry.</a></li><li><a href="https://arxiv.org/abs/2007.11898" target="_blank" rel="noopener">Campos et al. ORB-SLAM3: An Accurate Open-Source Library for Visual, Visual–Inertial, and Multimap SLAM. IEEE Transactions on Robotics, 2021.</a></li><li><a href="https://arxiv.org/abs/2511.10647" target="_blank" rel="noopener">Depth Anything 3: Recovering the Visual Space from Any Views.</a></li><li><a href="https://doi.org/10.1109/TIP.2003.819861" target="_blank" rel="noopener">Wang et al. Image Quality Assessment: From Error Visibility to Structural Similarity. IEEE Transactions on Image Processing, 2004.</a></li><li><a href="https://arxiv.org/abs/1801.03924" target="_blank" rel="noopener">Zhang et al. The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. CVPR, 2018.</a></li><li><a href="https://www.blender.org/" target="_blank" rel="noopener">Blender Foundation. Blender: Free and Open Source 3D Creation Suite.</a></li><li><a href="https://arxiv.org/abs/1708.03852" target="_blank" rel="noopener">Qin et al. VINS-Mono: A Robust and Versatile Monocular Visual-Inertial State Estimator. TRO, 2018.</a></li><li><a href="https://arxiv.org/abs/2509.13414" target="_blank" rel="noopener">Keetha et al. MapAnything: Universal Feed-Forward Metric 3D Reconstruction. 2025.</a></li><li><a href="https://arxiv.org/abs/2603.20530" target="_blank" rel="noopener">Zhou et al. Memory Over Maps: 3D Object Localization Without Reconstruction. 2026.</a></li><li><a href="https://arxiv.org/abs/2411.17735" target="_blank" rel="noopener">Yang et al. 3D-Mem: 3D Scene Memory for Embodied Exploration and Reasoning. 2024.</a></li><li><a href="https://arxiv.org/abs/2404.06609" target="_blank" rel="noopener">Khanna et al. GOAT-Bench: A Benchmark for Multi-Modal Lifelong Navigation. CVPR, 2024.</a></li><li><a href="https://arxiv.org/abs/2309.16650" target="_blank" rel="noopener">ConceptGraphs: Open-Vocabulary 3D Scene Graphs for Perception and Planning. ICRA, 2024.</a></li><li><a href="https://arxiv.org/abs/2404.13696" target="_blank" rel="noopener">Clio: Real-time Task-Driven Open-Set 3D Scene Graphs. RA-L, 2024.</a></li></ol></section><noscript>JavaScript is required for tables and interactive figures.</noscript></main></div><footer>Phygital AI · AWSM · frozen evidence, editable outputs · <a href="index.html">English</a></footer></body></html>