AWSM / index.html
Ooliva's picture
Upload AWSM article, metrics, manifests, and scripts
4c427dc verified
Raw History Blame Contribute Delete
41.1 kB
<!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>