-
Notifications
You must be signed in to change notification settings - Fork 1
Expand file tree
/
Copy pathindex.html
More file actions
616 lines (576 loc) · 23.4 KB
/
index.html
File metadata and controls
616 lines (576 loc) · 23.4 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
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="description"
content="DiT4DiT: Jointly Modeling Video Dynamics and Actions for Generalizable Robot Control">
<meta name="keywords"
content="Video Generation, Diffusion Transformer, Robot Control, Vision-Language-Action, Flow Matching, Robotic Manipulation">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>DiT4DiT | Jointly Modeling Video Dynamics and Actions for Generalizable Robot Control</title>
<!-- Thumbnail for social media sharing -->
<meta property="og:image" content="media/figures/pipeline.jpg">
<!-- Favicon -->
<link rel="icon" href="media/figures/favicon.png" type="image/png">
<!-- Fonts: Geist (Google) + Computer Modern Serif (local) -->
<link href="https://fonts.googleapis.com/css2?family=Geist:wght@300;400;500;600;700;800&display=swap" rel="stylesheet">
<link rel="stylesheet" href="static/fonts/cmun.css">
<!-- Keep Bulma for table utility classes -->
<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">
<!-- DreamDojo-style overrides (loaded last) -->
<link rel="stylesheet" href="static/css/index.css">
<link rel="stylesheet" href="static/fontawesome/css/fontawesome.css">
<link rel="stylesheet" href="static/fontawesome/css/brands.css">
<link rel="stylesheet" href="static/fontawesome/css/light.css">
<script src="https://ajax.googleapis.com/ajax/libs/jquery/3.5.1/jquery.min.js"></script>
<script defer src="static/js/fontawesome.all.min.js"></script>
<!-- MathJax for LaTeX rendering -->
<script>
MathJax = {
tex: { inlineMath: [['$', '$'], ['\\(', '\\)']] }
};
</script>
<script src="https://cdn.jsdelivr.net/npm/mathjax@3/es5/tex-chtml.js" async></script>
</head>
<body>
<article class="blogPost">
<!-- ===== Title ===== -->
<div class="center blogTitle">
<h1 class="dreamTitle"><span class="projectName">DiT4DiT</span></h1>
<p class="subtitleText">Jointly Modeling Video Dynamics and Actions for Generalizable Robot Control</p>
</div>
<!-- ===== Authors ===== -->
<div class="center authors">
<div class="authorList">
<span class="authorName"><a class="authorLink" href="https://teleema.github.io/" target="_blank">Teli Ma</a><sup>1,2</sup></span>
<span class="authorName"><a class="authorLink" href="https://jiaazheng.github.io/" target="_blank">Jia Zheng</a><sup>1,2</sup></span>
<span class="authorName"><a class="authorLink" href="https://scholar.google.com/citations?user=GaJXZ-UAAAAJ&hl=en" target="_blank">Zifan Wang</a><sup>1,2</sup></span>
<span class="authorName"><a class="authorLink" href="https://scholar.google.com/citations?user=nvzF-RMAAAAJ&hl=en" target="_blank">Chunli Jiang</a><sup>1</sup></span>
<span class="authorName"><a class="authorLink">Andy Cui</a><sup>1</sup></span>
<span class="authorName"><a class="authorLink" href="https://junweiliang.me/index.html" target="_blank">Junwei Liang</a><sup>2,3,*</sup></span>
<span class="authorName"><a class="authorLink" href="https://shuoyangrobotics.github.io/" target="_blank">Shuo Yang</a><sup>1,*</sup></span>
</div>
<!-- Affiliations -->
<div class="affiliationContainer">
<span><sup>1</sup>Mondo Robotics</span>
<span><sup>2</sup>HKUST(GZ)</span>
<span><sup>3</sup>HKUST</span>
</div>
<p class="affiliationNote"><sup>*</sup>Corresponding author, Co-advising</p>
</div>
<!-- ===== Action buttons ===== -->
<div class="center">
<div class="linkContainer">
<a class="borderedLink" href="https://my.feishu.cn/wiki/G8BMwxbfmiBLlYkc58KcIF76nxb?from=from_copylink" target="_blank"><i class="fa-light fa-file-pdf"></i> Paper</a>
<a class="borderedLink" href="https://arxiv.org/abs/2603.10448" target="_blank"><i class="ai ai-arxiv"></i> arXiv</a>
<a class="borderedLink" href="https://github.com/Mondo-Robotics/DiT4DiT" target="_blank"><i class="fa-light fa-code"></i> Code </a>
<a class="borderedLink" href="https://x.com/ShuoYangAIR/status/2031992572227555526" target="_blank"><i class="fa-brands fa-x-twitter"></i> Thread</a>
</div>
</div>
<!-- ===== Content ===== -->
<div class="blogContent">
<h2>Whole-body Control</h2>
<div class="videoContainer">
<div class="videoGrid cols-1">
<div class="videoCell">
<video controls autoplay loop muted playsinline>
<source src="media/videos/shelf_organization.mp4" type="video/mp4">
</video>
<p class="videoLabel">Shelf Organization (1x speed)</p>
</div>
<div class="videoCell">
<video controls autoplay loop muted playsinline>
<source src="media/videos/relocate_chair.mp4" type="video/mp4">
</video>
<p class="videoLabel">Relocate Chair (1x speed)</p>
</div>
</div>
</div>
<h2>Tabletop Tasks</h2>
<!-- Video demos: Row 1 (2 videos) -->
<div class="videoContainer">
<div class="videoGrid cols-2">
<div class="videoCell">
<video controls autoplay loop muted playsinline>
<source src="media/videos/packing.mp4" type="video/mp4">
</video>
<p class="videoLabel">Box Packing (1x speed)</p>
</div>
<div class="videoCell">
<video controls autoplay loop muted playsinline>
<source src="media/videos/drawer.mp4" type="video/mp4">
</video>
<p class="videoLabel">Drawer Interaction (1x speed)</p>
</div>
</div>
</div>
<!-- Video demos: Row 2 (2 videos) -->
<div class="videoContainer">
<div class="videoGrid cols-2">
<div class="videoCell">
<video controls autoplay loop muted playsinline>
<source src="media/videos/cups.mp4" type="video/mp4">
</video>
<p class="videoLabel">Stack up the Cups (1x speed)</p>
</div>
<div class="videoCell">
<video controls autoplay loop muted playsinline>
<source src="media/videos/plate.mp4" type="video/mp4">
</video>
<p class="videoLabel">Insert Plate into the Rack (1x speed)</p>
</div>
</div>
</div>
<!-- Video demos: Row 3 (2 videos) -->
<div class="videoContainer">
<div class="videoGrid cols-2">
<div class="videoCell">
<video controls autoplay loop muted playsinline>
<source src="media/videos/eggplant.mp4" type="video/mp4">
</video>
<p class="videoLabel">Pick and Place (1x speed)</p>
</div>
<div class="videoCell">
<video controls autoplay loop muted playsinline>
<source src="media/videos/spoon.mp4" type="video/mp4">
</video>
<p class="videoLabel">Move the Spoon (1x speed)</p>
</div>
</div>
</div>
<!-- Video demos: Row 4 (2 videos side by side) -->
<div class="videoContainer">
<div class="videoGrid cols-2">
<div class="videoCell">
<video controls autoplay loop muted playsinline>
<source src="media/videos/flower.mp4" type="video/mp4">
</video>
<p class="videoLabel">Arrange the Flower (1x speed)</p>
</div>
<div class="videoCell">
<video controls autoplay loop muted playsinline>
<source src="media/videos/twist_cap.mp4" type="video/mp4">
</video>
<p class="videoLabel">Twist Cap (1x speed)</p>
</div>
</div>
</div>
<!-- Abstract -->
<h2>Highlights</h2>
<ol>
<li><b>A cascaded video-action architecture.</b> An end-to-end framework that unifies video and action Diffusion Transformers, leveraging generative video dynamics as a rich physical proxy to replace static vision-language priors.</li>
<li><b>An joint dual flow-matching objective.</b> A joint dual flow-matching objective that decouples diffusion timesteps, enabling the extraction of highly structured, physics-aware features in a single deterministic step for efficient closed-loop control.</li>
<li><b>High data efficiency and zero-shot generalization.</b> Achieving state-of-the-art performance across simulation and physical Unitree G1 deployments using a single ego-camera and merely 15% of the pre-training data of comparable baselines, with robust adaptation to unseen objects.</li>
</ol>
<!-- Overview of DiT4DiT -->
<h2>Method</h2>
<div class="imageWrapper">
<img src="media/figures/pipeline.jpg" style="width: 90%;" alt="DiT4DiT pipeline overview">
</div>
<p>
<b>DiT4DiT is an end-to-end Video-Action Model with a cascaded dual-DiT architecture.</b>
<b>(a)</b> The <b>Video DiT</b> (initialized from Cosmos-Predict2.5-2B) takes the current observation frames
and a language goal, encodes them via a causal video VAE into latent space, and models future visual dynamics
via flow matching. A forward hook mechanism intercepts intermediate hidden activations at a fixed flow
timestep from a specific transformer layer, converting the generative process into rich, physics-aware
visual tokens — without requiring full video reconstruction.
<b>(b)</b> The <b>Action DiT</b> takes the extracted visual tokens via cross-attention, along with
proprioceptive state embeddings and noisy action trajectories, and predicts a velocity field to generate
precise robot action trajectories.
<b>(c)</b> A <b>dual flow-matching objective</b> with a tri-timestep scheme jointly optimizes both branches
end-to-end, allowing the video branch to learn the full generative trajectory while the action branch
performs efficient generative inverse dynamics.
</p>
<div class="imageWrapper">
<img src="media/figures/tri_timestep.png" style="width: 90%;" alt="Tri-timestep mechanism">
</div>
<p>
<b>Asymmetric tri-timestep design.</b> We decouple the diffusion timesteps to optimize joint video-action generation. The video module uses uniform sampling $\tau_v$ to capture the full denoising trajectory, while the action module uses Beta sampling $\tau_a$ to focus on critical control phases. Meanwhile, stable visual conditions are extracted at a fixed deterministic timestep $\tau_f$ from the evolving hidden states $h_t^1 \rightarrow h_t^0$.
</p>
<!-- Why Video Generation? -->
<!-- <h2>Why Video Generation?</h2>
<p>
We provide empirical evidence that <b>video generation is a vastly superior, data-efficient unsupervised
training objective</b> compared to semantic grounding or VLM-centric latent modeling.
Our study compares three proxy objectives: semantic grounding, FLARE-style latent prediction, and video
generation.
Video generation achieves <b>up to 10× higher data efficiency</b> and <b>7× faster
convergence</b>.
</p>
<div class="imageWrapper">
<img src="media/figures/intro.jpg" style="width: 85%;" alt="Video generation vs other objectives">
</div> -->
<!-- Simulation Results -->
<h2>Simulation Results</h2>
<!-- <p> -->
<!-- DiT4DiT achieves <b>state-of-the-art performance</b> on two challenging simulation benchmarks.
</p> -->
<h3>LIBERO Benchmark</h3>
<p>
DiT4DiT achieves <b>98.6% average success rate</b> across four LIBERO suites, surpassing all baselines
including π<sub>0.5</sub> (96.9%), CogVLA (97.4%), and OpenVLA-OFT (97.1%). Particularly strong on
LIBERO-Long (extended horizon): <b>97.6%</b>.
</p>
<!-- <div class="table-container">
<table class="table is-bordered is-striped is-hoverable is-fullwidth">
<thead>
<tr>
<th>Method</th>
<th>LIBERO-Spatial</th>
<th>LIBERO-Object</th>
<th>LIBERO-Goal</th>
<th>LIBERO-Long</th>
<th>Average</th>
</tr>
</thead>
<tbody>
<tr>
<td>OpenVLA</td>
<td>84.7</td>
<td>88.4</td>
<td>79.2</td>
<td>53.7</td>
<td>76.5</td>
</tr>
<tr>
<td>CogACT</td>
<td>92.5</td>
<td>92.7</td>
<td>91.3</td>
<td>74.7</td>
<td>87.8</td>
</tr>
<tr>
<td>OpenVLA-OFT</td>
<td>97.8</td>
<td>97.8</td>
<td>98.0</td>
<td>94.8</td>
<td>97.1</td>
</tr>
<tr>
<td>π<sub>0.5</sub></td>
<td>98.0</td>
<td>98.4</td>
<td>96.0</td>
<td>95.2</td>
<td>96.9</td>
</tr>
<tr>
<td>CogVLA</td>
<td>98.4</td>
<td>99.6</td>
<td>97.6</td>
<td>93.6</td>
<td>97.4</td>
</tr>
<tr>
<td>Qwen3DiT</td>
<td>97.0</td>
<td>97.2</td>
<td>97.8</td>
<td>94.2</td>
<td>96.6</td>
</tr>
<tr style="font-weight: bold; background-color: #fdf6e3;">
<td>DiT4DiT (Ours)</td>
<td>99.2</td>
<td>99.2</td>
<td>98.4</td>
<td>97.6</td>
<td>98.6</td>
</tr>
</tbody>
</table>
</div> -->
<h3>RoboCasa-GR1 Benchmark</h3>
<p>
DiT4DiT achieves <b>50.8% average success rate</b> across 24 tasks, surpassing GR00T-N1.5 (41.8%) by 9.0
points and GR00T-N1.6 (40.8%) by 10.0 points. Outperforms Qwen3DiT (36.2%) by <b>14.6% absolute</b>,
confirming that video-generative priors are vastly superior to static VLM priors.
</p>
<!-- <div class="table-container">
<table class="table is-bordered is-striped is-hoverable is-fullwidth">
<thead>
<tr>
<th>Task</th>
<th>GR00T-N1.5</th>
<th>GR00T-N1.6</th>
<th>Qwen3DiT</th>
<th style="background-color: #fdf6e3;">DiT4DiT</th>
</tr>
</thead>
<tbody>
<tr>
<td>BottleToCabinetClose</td>
<td>40.0</td>
<td>36.0</td>
<td><b>50.0</b></td>
<td>48.0</td>
</tr>
<tr>
<td>CanToDrawerClose</td>
<td>56.0</td>
<td>28.0</td>
<td>48.0</td>
<td><b>74.0</b></td>
</tr>
<tr>
<td>CupToDrawerClose</td>
<td>50.0</td>
<td>12.0</td>
<td>42.0</td>
<td><b>52.0</b></td>
</tr>
<tr>
<td>MilkToMicrowaveClose</td>
<td><b>52.0</b></td>
<td>20.0</td>
<td>38.0</td>
<td>50.0</td>
</tr>
<tr>
<td>PotatoToMicrowaveClose</td>
<td>22.0</td>
<td>28.0</td>
<td>18.0</td>
<td><b>36.0</b></td>
</tr>
<tr>
<td>WineToCabinetClose</td>
<td><b>44.0</b></td>
<td>18.0</td>
<td>28.0</td>
<td>42.0</td>
</tr>
<tr>
<td>FromCuttingboardToBasket</td>
<td>46.0</td>
<td>42.0</td>
<td>42.0</td>
<td><b>52.0</b></td>
</tr>
<tr>
<td>FromCuttingboardToCardboardbox</td>
<td>44.0</td>
<td>40.0</td>
<td>30.0</td>
<td><b>48.0</b></td>
</tr>
<tr>
<td>FromCuttingboardToPan</td>
<td>58.0</td>
<td>62.0</td>
<td>50.0</td>
<td><b>76.0</b></td>
</tr>
<tr>
<td>FromCuttingboardToPot</td>
<td>48.0</td>
<td>60.0</td>
<td>44.0</td>
<td><b>62.0</b></td>
</tr>
<tr>
<td>FromCuttingboardToTieredbasket</td>
<td>28.0</td>
<td>48.0</td>
<td>36.0</td>
<td><b>50.0</b></td>
</tr>
<tr>
<td>FromPlacematToBasket</td>
<td>32.0</td>
<td>42.0</td>
<td>14.0</td>
<td><b>50.0</b></td>
</tr>
<tr>
<td>FromPlacematToBowl</td>
<td>52.0</td>
<td>34.0</td>
<td>28.0</td>
<td><b>56.0</b></td>
</tr>
<tr>
<td>FromPlacematToPlate</td>
<td><b>42.0</b></td>
<td><b>42.0</b></td>
<td>40.0</td>
<td>32.0</td>
</tr>
<tr>
<td>FromPlacematToTieredshelf</td>
<td>26.0</td>
<td>24.0</td>
<td><b>30.0</b></td>
<td>18.0</td>
</tr>
<tr>
<td>FromPlateToBowl</td>
<td>38.0</td>
<td>48.0</td>
<td>36.0</td>
<td><b>56.0</b></td>
</tr>
<tr>
<td>FromPlateToCardboardbox</td>
<td>40.0</td>
<td>44.0</td>
<td>36.0</td>
<td><b>58.0</b></td>
</tr>
<tr>
<td>FromPlateToPan</td>
<td>56.0</td>
<td>48.0</td>
<td>34.0</td>
<td><b>68.0</b></td>
</tr>
<tr>
<td>FromPlateToPlate</td>
<td>50.0</td>
<td><b>66.0</b></td>
<td>44.0</td>
<td>58.0</td>
</tr>
<tr>
<td>FromTrayToCardboardbox</td>
<td>36.0</td>
<td>42.0</td>
<td><b>48.0</b></td>
<td>38.0</td>
</tr>
<tr>
<td>FromTrayToPlate</td>
<td>54.0</td>
<td>52.0</td>
<td>44.0</td>
<td><b>56.0</b></td>
</tr>
<tr>
<td>FromTrayToPot</td>
<td>36.0</td>
<td><b>64.0</b></td>
<td>34.0</td>
<td>54.0</td>
</tr>
<tr>
<td>FromTrayToTieredbasket</td>
<td>34.0</td>
<td>42.0</td>
<td>36.0</td>
<td><b>46.0</b></td>
</tr>
<tr>
<td>FromTrayToTieredshelf</td>
<td>22.0</td>
<td><b>38.0</b></td>
<td>18.0</td>
<td><b>38.0</b></td>
</tr>
<tr style="font-weight: bold; background-color: #fdf6e3;">
<td>Average</td>
<td>41.8</td>
<td>40.8</td>
<td>36.2</td>
<td>50.8</td>
</tr>
</tbody>
</table>
</div> -->
<!-- Real-World Results -->
<h2>Real-World Results</h2>
<p>
We evaluate DiT4DiT on seven household tasks with the <b>Unitree G1 humanoid robot</b>:
<i>pick-and-place</i>, <i>arrange flower</i>, <i>stack cups</i>, <i>insert plate</i>, <i>box packaging</i>, <i>move spoon</i>, and <i>drawer interaction</i>.
DiT4DiT comprehensively outperforms both GR00T-N1.5 (pre-trained with more data) and Qwen3DiT across all
tasks.
</p>
<div class="imageWrapper">
<img src="media/figures/real_task_res.jpg" style="width: 85%;" alt="Real-world task results">
</div>
<!-- Zero-Shot Generalization -->
<h2>Generalization</h2>
<p>
DiT4DiT demonstrates <b>strong zero-shot generalization</b> under severe distribution shifts in both
simulation and real-world settings.
</p>
<div class="imageWrapper">
<img src="media/figures/task_gene.jpg" style="width: 85%;" alt="Zero-shot generalization tasks">
</div>
<h3>Simulation Generalization</h3>
<p>
In RoboCasa object-substitution tests, DiT4DiT achieves <b>54.5%</b> on unseen objects vs. 32.0% for
Qwen3DiT.
</p>
<h3>Real-World Generalization</h3>
<p>
Tested on category changes (unseen cups/vases), object substitution (corn instead of eggplant), and
quantity variation (4 cups instead of 3). DiT4DiT achieves <b>70%</b> on Arrange Flower (Category) vs. 0%
for Qwen3DiT and 10% for GR00T-N1.5.
</p>
<div class="imageWrapper">
<img src="media/figures/real_task_gene_exp.png" style="width: 100%;" alt="Real-world generalization experiments">
</div>
<!-- Generated Video Plans -->
<h2>Generated Video Plans</h2>
<p>
DiT4DiT can optionally generate full video plans showing predicted future dynamics. The video branch
produces realistic visual plans that demonstrate the model's understanding of physical dynamics and
task-relevant behaviors.
</p>
<div class="imageWrapper">
<img src="media/figures/video_gen.jpg" style="width: 85%;" alt="Generated video plans">
</div>
<!-- Model Efficiency -->
<h2>Efficiency</h2>
<div class="table-container">
<table class="table is-bordered is-striped is-hoverable" style="margin: 0 auto; width: auto;">
<thead>
<tr>
<th>Model</th>
<th>Trainable Params</th>
<th>Deployment Freq.</th>
</tr>
</thead>
<tbody>
<tr>
<td>GR00T-N1.5</td>
<td>2.7B</td>
<td>13 Hz</td>
</tr>
<tr>
<td>Qwen3DiT</td>
<td>2.3B</td>
<td>9 Hz</td>
</tr>
<tr style="font-weight: bold; background-color: #fdf6e3;">
<td>DiT4DiT (Ours)</td>
<td>2.2B</td>
<td>6 Hz</td>
</tr>
</tbody>
</table>
</div>
<p>
DiT4DiT is the most parameter-efficient model. While the 6 Hz deployment frequency is lower than
alternatives, it is sufficient for real-time closed-loop control, and the superior action quality
more than compensates for the lower frequency.
</p>
<!-- Citation -->
<h2>BibTeX</h2>
<!-- <p>If you find our work useful, please consider citing:</p> -->
<pre class="bibtex">@article{ma2026dit4dit,
title={DiT4DiT: Jointly Modeling Video Dynamics and Actions for Generalizable Robot Control},
author={Ma, Teli and Zheng, Jia and Wang, Zifan and Jiang, Chunli and Cui, Andy and Liang, Junwei and Yang, Shuo},
journal={arXiv preprint arXiv:2603.10448},
year={2026}
}</pre>
</div><!-- .blogContent -->
</article>
<footer class="siteFooter"></footer>
</body>
</html>