-
Notifications
You must be signed in to change notification settings - Fork 23
Expand file tree
/
Copy pathindex.html
More file actions
591 lines (556 loc) · 31.3 KB
/
Copy pathindex.html
File metadata and controls
591 lines (556 loc) · 31.3 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
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8" />
<meta name="viewport" content="width=device-width, initial-scale=1.0" />
<meta
name="description"
content="World Action Models: The Next Frontier in Embodied AI. A survey of definitions and formalism, VLAs and world models, architecture, training data, evaluation, and open challenges."
/>
<meta name="theme-color" content="#f6f7f3" />
<title>WAM Survey | World Action Models: The Next Frontier in Embodied AI</title>
<link rel="stylesheet" href="./styles.css" />
<link rel="preconnect" href="https://fonts.googleapis.com" />
<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin />
<link
href="https://fonts.googleapis.com/css2?family=Sora:wght@300;400;500;600;700&family=Geist+Mono:wght@400;500;600&display=swap"
rel="stylesheet"
/>
<script defer src="https://cloud.umami.is/script.js" data-website-id="9e371790-3ee3-4359-801c-0df46be17255"></script>
<link rel="icon" href="./favicon.ico?v=20260513b" type="image/x-icon">
</head>
<body>
<a class="skip-link" href="#library">Skip to paper library</a>
<nav class="sidebar" aria-label="Page navigation">
<a class="sidebar-brand" href="#overview" aria-label="WAM Survey">
<span class="sidebar-mark">WAM</span>
<span class="sidebar-name">Survey</span>
</a>
<div class="sidebar-links">
<a class="sidebar-link active" href="#overview">Overview</a>
<a class="sidebar-link" href="#definition">Definition</a>
<a class="sidebar-link" href="#vlas-wms">VLAs and WMs</a>
<a class="sidebar-link" href="#architecture">Architecture</a>
<a class="sidebar-link" href="#training-data">Training Data</a>
<a class="sidebar-link" href="#evaluation">Evaluation</a>
<a class="sidebar-link" href="#challenges">Challenges</a>
<a class="sidebar-link" href="#library">Library</a>
</div>
</nav>
<header class="mobile-header">
<div class="mobile-topbar">
<a class="mobile-brand" href="#overview">WAM Survey</a>
<a class="btn btn-dark btn-sm" href="#library">Library</a>
</div>
<nav class="mobile-section-nav" aria-label="Mobile page sections">
<a href="#overview">Overview</a>
<a href="#definition">Definition</a>
<a href="#vlas-wms">VLAs and WMs</a>
<a href="#architecture">Architecture</a>
<a href="#training-data">Training Data</a>
<a href="#evaluation">Evaluation</a>
<a href="#challenges">Challenges</a>
<a href="#library">Library</a>
</nav>
</header>
<main class="content-main">
<header class="hero-section reveal" id="overview">
<div class="hero-copy">
<div class="paper-tags">
<span class="tag">Embodied AI</span>
<span class="tag">VLA</span>
<span class="tag">World Models</span>
<span class="tag">World Action Models</span>
</div>
<p class="eyebrow">survey of world action models</p>
<h1>World Action Models: The Next Frontier in Embodied AI</h1>
<div class="paper-authors" aria-label="Paper authors and affiliations">
<p>
<a href="https://sinwang20.github.io/">Siyin Wang</a><sup>1,2,*,‡</sup>,
<a href="https://openmoss.github.io/Awesome-WAM">Junhao Shi</a><sup>1,2,*</sup>,
<a href="https://openmoss.github.io/Awesome-WAM">Zhaoyang Fu</a><sup>1,*</sup>,
<a href="https://openmoss.github.io/Awesome-WAM">Xinzhe He</a><sup>1,*</sup>,
<a href="https://openmoss.github.io/Awesome-WAM">Feihong Liu</a><sup>1,*</sup>, <br>
<a href="https://openmoss.github.io/Awesome-WAM">Chenchen Yang</a><sup>1,2</sup>,
<a href="https://openmoss.github.io/Awesome-WAM">Yikang Zhou</a><sup>2</sup>,
<a href="https://openmoss.github.io/Awesome-WAM">Zhaoye Fei</a><sup>1</sup>,
<a href="https://scholar.google.com/citations?user=IlNreT4AAAAJ&hl=zh-TW">Jingjing Gong</a><sup>2</sup>,
<a href="https://jinlanfu.github.io/">Jinlan Fu</a><sup>1</sup>, <br>
<a href="https://sites.google.com/view/showlab">Mike Zheng Shou</a><sup>3</sup>,
<a href="https://xuanjing-huang.github.io/">Xuanjing Huang</a><sup>1,2</sup>,
<a href="https://xpqiu.github.io/">Xipeng Qiu</a><sup>1,2</sup>,
<a href="https://teai.fudan.edu.cn/">Yu-Gang Jiang</a><sup>1,†</sup>
</p>
<p>
<sup>1</sup>Fudan University
<sup>2</sup>Shanghai Innovation Institute
<sup>3</sup>National University of Singapore
</p>
<p><sup>*</sup>Equal Contribution, <sup>‡</sup>Project Lead, <sup>†</sup>Corresponding Author</p>
</div>
<div class="paper-actions">
<a class="btn btn-dark btn-icon" href="https://arxiv.org/abs/2605.12090" target="_blank" rel="noopener">
<svg viewBox="0 0 24 24" aria-hidden="true">
<path d="M7 3.75h7.5L19.25 8.5v11.75a1 1 0 0 1-1 1H7a1 1 0 0 1-1-1v-15.5a1 1 0 0 1 1-1Z" fill="none" stroke="currentColor" stroke-linejoin="round" stroke-width="1.8"/>
<path d="M14.5 3.75V8.5h4.75" fill="none" stroke="currentColor" stroke-linejoin="round" stroke-width="1.8"/>
<path d="M8.75 15.25h6.5M8.75 18.25h4.75" fill="none" stroke="currentColor" stroke-linecap="round" stroke-width="1.8"/>
</svg>
<span>Paper</span>
</a>
<a class="btn btn-dark btn-icon" href="https://github.com/OpenMOSS/Awesome-WAM" target="_blank" rel="noopener">
<svg viewBox="0 0 24 24" aria-hidden="true">
<path d="M12 3.75a8.75 8.75 0 0 0-2.77 17.05c.44.08.6-.19.6-.42v-1.66c-2.46.53-2.98-1.18-2.98-1.18-.4-1.03-.98-1.3-.98-1.3-.8-.54.06-.53.06-.53.89.06 1.36.92 1.36.92.79 1.35 2.07.96 2.58.73.08-.57.31-.96.56-1.18-1.96-.22-4.02-.98-4.02-4.37 0-.97.35-1.77.92-2.39-.09-.23-.4-1.13.09-2.36 0 0 .75-.24 2.46.91a8.5 8.5 0 0 1 4.48 0c1.71-1.15 2.46-.91 2.46-.91.49 1.23.18 2.13.09 2.36.57.62.92 1.42.92 2.39 0 3.4-2.07 4.15-4.05 4.36.32.27.6.8.6 1.61v2.39c0 .23.16.5.61.42A8.75 8.75 0 0 0 12 3.75Z" fill="currentColor"/>
</svg>
<span>Project</span>
</a>
<a class="btn btn-dark btn-icon" href="./leaderboard/">
<svg viewBox="0 0 24 24" aria-hidden="true">
<path d="M4.5 19.5h15" fill="none" stroke="currentColor" stroke-linecap="round" stroke-width="1.8"/>
<path d="M6.25 16.5V10.5M12 16.5V5.5M17.75 16.5V8" fill="none" stroke="currentColor" stroke-linecap="round" stroke-width="1.8"/>
<path d="M5.25 10.5l3-3 3 2.5 4.25-4.5 3.25 2.5" fill="none" stroke="currentColor" stroke-linecap="round" stroke-linejoin="round" stroke-width="1.8"/>
</svg>
<span>Leaderboard</span>
</a>
</div>
</div>
<p class="paper-subtitle">
This is the <strong>first systematic survey</strong> on <strong>World Action Models (WAMs)</strong> — embodied foundation models that unify
predictive world modeling with action generation. We formalize the definition of WAMs,
trace how Vision-Language-Action (VLA) models and world models converge, and organize existing methods into
Cascaded and Joint architectures. We also review the training data ecosystem, synthesize evaluation protocols,
and discuss open challenges, offering a comprehensive roadmap for this emerging field.
</p>
<figure class="architecture-figure section-stack section-figure overview-figure" aria-label="Overview roadmap">
<img
src="./figs/roadmap.svg"
alt="Temporal evolution and taxonomy of representative World Action Model methods"
loading="lazy"
/>
<figcaption>
Temporal evolution and taxonomy of representative World Action Models. The figure organizes
the field into Joint and Cascaded WAM branches and highlights the dominant architectural
directions explored across recent work.
</figcaption>
</figure>
</header>
<section class="paper-section reveal" id="definition">
<div class="section-heading">
<p class="eyebrow">definition</p>
<h2>Definitions and Formalism</h2>
<p class="section-note">
World Action Models are defined by forward predictive modeling and coupled action
generation. The key boundary is that future-state prediction must be part of the policy,
not just an auxiliary backbone or external simulator.
</p>
</div>
<figure class="architecture-figure section-stack section-figure definition-figure" aria-label="Definition diagram">
<img
src="./figs/definition.svg"
alt="Conceptual definition of VLA, WM, and WAM"
loading="lazy"
/>
<figcaption>
Conceptual definition and comparison of VLA, WM, and WAM. The figure contrasts their
input-output formulations and highlights that WAMs jointly predict future observations and
executable actions rather than modeling either side alone.
</figcaption>
</figure>
<div class="argument-stack section-stack">
<article class="argument-card">
<span class="argument-index">01</span>
<h3>Vision-Language-Action (VLA)</h3>
<p>
VLAs are embodied foundation models that frame robot control as a multimodal sequence
modeling task. They process current observation <code>o</code> and language instruction
<code>l</code> to generate actions under the objective <code>p(a | o, l)</code>.
</p>
<p>
In this formulation, observation and language map directly to action. Semantic grounding
is strong, but the model remains fundamentally reactive because future physical evolution
is not explicitly represented.
</p>
</article>
<article class="argument-card">
<span class="argument-index">02</span>
<h3>World Models (WM)</h3>
<p>
World models are predictive transition functions that internalize environment dynamics.
Their role is to model how a current state evolves under intervention, typically as
<code>p(o' | o, a)</code>, so they can simulate future observations rather than output a policy by themselves.
</p>
<p>
This makes them predictive rather than directly executable: they imagine what the world
will look like after action, but do not by themselves define a robot policy.
</p>
</article>
<article class="argument-card emphasis">
<span class="argument-index">03</span>
<h3>World Action Models (WAMs)</h3>
<p>
WAMs unify environmental dynamics modeling with motor control. A model qualifies as a
WAM only when it performs forward predictive modeling of future state and couples action
generation to that anticipated future, targeting <code>p(o', a | o, l)</code>.
</p>
<p>
The defining shift is that future-state synthesis and executable action are learned
together inside one embodied policy framework rather than being treated as unrelated outputs.
</p>
</article>
</div>
</section>
<section class="paper-section reveal figures-section" id="vlas-wms">
<div class="section-heading">
<p class="eyebrow">vlas and wms</p>
<h2>VLAs and World Models: Foundations and Early Integration</h2>
<p class="section-note">
WAMs emerge from the convergence of Vision-Language-Action policies and predictive world
models. The survey positions VLA, world-modeling, and WM-for-VLA integration as the
background that made predictive embodied policies possible.
</p>
</div>
<div class="storyline" aria-label="Key survey figures">
<article class="story-step story-step-textonly">
<div class="story-copy">
<span class="story-number">01</span>
<p class="figure-kicker">Vision-Language-Action Models</p>
<h3>VLAs scale robot control through language-conditioned foundation policies</h3>
<p>
VLAs evolved from task-specific imitation learning into language-conditioned policies
that fuse visual observations and task prompts. Modern systems inherit LVLM priors,
then generate actions through autoregressive tokenization or diffusion-style action heads.
</p>
<dl class="figure-logic">
<div>
<dt>Key point</dt>
<dd>VLA models bring internet-scale semantic understanding into robot control and support open-vocabulary, long-horizon manipulation.</dd>
</div>
<div>
<dt>Limitation</dt>
<dd>Even with 3D, depth, force, and tactile inputs, most VLAs remain reactive image-to-action mappings without explicit world dynamics.</dd>
</div>
</dl>
</div>
</article>
<article class="story-step story-step-textonly">
<div class="story-copy">
<span class="story-number">02</span>
<p class="figure-kicker">World Models</p>
<h3>World models provide predictive structure over future state evolution</h3>
<p>
World models learn internal representations of environment dynamics and predict the
consequences of actions, language instructions, or multimodal context. The survey
distinguishes action-conditioned, language-conditioned, and embodied world models.
</p>
<dl class="figure-logic">
<div>
<dt>Role</dt>
<dd>World models enable simulation, planning, and decision-making by forecasting future states before real execution.</dd>
</div>
<div>
<dt>Design</dt>
<dd>Explicit models predict pixels or videos, while implicit models learn compact latent dynamics for efficiency and abstraction.</dd>
</div>
</dl>
</div>
</article>
<article class="story-step story-step-stack">
<div class="story-copy">
<span class="story-number">03</span>
<p class="figure-kicker">World Models for VLA</p>
<h3>World models extend VLAs beyond direct policy learning</h3>
<p>
World models let VLA agents imagine future observations, generate trajectories, estimate
outcomes, and test policies before physical execution. The survey frames their contribution
through two routes: improving learning and enabling scalable evaluation.
</p>
<dl class="figure-logic">
<div>
<dt>Learning</dt>
<dd>World models augment imitation data, support model-based reinforcement learning, and derive reward signals from predicted futures.</dd>
</div>
<div>
<dt>Evaluation</dt>
<dd>World models act as data-driven simulators for reproducible, safety-aware rollout testing with less real-world deployment cost.</dd>
</div>
</dl>
</div>
<figure class="story-figure story-figure-pdf" aria-label="WM for VLA diagram">
<img
src="./figs/wm4vla.svg"
alt="World models for VLA overview diagram"
loading="lazy"
/>
</figure>
</article>
</div>
</section>
<section class="paper-section reveal" id="architecture">
<div class="section-heading">
<p class="eyebrow">architecture</p>
<h2>Architecture</h2>
<p class="section-note">
The decisive design question is how world prediction and action generation are structurally
coupled. The survey organizes methods into Cascaded WAMs and Joint WAMs, each with
different training regimes, representations, and latency trade-offs.
</p>
</div>
<div class="architecture-grid">
<article class="paradigm-card cascaded-card">
<div class="paradigm-topline">
<h3>Cascaded WAM</h3>
<code>future plan -> action</code>
</div>
<p>
A world model first synthesizes the anticipated future state, then a separate action model
decodes executable commands from that plan. This factorization gives clear modular structure,
but makes coupling quality between the two stages the central bottleneck.
</p>
<div class="stage-list">
<div><strong>Explicit planning</strong><span>pixel-space carriers such as video, flow, depth, normals, or 4D structure</span></div>
<div><strong>Implicit planning</strong><span>latent features, future tokens, masks, and hidden-state conditioning</span></div>
<!-- <div><strong>Trade-off</strong><span>Natural modularity and interpretable future plans, but two-stage coupling, generation cost, and stage mismatch can limit closed-loop control.</span></div> -->
</div>
</article>
<article class="paradigm-card joint-card">
<div class="paradigm-topline">
<h3>Joint WAM</h3>
<code>future + action</code>
</div>
<p>
Future states and actions are predicted within one shared model and trained under joint
supervision. The main question becomes how coupling is realized across discrete tokenization,
autoregressive sequence modeling, diffusion, flow matching, or parallel generation.
</p>
<div class="stage-list">
<div><strong>Autoregressive generation</strong><span>shared token spaces, multi-head routing, unified vocabularies</span></div>
<div><strong>Diffusion-based generation</strong><span>single-engine or multi-engine diffusion and non-autoregressive generation</span></div>
<!-- <div><strong>Trade-off</strong><span>Unified training strengthens world-action coupling, but sequential error propagation, latency, and unclear coupling mechanisms remain open issues.</span></div> -->
</div>
</article>
</div>
<div class="architecture-figures section-stack" aria-label="Architecture figures">
<figure class="architecture-figure">
<img
src="./figs/cascaded.svg"
alt="Cascaded World Action Model architecture diagram"
loading="lazy"
/>
<figcaption>
Cascaded WAM architectures first predict future state representations and then derive executable actions from that predicted plan.
</figcaption>
</figure>
<figure class="architecture-figure">
<img
src="./figs/joint.svg"
alt="Joint World Action Model architecture diagram"
loading="lazy"
/>
<figcaption>
Joint WAM architectures model future world state and action generation inside one shared predictive system with tighter coupling.
</figcaption>
</figure>
</div>
</section>
<section class="paper-section reveal" id="training-data">
<div class="section-heading">
<p class="eyebrow">training data</p>
<h2>Training Data</h2>
<p class="section-note">
WAM data is a mixture-design problem, not just scale. The survey groups
data into robot-centric teleoperation, portable human demonstration, simulation,
and human or egocentric video.
</p>
</div>
<div class="training-data-layout section-stack">
<figure class="data-figure" aria-label="Training data overview">
<img
src="./figs/data.svg"
alt="Overview of embodied data sources for training World Action Models"
loading="lazy"
/>
</figure>
<div class="data-matrix">
<article class="data-row">
<span>Robot-Centric Teleoperation</span>
<h3>Aligned state-action supervision</h3>
<p>High-frequency robot trajectories give aligned state-action pairs, kinematics, proprioception, and contact-rich control with low sim-to-real gap.</p>
</article>
<article class="data-row">
<span>Portable Human Demonstration</span>
<h3>Low-cost real-world diversity</h3>
<p>UMI-style collection brings retargeted human demonstrations from scenes, adding low-cost diversity, dexterous motion, and multi-view interaction.</p>
</article>
<article class="data-row">
<span>Simulation</span>
<h3>Privileged physics supervision</h3>
<p>Simulators provide scalable variation with privileged depth, pose, collision, 3D structure, and controllable physics for dynamics learning.</p>
</article>
<article class="data-row">
<span>Human and Ego-Centric Data</span>
<h3>Massive-scale world priors</h3>
<p>Web and egocentric video supply passive dynamics, long-horizon activity, semantic diversity, and open-world priors beyond robot labs.</p>
</article>
</div>
</div>
</section>
<section class="paper-section reveal" id="evaluation">
<div class="section-heading">
<p class="eyebrow">evaluation</p>
<h2>Evaluation</h2>
<p class="section-note">
Evaluation has two axes: world modeling capability and action policy capability. WAMs
cannot be judged by visual plausibility alone, and they also cannot be judged by task success
alone without testing whether imagined futures are physically and causally meaningful.
</p>
</div>
<div class="eval-sections">
<article class="eval-section-card">
<div class="eval-section-head">
<p class="eyebrow">world modeling capability</p>
<h3>How to Evaluate World Modeling Capability?</h3>
<p>
The survey treats world-model evaluation as a three-part question: whether generated
futures are visually faithful, physically plausible, and still informative enough to recover executable control.
</p>
</div>
<div class="eval-subgrid">
<article class="eval-card">
<span>Visual fidelity</span>
<h3>Reconstruction, perception, and realism</h3>
<p>PSNR and SSIM cover low-level fidelity; LPIPS, DreamSim, and DINO test perceptual and semantic consistency; FVD measures distribution-level realism and temporal quality.</p>
</article>
<article class="eval-card">
<span>Physical commonsense</span>
<h3>Object dynamics and trajectory plausibility</h3>
<p>Physical commonsense asks whether generated worlds obey object continuity, material behavior, causal order, and plausible motion. Benchmarks such as VideoPhy, PhyGenBench, VBench-2.0, WorldModelBench, Physics-IQ, WorldScore, and EWMBench cover both physical interactions and long-horizon motion consistency.</p>
</article>
<article class="eval-card">
<span>Action plausibility</span>
<h3>Can imagined futures support control?</h3>
<p>WorldSimBench and the IDM Turing Test ask whether generated futures preserve enough action-relevant information to infer correct controls and support downstream execution.</p>
</article>
</div>
</article>
<article class="eval-section-card">
<div class="eval-section-head">
<p class="eyebrow">action policy capability</p>
<h3>How to Evaluate Action Policy?</h3>
<p>
The survey reviews benchmark families by robot morphology and manipulation setting,
emphasizing generalization, long-horizon control, dexterity, sim-to-real transfer, and real-device performance.
</p>
</div>
<div class="eval-subgrid eval-subgrid-policy">
<article class="eval-card">
<span>General manipulation</span>
<h3>Multi-task and broad policy capability</h3>
<p>Meta-World, RLBench, ManiSkill, LIBERO, RoboCasa, GemBench, CALVIN, and related suites test multi-task learning, scaling, language conditioning, robustness, and long-horizon execution.</p>
</article>
<article class="eval-card">
<span>Bimanual and humanoid form</span>
<h3>Higher-DoF coordinated control</h3>
<p>RoboTwin, BiGym, HumanoidBench, and HumanoidGen raise the difficulty through dual-arm coordination, humanoid locomotion, tactile sensing, and large action spaces.</p>
</article>
<article class="eval-card">
<span>Mobile manipulation</span>
<h3>Navigation plus manipulation</h3>
<p>ManipulaTHOR, HomeRobot, and BEHAVIOR-1K evaluate policies that must combine scene navigation, open-vocabulary perception, and object interaction inside larger environments.</p>
</article>
<article class="eval-card">
<span>Contact and deformation</span>
<h3>Physics beyond rigid-body control</h3>
<p>SoftGym, PlasticineLab, DaXBench, TacSL, and ManiFeel assess cloth, liquid, deformable objects, and tactile-driven fine manipulation where contact modeling becomes central.</p>
</article>
<article class="eval-card">
<span>Real-device evaluation</span>
<h3>Deployment in physical environments</h3>
<p>RoboArena, RoboChallenge, and Maniparena measure whether policy performance survives the reality gap and remains reliable when moved onto real robots.</p>
</article>
</div>
</article>
</div>
</section>
<section class="paper-section reveal" id="challenges">
<div class="section-heading">
<p class="eyebrow">open challenges</p>
<h2>Open Challenges and Opportunities</h2>
<p class="section-note">
WAMs are still a nascent paradigm. The next phase depends on resolving coupling, modality,
data-mixture, long-horizon, efficiency, evaluation, and safety problems.
</p>
</div>
<div class="challenge-grid">
<article><span>01</span><h3>Architectural Coupling</h3><p>Systematic comparisons under matched scale, data, and protocols are still missing.</p></article>
<article><span>02</span><h3>Multimodal Physical State</h3><p>RGB-only prediction misses tactile, force, acoustic, and deformation cues that matter most in contact-rich manipulation.</p></article>
<article><span>03</span><h3>Data Utilization and Mixture Design</h3><p>The marginal role of robot data, simulation, and egocentric human video is still poorly understood.</p></article>
<article><span>04</span><h3>Long-Horizon Planning</h3><p>Distribution drift, compounding action error, and weak temporal abstraction still block sustained predictive control.</p></article>
<article><span>05</span><h3>Inference Latency and Efficiency</h3><p>Diffusion and autoregressive prediction remain too slow for many closed-loop settings without aggressive compression.</p></article>
<article><span>06</span><h3>Evaluation and Safe Deployment</h3><p>The field still lacks joint metrics for causal consistency between imagined futures and executed actions, plus robust safety checks.</p></article>
</div>
</section>
<section class="paper-section" id="library">
<div class="section-heading reveal">
<p class="eyebrow">library</p>
<h2>Paper Library</h2>
<p class="section-note">
This library is populated from the survey repository and keeps the paper list searchable.
For quick browsing, the top-level filters focus only on the two architectural families
emphasized by the paper: Cascaded WAM and Joint WAM.
</p>
</div>
<div class="explorer-panel reveal">
<div class="library-tools">
<label class="search-box" for="paperSearch">
<span>Search</span>
<input
id="paperSearch"
type="search"
placeholder="DreamZero, JEPA, diffusion, 2602..."
autocomplete="off"
/>
</label>
<label class="sort-box" for="paperSort">
<span>Sort</span>
<select id="paperSort">
<option value="newest">Newest first</option>
<option value="oldest">Oldest first</option>
<option value="name-asc">Name A-Z</option>
</select>
</label>
</div>
<div class="filter-tabs" role="group" aria-label="Paper category filters">
<button class="filter-tab active" type="button" data-filter="all" aria-pressed="true">
<span>All</span><span class="filter-count" data-count-for="all">0</span>
</button>
<button class="filter-tab" type="button" data-filter="cascaded" aria-pressed="false">
<span>Cascaded WAM</span><span class="filter-count" data-count-for="cascaded">0</span>
</button>
<button class="filter-tab" type="button" data-filter="joint" aria-pressed="false">
<span>Joint WAM</span><span class="filter-count" data-count-for="joint">0</span>
</button>
</div>
</div>
<div class="explorer-status reveal" aria-live="polite">
<span id="resultCount">Loading related papers...</span>
<button class="text-button" id="resetFilters" type="button">Reset filters</button>
</div>
<div class="paper-grid" id="paperGrid"></div>
<div class="load-more-wrap">
<button class="btn btn-outline" id="loadMore" type="button" hidden>Load more</button>
</div>
</section>
</main>
<footer class="site-footer">
<div class="footer-inner">
<p>World Action Models: The Next Frontier in Embodied AI</p>
<p>OpenMOSS</p>
</div>
</footer>
<noscript>
<div class="noscript-banner">
Paper-library filtering requires JavaScript; the survey PDF and bibliography remain directly accessible.
</div>
</noscript>
<script type="module" src="./script.js"></script>
</body>
</html>