-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathindex.html
More file actions
652 lines (625 loc) · 34.2 KB
/
Copy pathindex.html
File metadata and controls
652 lines (625 loc) · 34.2 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="description"
content="SpatialWorld: Benchmarking Interactive Spatial Reasoning of Multimodal Agents in Real-World Tasks">
<meta name="keywords"
content="SpatialWorld, Multimodal Agents, Spatial Reasoning, Embodied AI, 3D Environments, Benchmark">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>SpatialWorld: Benchmarking Interactive Spatial Reasoning of Multimodal Agents</title>
<link href="https://fonts.googleapis.com/css?family=Google+Sans|Noto+Sans|Castoro" rel="stylesheet">
<link rel="stylesheet" href="static/css/bulma.min.css">
<link rel="stylesheet" href="static/css/fontawesome.all.min.css">
<link rel="stylesheet" href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css">
<link rel="stylesheet" href="static/css/index.css">
<link rel="stylesheet" href="css/custom.css">
<script defer src="static/js/fontawesome.all.min.js"></script>
<link href="https://cdnjs.cloudflare.com/ajax/libs/prism/1.29.0/themes/prism.min.css" rel="stylesheet" />
<link href="https://cdnjs.cloudflare.com/ajax/libs/prism/1.29.0/plugins/line-numbers/prism-line-numbers.min.css"
rel="stylesheet" />
<script src="https://cdnjs.cloudflare.com/ajax/libs/prism/1.29.0/prism.min.js"></script>
<script src="https://cdnjs.cloudflare.com/ajax/libs/prism/1.29.0/plugins/line-numbers/prism-line-numbers.min.js"></script>
<script src="https://cdnjs.cloudflare.com/ajax/libs/prism/1.29.0/components/prism-bibtex.min.js"></script>
<base target="_blank">
</head>
<body>
<section class="hero hero-gradient">
<div class="hero-body">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column has-text-centered">
<h1 class="title is-2 publication-title">
SpatialWorld: Benchmarking Interactive Spatial Reasoning of Multimodal Agents in Real-World Tasks
</h1>
<div class="is-size-5 publication-authors">
<span class="author-block"><a href="#">Hongcheng Gao</a><sup>*†1,4</sup>,</span>
<span class="author-block"><a href="#">Hailong Qu</a><sup>*2</sup>,</span>
<span class="author-block"><a href="#">Jingyi Tang</a><sup>3</sup>,</span>
<span class="author-block"><a href="#">Jiahao Wang</a><sup>5</sup>,</span>
<span class="author-block"><a href="#">Hengkang Qiao</a><sup>2</sup>,</span>
<br>
<span class="author-block"><a href="#">Zihao Huang</a><sup>6</sup>,</span>
<span class="author-block"><a href="#">Shihong Huang</a><sup>3</sup>,</span>
<span class="author-block"><a href="#">Junming Yang</a><sup>7</sup>,</span>
<span class="author-block"><a href="#">Yi Li</a><sup>1</sup>,</span>
<span class="author-block"><a href="#">Hongyixuan Yuan</a><sup>2</sup>,</span>
<br>
<span class="author-block"><a href="#">Wenjie Li</a><sup>8</sup>,</span>
<span class="author-block"><a href="#">Bohan Zeng</a><sup>3</sup>,</span>
<span class="author-block"><a href="#">Wenbo Li</a><sup>9</sup>,</span>
<span class="author-block"><a href="#">Bo Wang</a><sup>6</sup>,</span>
<span class="author-block"><a href="#">Jianhui Liu</a><sup>10</sup>,</span>
<span class="author-block"><a href="#">Zhiqi Huang</a><sup>3</sup>,</span>
<br>
<span class="author-block"><a href="#">Haoyang Huang</a><sup>9</sup>,</span>
<span class="author-block"><a href="#">Wentao Zhang</a><sup>3</sup>,</span>
<span class="author-block"><a href="#">Guoqing Huang</a><sup>2</sup>,</span>
<span class="author-block"><a href="#">Nan Duan</a><sup>9</sup>,</span>
<span class="author-block"><a href="#">Yinpeng Dong</a><sup>†1</sup></span>
</div>
<div class="is-size-6 publication-authors" style="color: #666; margin-top: 8px;">
<span class="author-block"><sup>1</sup>Tsinghua University,</span>
<span class="author-block"><sup>2</sup>Chongqing University,</span>
<span class="author-block"><sup>3</sup>Peking University,</span>
<span class="author-block"><sup>4</sup>ZenoMind AI,</span>
<br>
<span class="author-block"><sup>5</sup>Xi'an Jiaotong University,</span>
<span class="author-block"><sup>6</sup>Beijing Institute of Technology,</span>
<span class="author-block"><sup>7</sup>Southeast University,</span>
<br>
<span class="author-block"><sup>8</sup>Shanghai Jiao Tong University,</span>
<span class="author-block"><sup>9</sup>Joy Future Academy,</span>
<span class="author-block"><sup>10</sup>The University of Hong Kong</span>
</div>
<p class="is-size-6" style="margin-top: 8px; color: #888;">
<sup>*</sup>Equal contribution · <sup>†</sup>Corresponding author
</p>
<div class="column has-text-centered" style="margin-top: 16px;">
<div class="publication-links">
<span class="link-block">
<a href="#" class="external-link button is-normal is-rounded is-dark">
<span class="icon"><i class="ai ai-arxiv"></i></span>
<span>Paper</span>
</a>
</span>
<span class="link-block">
<a href="#" class="external-link button is-normal is-rounded is-dark">
<span class="icon"><i class="fab fa-github"></i></span>
<span>Code</span>
</a>
</span>
<span class="link-block">
<a href="#" class="external-link button is-normal is-rounded is-dark">
<span class="icon"><i class="fas fa-database"></i></span>
<span>Benchmark</span>
</a>
</span>
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<!-- Abstract -->
<section class="section">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column">
<h2 class="title is-3 has-text-centered">Abstract</h2>
<div class="content has-text-justified">
<p>
Spatial reasoning is a foundational capability for multimodal large language models (MLLMs) to perceive
and operate within the physical world. However, existing benchmarks predominantly rely on passive
evaluation (e.g., static VQA) or simulator-specific pipelines, failing to assess general interactive
spatial understanding. We introduce <strong>SpatialWorld</strong>, a unified benchmark designed
specifically for evaluating the interactive spatial understanding of multimodal agents in complex
real-world tasks.
</p>
<p>
Integrating eight heterogeneous simulation backends under a shared, simulator-agnostic protocol,
SpatialWorld features <strong>760 human-annotated tasks</strong> across diverse domains (e.g.,
household routines, travel, social collaboration). Agents must solve tasks under vision-only partial
observability, actively gathering egocentric visual evidence and expressing decisions via a unified,
text-based action interface native to MLLMs. For reliable evaluation, each task includes a
human-validated initial state, a reference trajectory, and a terminal-state verifier.
</p>
<p>
Evaluating 15 advanced agents reveals that robust spatial task solving remains challenging: the strongest
model, <strong>GPT-5</strong>, achieves an average task success rate (TSR) of only
<strong>20.0%</strong>, while the leading open-source model, <strong>Qwen-3.5</strong>, reaches
<strong>15.6%</strong>. Further analysis exposes a clear mismatch between task success and execution
efficiency, alongside substantial domain-specific performance variations. These bottlenecks in active
exploration and long-horizon planning position SpatialWorld as a rigorous testbed for future spatial
agents.
</p>
</div>
<div class="stat-grid">
<div class="stat-card">
<div class="stat-value">760</div>
<div class="stat-label">Human-Annotated Tasks</div>
</div>
<div class="stat-card">
<div class="stat-value">8</div>
<div class="stat-label">Simulation Backends</div>
</div>
<div class="stat-card">
<div class="stat-value">6</div>
<div class="stat-label">Scenario Categories</div>
</div>
<div class="stat-card">
<div class="stat-value">15</div>
<div class="stat-label">Evaluated MLLM Agents</div>
</div>
</div>
</div>
</div>
</div>
</section>
<!-- Framework Overview -->
<section class="section" style="padding-top: 0;">
<div class="container is-max-desktop">
<h2 class="title is-3 has-text-centered">Benchmark Overview</h2>
<div class="content has-text-justified">
<p>
SpatialWorld is a scalable, general-purpose evaluation framework for multimodal agents, supporting
end-to-end task solving and structured plan generation. It unifies diverse 3D backends under a standardized
observation-action interface, enabling rigorous assessment of interactive spatial reasoning via reproducible
benchmarks and automated efficiency metrics.
</p>
</div>
<figure class="has-text-centered" id="framework-figure">
<img src="imgs/spatialworld_main.png" alt="SpatialWorld Framework Overview" class="teaser-image"
onerror="this.parentElement.style.display='none'">
<figcaption class="is-size-7 has-text-grey" style="margin-top: 10px;">
SpatialWorld unifies eight 3D simulation backends under a shared closed-loop evaluation protocol.
</figcaption>
</figure>
</div>
</section>
<!-- Eight Environments -->
<section class="section" style="padding-top: 0;">
<div class="container is-max-desktop">
<h2 class="title is-3 has-text-centered">Eight Simulation Backends</h2>
<div class="env-grid">
<div class="env-card">
<h4>🏠 AI2-THOR</h4>
<p>Near-photorealistic indoor scenes with rich object affordances for household manipulation. <strong>311 tasks</strong></p>
</div>
<div class="env-card">
<h4>🏗️ ProcTHOR</h4>
<p>Procedurally generated indoor layouts testing generalization across diverse room configurations. <strong>127 tasks</strong></p>
</div>
<div class="env-card">
<h4>🛋️ VirtualHome</h4>
<p>Daily activity scripts in home environments with multi-step routines. <strong>38 tasks</strong></p>
</div>
<div class="env-card">
<h4>🚗 CARLA</h4>
<p>Urban traffic simulation for outdoor navigation and travel-oriented tasks. <strong>80 tasks</strong></p>
</div>
<div class="env-card">
<h4>🌆 EmbodiedCity</h4>
<p>Large-scale city navigation with realistic pedestrian and vehicle dynamics. <strong>53 tasks</strong></p>
</div>
<div class="env-card">
<h4>👥 Multi-AI2THOR</h4>
<p>Multi-agent social collaboration in shared indoor environments. <strong>29 tasks</strong></p>
</div>
<div class="env-card">
<h4>👥 Multi-ProcTHOR</h4>
<p>Coordinated multi-agent tasks in procedurally generated scenes. <strong>17 tasks</strong></p>
</div>
<div class="env-card">
<h4>🎮 3D Games</h4>
<p>Lightweight digital environments (Block 3D, Maze 3D, Snake, Rubik's Cube) for abstract spatial reasoning. <strong>105 tasks</strong></p>
</div>
</div>
<h3 class="title is-4" style="margin-top: 40px;">Complexity Levels</h3>
<div class="complexity-grid">
<div class="complexity-card">
<h4>🧭 Navigation</h4>
<p>Explore the 3D environment and reach a target location or object without manipulating environment state.</p>
</div>
<div class="complexity-card">
<h4>🤲 Interaction</h4>
<p>Perform object-level state changes (pick, place, open, toggle) without extensive spatial exploration.</p>
</div>
<div class="complexity-card">
<h4>🔀 Hybrid</h4>
<p>Combine long-horizon navigation with multi-step manipulation, demanding both exploration and fine-grained interaction.</p>
</div>
</div>
</div>
</section>
<!-- Task Gallery -->
<section class="section" style="padding-top: 0;">
<div class="container is-max-widescreen">
<h2 class="title is-3 has-text-centered">Task Examples</h2>
<p class="content has-text-centered" style="max-width: 800px; margin: 0 auto 20px; color: #64748b;">
Representative tasks from SpatialWorld. Agents receive only a natural-language instruction and egocentric RGB observations.
</p>
<div class="tabs is-centered gallery-tabs" id="galleryTabs">
<ul>
<li class="is-active" data-tab="daily"><a>Daily</a></li>
<li data-tab="travel"><a>Travel</a></li>
<li data-tab="work"><a>Work & Study</a></li>
<li data-tab="social"><a>Social</a></li>
<li data-tab="digital"><a>Digital Games</a></li>
</ul>
</div>
<div class="gallery-panel" data-panel="daily">
<div class="task-gallery">
<div class="task-card">
<div class="task-card-visual" style="background: linear-gradient(135deg, #0f766e, #14b8a6);">AI2-THOR</div>
<div class="task-card-body">
<div class="task-card-env">AI2-THOR · Daily</div>
<div class="task-card-instruction">"I found the lettuce was rotten; please help me throw it in the trash."</div>
<div class="task-card-meta">
<span class="tag interact">Interaction</span>
<span class="tag">object_in_receptacle</span>
</div>
</div>
</div>
<div class="task-card">
<div class="task-card-visual" style="background: linear-gradient(135deg, #0369a1, #38bdf8);">ProcTHOR</div>
<div class="task-card-body">
<div class="task-card-env">ProcTHOR · Daily</div>
<div class="task-card-instruction">"Please put the dirty plate into the dishwasher and close it."</div>
<div class="task-card-meta">
<span class="tag hybrid">Hybrid</span>
<span class="tag">multi-step manipulation</span>
</div>
</div>
</div>
<div class="task-card">
<div class="task-card-visual" style="background: linear-gradient(135deg, #7c3aed, #a78bfa);">VirtualHome</div>
<div class="task-card-body">
<div class="task-card-env">VirtualHome · Daily</div>
<div class="task-card-instruction">"I need to tidy up the kitchen. Please open the refrigerator door and put the salmon inside, but do not close the refrigerator door."</div>
<div class="task-card-meta">
<span class="tag hybrid">Hybrid</span>
<span class="tag">object_state</span>
</div>
</div>
</div>
</div>
</div>
<div class="gallery-panel" data-panel="travel" style="display:none;">
<div class="task-gallery">
<div class="task-card">
<div class="task-card-visual" style="background: linear-gradient(135deg, #b45309, #fbbf24);">CARLA</div>
<div class="task-card-body">
<div class="task-card-env">CARLA · Travel</div>
<div class="task-card-instruction">"Walk to the position marked by the red line in the screenshot. You can turn and move in any direction."</div>
<div class="task-card-meta">
<span class="tag nav">Navigation</span>
<span class="tag">distance_to_waypoint</span>
</div>
</div>
</div>
<div class="task-card">
<div class="task-card-visual" style="background: linear-gradient(135deg, #1e40af, #60a5fa);">EmbodiedCity</div>
<div class="task-card-body">
<div class="task-card-env">EmbodiedCity · Travel</div>
<div class="task-card-instruction">"Navigate to the bus stop at the end of the street and wait there."</div>
<div class="task-card-meta">
<span class="tag nav">Navigation</span>
<span class="tag">urban navigation</span>
</div>
</div>
</div>
</div>
</div>
<div class="gallery-panel" data-panel="work" style="display:none;">
<div class="task-gallery">
<div class="task-card">
<div class="task-card-visual" style="background: linear-gradient(135deg, #0f766e, #2dd4bf);">AI2-THOR</div>
<div class="task-card-body">
<div class="task-card-env">AI2-THOR · Work & Study</div>
<div class="task-card-instruction">"Please organize the documents on the desk and place the laptop in the drawer."</div>
<div class="task-card-meta">
<span class="tag hybrid">Hybrid</span>
<span class="tag">office scene</span>
</div>
</div>
</div>
<div class="task-card">
<div class="task-card-visual" style="background: linear-gradient(135deg, #4338ca, #818cf8);">ProcTHOR</div>
<div class="task-card-body">
<div class="task-card-env">ProcTHOR · Work & Study</div>
<div class="task-card-instruction">"Find the whiteboard marker and place it on the conference table."</div>
<div class="task-card-meta">
<span class="tag hybrid">Hybrid</span>
<span class="tag">search & place</span>
</div>
</div>
</div>
</div>
</div>
<div class="gallery-panel" data-panel="social" style="display:none;">
<div class="task-gallery">
<div class="task-card">
<div class="task-card-visual" style="background: linear-gradient(135deg, #be185d, #f472b6);">Multi-AI2THOR</div>
<div class="task-card-body">
<div class="task-card-env">Multi-AI2THOR · Social Collaboration</div>
<div class="task-card-instruction">"Work with your partner to set the dining table: you handle the plates while they handle the utensils."</div>
<div class="task-card-meta">
<span class="tag hybrid">Hybrid</span>
<span class="tag">multi-agent</span>
</div>
</div>
</div>
<div class="task-card">
<div class="task-card-visual" style="background: linear-gradient(135deg, #9333ea, #c084fc);">Multi-ProcTHOR</div>
<div class="task-card-body">
<div class="task-card-env">Multi-ProcTHOR · Social Collaboration</div>
<div class="task-card-instruction">"Coordinate with the other agent to move the heavy box to the storage room together."</div>
<div class="task-card-meta">
<span class="tag hybrid">Hybrid</span>
<span class="tag">coordination</span>
</div>
</div>
</div>
</div>
</div>
<div class="gallery-panel" data-panel="digital" style="display:none;">
<div class="task-gallery">
<div class="task-card">
<div class="task-card-visual" style="background: linear-gradient(135deg, #059669, #34d399);">Maze 3D</div>
<div class="task-card-body">
<div class="task-card-env">3D Games · Digital</div>
<div class="task-card-instruction">"Navigate through the 3D maze from the green start point to the red exit."</div>
<div class="task-card-meta">
<span class="tag nav">Navigation</span>
<span class="tag">abstract spatial</span>
</div>
</div>
</div>
<div class="task-card">
<div class="task-card-visual" style="background: linear-gradient(135deg, #dc2626, #f87171);">Rubik's Cube</div>
<div class="task-card-body">
<div class="task-card-env">3D Games · Digital</div>
<div class="task-card-instruction">"Rotate the faces to solve the Rubik's cube and match all sides."</div>
<div class="task-card-meta">
<span class="tag interact">Interaction</span>
<span class="tag">geometric reasoning</span>
</div>
</div>
</div>
<div class="task-card">
<div class="task-card-visual" style="background: linear-gradient(135deg, #ca8a04, #fde047);">Snake 3D</div>
<div class="task-card-body">
<div class="task-card-env">3D Games · Digital</div>
<div class="task-card-instruction">"Control the snake to eat the food without hitting the walls or itself."</div>
<div class="task-card-meta">
<span class="tag hybrid">Hybrid</span>
<span class="tag">planning</span>
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<!-- Scenario Distribution -->
<section class="section" style="padding-top: 0;">
<div class="container is-max-desktop">
<h2 class="title is-3 has-text-centered">Scenario Distribution</h2>
<div class="performance-table-wrapper">
<table class="scenario-table" style="width:100%; border-collapse: collapse;">
<thead>
<tr>
<th>Environment</th>
<th>Daily</th>
<th>Work</th>
<th>Entertain.</th>
<th>Travel</th>
<th>Social</th>
<th>Total</th>
</tr>
</thead>
<tbody>
<tr>
<td style="text-align:left; font-weight:500;">AI2-THOR</td>
<td>219</td><td>41</td><td>40</td><td>11</td><td>0</td><td>311</td>
</tr>
<tr>
<td style="text-align:left; font-weight:500;">ProcTHOR</td>
<td>92</td><td>10</td><td>23</td><td>2</td><td>0</td><td>127</td>
</tr>
<tr>
<td style="text-align:left; font-weight:500;">VirtualHome</td>
<td>27</td><td>8</td><td>3</td><td>0</td><td>0</td><td>38</td>
</tr>
<tr>
<td style="text-align:left; font-weight:500;">CARLA</td>
<td>0</td><td>0</td><td>0</td><td>80</td><td>0</td><td>80</td>
</tr>
<tr>
<td style="text-align:left; font-weight:500;">EmbodiedCity</td>
<td>12</td><td>0</td><td>2</td><td>39</td><td>0</td><td>53</td>
</tr>
<tr>
<td style="text-align:left; font-weight:500;">Multi-AI2THOR</td>
<td>0</td><td>0</td><td>0</td><td>0</td><td>29</td><td>29</td>
</tr>
<tr>
<td style="text-align:left; font-weight:500;">Multi-ProcTHOR</td>
<td>0</td><td>0</td><td>0</td><td>0</td><td>17</td><td>17</td>
</tr>
<tr>
<td style="text-align:left; font-weight:500;">3D Games</td>
<td>0</td><td>0</td><td>105</td><td>0</td><td>0</td><td>105</td>
</tr>
<tr>
<td style="text-align:left;">Total</td>
<td>350</td><td>59</td><td>173</td><td>132</td><td>46</td><td>760</td>
</tr>
</tbody>
</table>
</div>
</div>
</section>
<!-- Main Results TSR -->
<section class="section" style="padding-top: 0;">
<div class="container is-max-widescreen">
<h2 class="title is-3 has-text-centered">Main Results — Task Success Rate (TSR %)</h2>
<p class="has-text-centered is-size-6" style="color:#64748b; margin-bottom: 16px;">
Performance across physical scenario categories and digital 3D games. Bold = best, underline = second-best per column.
</p>
<div class="performance-table-wrapper">
<table class="performance-table">
<thead>
<tr>
<th rowspan="2" style="vertical-align:middle;">Model</th>
<th colspan="6">Physical</th>
<th>Digital</th>
</tr>
<tr>
<th>Daily</th>
<th>Work</th>
<th>Entertain.</th>
<th>Travel</th>
<th>Social</th>
<th>Overall</th>
<th>Entertain.</th>
</tr>
</thead>
<tbody>
<tr class="group-header"><td colspan="8">(A) Open-Source Models</td></tr>
<tr><td>Qwen2.5-VL-72B</td><td>3.8</td><td>8.5</td><td>3.0</td><td>1.1</td><td>2.2</td><td>3.7</td><td>7.6</td></tr>
<tr><td>Qwen3-VL-30B-A3B</td><td>5.3</td><td>5.1</td><td>4.5</td><td>1.1</td><td>4.3</td><td>4.1</td><td>7.9</td></tr>
<tr><td>Qwen3-VL-235B-Instruct</td><td>6.9</td><td>8.5</td><td>7.4</td><td>4.5</td><td>10.9</td><td>7.6</td><td>5.0</td></tr>
<tr><td>Qwen3-VL-235B-Thinking</td><td>5.7</td><td>8.5</td><td>7.4</td><td>3.8</td><td>10.9</td><td>7.3</td><td>28.3</td></tr>
<tr><td>Qwen-3.5-397B-A17B</td><td class="second">13.1</td><td class="best">16.9</td><td class="best">13.2</td><td>4.5</td><td class="second">19.6</td><td class="second">13.5</td><td>26.0</td></tr>
<tr><td>GLM-4.5V</td><td>3.6</td><td>3.4</td><td>4.5</td><td>2.2</td><td>13.0</td><td>5.3</td><td>14.5</td></tr>
<tr><td>GLM-4.6V</td><td>2.9</td><td>5.1</td><td>4.4</td><td>1.5</td><td>0.0</td><td>2.8</td><td>8.1</td></tr>
<tr><td>Kimi-VL-A3B</td><td>1.1</td><td>3.4</td><td>0.0</td><td>0.0</td><td>0.0</td><td>0.9</td><td>3.3</td></tr>
<tr><td>Kimi-K2.5</td><td>11.1</td><td>8.5</td><td>4.4</td><td>3.8</td><td>17.4</td><td>9.0</td><td>31.0</td></tr>
<tr class="group-header"><td colspan="8">(B) Closed-Source Models</td></tr>
<tr><td>Gemini-2.5-Pro</td><td>7.4</td><td class="second">11.9</td><td>1.5</td><td>3.8</td><td>10.9</td><td>7.1</td><td>32.6</td></tr>
<tr><td>Gemini-3-Flash</td><td>8.0</td><td>10.2</td><td>4.4</td><td class="second">6.1</td><td>4.3</td><td>6.6</td><td class="second">38.1</td></tr>
<tr><td>Gemini-3-Pro</td><td>11.4</td><td>10.2</td><td>5.9</td><td>4.5</td><td>8.7</td><td>8.1</td><td class="best">39.0</td></tr>
<tr><td>GPT-5</td><td class="best">14.9</td><td class="best">16.9</td><td class="second">10.3</td><td class="best">6.8</td><td class="best">34.8</td><td class="best">16.7</td><td>36.4</td></tr>
<tr><td>GPT-5.4</td><td>8.0</td><td>5.1</td><td>5.9</td><td>3.8</td><td>6.5</td><td>5.9</td><td>11.9</td></tr>
<tr><td>Doubao-2.0-Lite</td><td>5.7</td><td>6.8</td><td>5.9</td><td>3.0</td><td>13.0</td><td>6.9</td><td>24.8</td></tr>
</tbody>
</table>
</div>
</div>
</section>
<!-- SE Results -->
<section class="section" style="padding-top: 0;">
<div class="container is-max-widescreen">
<h2 class="title is-3 has-text-centered">Step Efficiency (SE)</h2>
<p class="has-text-centered is-size-6" style="color:#64748b; margin-bottom: 16px;">
SE = reference step count / actual step count on successful trajectories. Higher is more efficient.
"—" indicates no successful trajectory in that category.
</p>
<div class="performance-table-wrapper">
<table class="performance-table">
<thead>
<tr>
<th rowspan="2" style="vertical-align:middle;">Model</th>
<th colspan="6">Physical</th>
<th>Digital</th>
</tr>
<tr>
<th>Daily</th>
<th>Work</th>
<th>Entertain.</th>
<th>Travel</th>
<th>Social</th>
<th>Overall</th>
<th>Entertain.</th>
</tr>
</thead>
<tbody>
<tr class="group-header"><td colspan="8">(A) Open-Source Models</td></tr>
<tr><td>Qwen2.5-VL-72B</td><td>0.545</td><td>0.510</td><td>0.458</td><td class="second">0.889</td><td>0.143</td><td>0.509</td><td>0.688</td></tr>
<tr><td>Qwen3-VL-30B-A3B</td><td>0.686</td><td class="second">0.667</td><td>0.500</td><td class="best">0.968</td><td>0.174</td><td class="second">0.599</td><td>0.765</td></tr>
<tr><td>Qwen3-VL-235B-Instruct</td><td class="second">0.708</td><td>0.574</td><td>0.529</td><td>0.449</td><td>0.243</td><td>0.501</td><td>0.397</td></tr>
<tr><td>Qwen3-VL-235B-Thinking</td><td>0.536</td><td>0.453</td><td>0.424</td><td>0.524</td><td>0.218</td><td>0.431</td><td>0.747</td></tr>
<tr><td>Qwen-3.5-397B-A17B</td><td>0.552</td><td>0.477</td><td>0.453</td><td>0.633</td><td>0.290</td><td>0.481</td><td>0.737</td></tr>
<tr><td>GLM-4.5V</td><td>0.677</td><td>0.583</td><td>0.482</td><td>0.450</td><td>0.270</td><td>0.492</td><td class="second">0.809</td></tr>
<tr><td>GLM-4.6V</td><td>0.705</td><td>0.381</td><td>0.444</td><td>0.417</td><td>—</td><td>0.487</td><td>0.920</td></tr>
<tr><td>Kimi-VL-A3B</td><td>0.636</td><td>0.333</td><td>—</td><td>—</td><td>—</td><td>0.485</td><td class="best">0.948</td></tr>
<tr><td>Kimi-K2.5</td><td>0.519</td><td>0.556</td><td>0.517</td><td>0.553</td><td>0.226</td><td>0.474</td><td>0.626</td></tr>
<tr class="group-header"><td colspan="8">(B) Closed-Source Models</td></tr>
<tr><td>Gemini-2.5-Pro</td><td>0.615</td><td>0.567</td><td class="best">0.667</td><td>0.483</td><td class="second">0.399</td><td>0.546</td><td>0.518</td></tr>
<tr><td>Gemini-3-Flash</td><td>0.575</td><td>0.390</td><td>0.504</td><td>0.612</td><td>0.183</td><td>0.453</td><td>0.657</td></tr>
<tr><td>Gemini-3-Pro</td><td class="second">0.708</td><td>0.544</td><td>0.466</td><td>0.732</td><td>0.281</td><td>0.546</td><td>0.717</td></tr>
<tr><td>GPT-5</td><td>0.597</td><td>0.540</td><td>0.387</td><td>0.544</td><td>0.248</td><td>0.463</td><td>0.583</td></tr>
<tr><td>GPT-5.4</td><td>0.617</td><td class="second">0.667</td><td>0.427</td><td>0.513</td><td>0.305</td><td>0.506</td><td>0.720</td></tr>
<tr><td>Doubao-2.0-Lite</td><td class="best">0.776</td><td class="best">0.708</td><td class="second">0.604</td><td class="second">0.708</td><td class="best">0.522</td><td class="best">0.664</td><td>0.599</td></tr>
</tbody>
</table>
</div>
</div>
</section>
<!-- Key Findings -->
<section class="section" style="padding-top: 0;">
<div class="container is-max-desktop">
<h2 class="title is-3 has-text-centered">Key Findings</h2>
<div class="finding-item">
<strong>Far from reliable 3D task solving.</strong>
GPT-5 achieves only 16.7% Physical Overall TSR; Qwen-3.5-397B-A17B reaches 13.5%. Even the strongest models struggle across all scenario categories.
</div>
<div class="finding-item">
<strong>Success ≠ efficiency.</strong>
Models with higher TSR do not necessarily achieve higher step efficiency. Doubao-2.0-Lite leads in SE (0.664 Physical Overall) despite moderate TSR, suggesting many successful trajectories involve redundant exploration.
</div>
<div class="finding-item">
<strong>Domain-specific strengths.</strong>
GPT-5 leads daily household and social collaboration; Qwen-3.5-397B-A17B ties GPT-5 in Work & Study and leads physical entertainment; Gemini-3-Pro achieves the highest scores on digital 3D games (39.0% TSR).
</div>
<div class="finding-item">
<strong>Vision-only closed-loop evaluation.</strong>
Unlike static VQA or simulator-specific pipelines, SpatialWorld requires agents to actively explore under partial observability using only egocentric RGB and a text-based action interface.
</div>
</div>
</section>
<!-- Citation -->
<section class="section">
<div class="container is-max-desktop">
<h2 class="title is-3 has-text-centered">Citation</h2>
<div class="content">
<pre class="line-numbers"><code class="language-bibtex">@article{gao2026spatialworld,
title={SpatialWorld: Benchmarking Interactive Spatial Reasoning of Multimodal Agents in Real-World Tasks},
author={Gao, Hongcheng and Qu, Hailong and Tang, Jingyi and Wang, Jiahao and Qiao, Hengkang and Huang, Zihao and Huang, Shihong and Yang, Junming and Li, Yi and Yuan, Hongyixuan and Li, Wenjie and Zeng, Bohan and Li, Wenbo and Wang, Bo and Liu, Jianhui and Huang, Zhiqi and Huang, Haoyang and Zhang, Wentao and Huang, Guoqing and Duan, Nan and Dong, Yinpeng},
journal={arXiv preprint arXiv:XXXX.XXXXX},
year={2026}
}</code></pre>
</div>
</div>
</section>
<footer class="footer">
<div class="container">
<div class="content has-text-centered">
<p style="color:#64748b;">
Website template adapted from
<a href="https://github.com/TuringEyeTest/turingeyetest.github.io">TuringEyeTest Project Page</a>.
</p>
<p style="color:#64748b; font-size:0.85rem;">
Project page for <a href="https://spatialworld.github.io">SpatialWorld</a>.
</p>
</div>
</div>
</footer>
<script>
document.querySelectorAll('#galleryTabs li').forEach(function (tab) {
tab.addEventListener('click', function () {
var target = tab.dataset.tab;
document.querySelectorAll('#galleryTabs li').forEach(function (t) {
t.classList.toggle('is-active', t === tab);
});
document.querySelectorAll('.gallery-panel').forEach(function (panel) {
panel.style.display = panel.dataset.panel === target ? 'block' : 'none';
});
});
});
</script>
</body>
</html>