-
Notifications
You must be signed in to change notification settings - Fork 1
Expand file tree
/
Copy pathonevl.html
More file actions
1060 lines (930 loc) · 47.1 KB
/
Copy pathonevl.html
File metadata and controls
1060 lines (930 loc) · 47.1 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
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
<!doctype html>
<html lang="en">
<head>
<title>OneVL: One-Step Latent Reasoning and Planning with Vision-Language Explanation</title>
<link rel="icon" href="assets/onevl/figures/xiaomi.png" type="image/x-icon">
<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=Rubik&display=swap" rel="stylesheet">
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<script src="./static/js/distill_template.v2.js"></script>
<script id="MathJax-script" async src="https://cdn.jsdelivr.net/npm/mathjax@3/es5/tex-mml-chtml.js"></script>
<script defer="" src="./static/js/hider.js"></script>
<link rel="stylesheet" href="./static/css/style.css">
<link rel="stylesheet" href="./static/css/onevl.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="https://cdn.jsdelivr.net/npm/katex@0.10.2/dist/katex.min.css" integrity="sha384-yFRtMMDnQtDRO8rLpMIKrtPCD5jdktao2TV19YiZYWMDkUR5GQZR/NOVTdquEx1j" crossorigin="anonymous">
<script defer src="https://cdn.jsdelivr.net/npm/katex@0.10.2/dist/katex.min.js" integrity="sha384-9Nhn55MVVN0/4OFx7EE5kpFBPsEMZxKTCnA+4fqDmg12eCTqGi6+BB2LjY8brQxJ" crossorigin="anonymous"></script>
<script defer src="https://cdn.jsdelivr.net/npm/katex@0.10.2/dist/contrib/auto-render.min.js" integrity="sha384-kWPLUVMOks5AQFrykwIup5lo0m3iMkkHrD0uJ4H5cjeGihAutqP0yW0J6dpFiVkI" crossorigin="anonymous"
onload="renderMathInElement(document.body);"></script>
<script defer src="./static/js/fontawesome.all.min.js"></script>
<script src="https://cdn.jsdelivr.net/npm/jquery@3.7.1/dist/jquery.min.js"></script>
<script defer src="./static/js/medium-zoom.min.js"></script>
<script defer src="./static/js/zoom.js"></script>
<style>
.onevl-name {
font-weight: bold;
}
.onevl-name .o1 { color: #ff6900; }
.onevl-name .n { color: #f47216; }
.onevl-name .e { color: #ea7c2b; }
.onevl-name .v { color: #df8540; }
.onevl-name .l { color: #d48f55; }
</style>
</head>
<body>
<div class="header-wrapper">
<div class="header-container", id="header-container">
<div class="header-content">
<h1 style="font-size:3.2rem;">
<span class="onevl-name"><span class="o1">O</span><span class="n">n</span><span class="e">e</span><span class="v">V</span><span class="l">L</span></span>:
One-Step Latent Reasoning and Planning with Vision-Language Explanation
</h1>
<h2 style="margin:2.2rem 0 1.8rem; display:flex; align-items:center; gap:0.4rem;">
<!-- TODO: Replace with Xiaomi EI team logo if available -->
<span class="icon is-small">
<img src="assets/onevl/figures/xiaomi.png" style="height:1.8em; vertical-align:text-top;">
</span> Xiaomi Embodied Intelligence Team
</h2>
<div class="button-container">
<a href="#" class="button paper-link" target="_blank">
<span class="icon is-small">
<i class="ai ai-arxiv" style="height: 1.5em;"></i>
</span><span>arXiv</span>
</a>
<a href="#" class="button paper-link" target="_blank">
<span class="icon is-small">
<img src="assets_common/icons/pdf.png" style="height: 1.4em;">
</span><span>PDF</span>
</a>
<a href="#" class="button" target="_blank">
<span class="icon is-small">
<img src="assets_common/icons/github.png" style="height: 1.4em;">
</span><span>GitHub</span>
</a>
<a href="#" class="button" target="_blank">
<span class="icon is-small">
<img src="assets_common/icons/hf.png" style="height: 1.5em;">
</span><span>HuggingFace</span>
</a>
</div>
</div>
<div class="header-image">
<!-- TODO: Replace with OneVL teaser bar chart (Fig 1 from paper) -->
<img src="assets/onevl/figures/teaser.png" alt="Teaser Image" class="teaser-image">
</div>
</div>
</div>
<d-article>
<d-contents>
<nav>
<h4>Contents</h4>
<div><a href="#introduction">Introduction</a></div>
<div><a href="#architecture">Architecture</a></div>
<div><a href="#pipeline">Training Pipeline</a></div>
<div><a href="#results">Results</a></div>
<div><a href="#qualitative">Qualitative</a></div>
<div><a href="#examples">Detailed Examples</a></div>
<div><a href="#contributors">Contributors</a></div>
</nav>
</d-contents>
<!-- ==================== Introduction ==================== -->
<section id="introduction">
<div class="wb-section">
<div class="wb-kicker">
Can latent reasoning outperform explicit CoT?
</div>
<div class="onevl-hr"></div>
<p class="wb-lead">
Chain-of-Thought (CoT) reasoning has become a powerful driver of trajectory prediction in VLA-based autonomous driving, yet its autoregressive nature imposes a latency cost that is prohibitive for real-time deployment.
</p>
<p class="wb-lead">
Latent CoT methods compress reasoning into hidden states to reduce latency, but consistently underperform explicit CoT — because purely linguistic latents encode symbolic abstractions, not the causal dynamics that govern driving.
</p>
<figure style="width: 100%; max-width: 1000px; display: block; margin: 24px auto;">
<img src="assets/onevl/figures/paradigm.png" alt="Comparison of three CoT paradigms">
</figure>
<p class="wb-lead">
We argue that the compression target itself must capture genuine causal relationships. A latent vector that compresses only language is merely compressing an abstraction of the world, not the underlying physical structure.
</p>
<p class="wb-lead">
<span class="onevl-name"><span class="o1">O</span><span class="n">n</span><span class="e">e</span><span class="v">V</span><span class="l">L</span></span> addresses this with dual auxiliary decoders: a <b>language decoder</b> that reconstructs text CoT, and a <b>visual world model decoder</b> that predicts future-frame tokens — forcing the latent space to internalize causal scene dynamics rather than symbolic summaries.
</p>
<p class="wb-lead">
Across four benchmarks, <span class="onevl-name"><span class="o1">O</span><span class="n">n</span><span class="e">e</span><span class="v">V</span><span class="l">L</span></span> is the <b>first</b> latent CoT method to surpass explicit CoT, delivering state-of-the-art accuracy at answer-only latency.
</p>
<br><br>
</div>
</section>
<!-- ==================== Architecture ==================== -->
<section id="architecture">
<div class="wb-section">
<h2>Model Architecture</h2>
<!-- TODO: Replace with Fig 3 from paper (full architecture diagram) -->
<figure style="width: 100%; max-width: 1000px; display: block; margin: 24px auto;">
<img src="assets/onevl/figures/onevl.png" alt="OneVL Architecture">
</figure>
<div class="wb-kicker">
<span class="onevl-name"><span class="o1">O</span><span class="n">n</span><span class="e">e</span><span class="v">V</span><span class="l">L</span></span>: Dual-Modal Latent Reasoning
</div>
<div class="onevl-hr"></div>
<p class="wb-lead">
OneVL augments a pretrained VLM with a compact latent token interface and dual auxiliary decoders for multimodal explanation.
During inference, the auxiliary decoders are discarded and all latent tokens are prefilled in a single parallel pass, matching answer-only AR prediction latency.
</p>
<style>
#architecture .wb-grid {
grid-template-columns: repeat(3, minmax(0, 1fr)) !important;
max-width: 960px;
margin: 0 auto;
gap: 2rem 2.5rem;
}
#architecture .wb-grid .wb-card {
margin: 0.75rem 0;
}
@media (max-width: 1024px) {
#architecture .wb-grid { grid-template-columns: 1fr !important; }
}
</style>
<div class="wb-grid">
<article class="wb-card">
<h3>
<a class="wb-anchor">Dual Latent Tokens</a>
</h3>
<p>
<b>35 visual</b> + <b>20 language</b> latent tokens create a tight information bottleneck that forces the model to distill only the causal structure of the scene — discouraging memorization in favor of generalizable representations.
</p>
</article>
<article class="wb-card">
<h3>
<a class="wb-anchor">Language Auxiliary Decoder</a>
</h3>
<p>
Recovers <b>human-readable CoT text</b> from language latent states, grounding the bottleneck in semantic intent: scene interpretation, object analysis, and driving decisions.
</p>
</article>
<article class="wb-card">
<h3>
<a class="wb-anchor">Visual Auxiliary Decoder</a>
</h3>
<p>
Predicts <b>future-frame visual tokens</b> at +0.5s and +1.0s, acting as a <b>world model auxiliary</b> that grounds the bottleneck in physical scene dynamics — a causal compression target that language alone cannot supply.
</p>
</article>
</div>
<br><br><br>
</div>
</section>
<!-- ==================== Training Pipeline ==================== -->
<section id="pipeline">
<div class="wb-section">
<h2>Three-Stage Training Pipeline</h2>
<div class="wb-kicker">
Progressive Alignment for Stable Compression
</div>
<div class="onevl-hr"></div>
<p class="wb-lead">
Training OneVL presents a unique optimization challenge: the main VLM, the language auxiliary decoder, and the visual auxiliary decoder must all be jointly optimized, yet they have fundamentally different learning objectives.
A principled three-stage pipeline progressively aligns these components.
</p>
<div class="stage-grid">
<article class="wb-card">
<h3>
<span class="stage-num">0</span>
<a class="wb-anchor">Main Model Warmup</a>
</h3>
<p>
Train the main VLM end-to-end on trajectory prediction with latent tokens embedded in each training sample.
The model learns to develop meaningful latent representations and establish information routing pathways.
</p>
</article>
<article class="wb-card">
<h3>
<span class="stage-num">1</span>
<a class="wb-anchor">Auxiliary Decoder Warmup</a>
</h3>
<p>
Freeze the main model and train the auxiliary decoders to align with the stable latent representations.
The language decoder learns to decode CoT text; the visual decoder learns to predict future frames.
</p>
</article>
<article class="wb-card">
<h3>
<span class="stage-num">2</span>
<a class="wb-anchor">Joint End-to-End Fine-tuning</a>
</h3>
<p>
Jointly fine-tune all three model components.
Gradients from both decoders flow back into the main model, creating a virtuous cycle that tightens the information bottleneck from both sides.
</p>
</article>
</div>
<br><br><br>
</div>
</section>
<!-- ==================== Results ==================== -->
<section id="results">
<div class="wb-section">
<h2>Key Results</h2>
<div class="wb-kicker">
State-of-the-Art Across Benchmarks
</div>
<div class="onevl-hr"></div>
<p class="wb-lead">
OneVL achieves state-of-the-art performance across NAVSIM, ROADWork, Impromptu, and APR1 with a <b>4B parameter model</b>, surpassing prior 8B methods.
Prefill inference matches answer-only prediction speed, and an MLP variant reaches <b>0.24s latency (4.16 Hz)</b> for real-world deployment.
</p>
<div class="results-grid">
<article class="wb-card result-card">
<div class="result-benchmark">NAVSIM</div>
<div class="result-value">88.84</div>
<div class="result-label">PDM-score</div>
</article>
<article class="wb-card result-card">
<div class="result-benchmark">ROADWork</div>
<div class="result-value">12.49</div>
<div class="result-label">ADE (pixels)</div>
</article>
<article class="wb-card result-card">
<div class="result-benchmark">Impromptu</div>
<div class="result-value">1.31</div>
<div class="result-label">ADE (meters)</div>
</article>
<article class="wb-card result-card">
<div class="result-benchmark">APR1</div>
<div class="result-value">2.62</div>
<div class="result-label">ADE (meters)</div>
</article>
</div>
<br><br><br>
</div>
</section>
<!-- ==================== Qualitative Examples ==================== -->
<section id="qualitative">
<div class="wb-section">
<div class="wb-kicker">
Trajectory Prediction & Interpretable Explanations
</div>
<div class="onevl-hr"></div>
<p class="wb-lead">
OneVL provides human-interpretable explanations in both language and vision.
The language auxiliary decoder recovers high-quality CoT text from compressed latents, while the visual auxiliary decoder generates spatially coherent future-frame previews.
</p>
<br>
<style>
.onevl-video-item {
max-width: 900px;
margin: 0 auto 2.5rem;
}
.onevl-video-item video {
width: 100%;
height: auto;
display: block;
border-radius: 0;
background: transparent;
}
.onevl-video-item p.onevl-video-caption {
text-align: center;
font-size: 0.92rem;
color: #6b7280;
margin: 0 0 0.3rem !important;
padding: 0;
font-family: "Rubik", sans-serif;
}
.onevl-video-item video {
margin-bottom: 0;
}
</style>
<div class="onevl-video-item">
<p class="onevl-video-caption">NAVSIM video log: OneVL trajectory prediction with multi-view and BEV visualization.</p>
<video autoplay loop muted playsinline>
<source src="assets/onevl/videos/demo1.mp4" type="video/mp4" />
</video>
</div>
<div class="onevl-video-item">
<p class="onevl-video-caption">NAVSIM video log: OneVL trajectory prediction with multi-view and BEV visualization.</p>
<video autoplay loop muted playsinline>
<source src="assets/onevl/videos/demo2.mp4" type="video/mp4" />
</video>
</div>
<div class="onevl-video-item">
<p class="onevl-video-caption">NAVSIM video log: OneVL trajectory prediction with multi-view and BEV visualization.</p>
<video autoplay loop muted playsinline>
<source src="assets/onevl/videos/demo3.mp4" type="video/mp4" />
</video>
</div>
<div class="wb-kicker">
Accuracy & Efficiency Across Benchmarks
</div>
<div class="onevl-hr"></div>
<p class="wb-lead">
OneVL consistently achieves the best accuracy across all four benchmarks while matching answer-only prediction latency. Existing latent CoT methods (COCONUT, CODI, SIM-CoT) underperform even the AR baseline, whereas OneVL surpasses explicit AR CoT at a fraction of the inference cost.
</p>
<figure style="width: 100%; max-width: 1000px; display: block; margin: 24px auto;">
<img src="assets/onevl/figures/benchmark.png" alt="Accuracy and efficiency comparison across four benchmarks">
</figure>
<br>
<p class="wb-lead">
On NAVSIM, OneVL achieves <b>88.84</b> PDM-score with a 4B model, surpassing 8B methods AdaThinkDrive (86.20) and LaST-VLA (87.30). Prefill inference reaches <b>4.46s</b> latency — matching answer-only prediction (4.49s) while being 32% faster than explicit AR CoT (6.58s).
</p>
<br>
<style>
.onevl-table-wrap {
max-width: 900px;
margin: 0 auto;
overflow-x: auto;
}
.onevl-table {
width: 100%;
border-collapse: collapse;
font-family: "Rubik", sans-serif;
font-size: 0.92rem;
line-height: 1.5;
}
.onevl-table th,
.onevl-table td {
padding: 0.5rem 0.75rem;
text-align: center;
border-bottom: 1px solid #e5e7eb;
}
.onevl-table th {
font-weight: 600;
color: #111827;
border-bottom: 2px solid #374151;
}
.onevl-table td:first-child,
.onevl-table th:first-child {
text-align: left;
}
.onevl-table .group-header td {
font-style: italic;
color: #6b7280;
text-align: center;
background: #fff3ea;
border-bottom: 1px solid #e5e7eb;
padding: 0.35rem 0.75rem;
font-size: 0.88rem;
}
.onevl-table .row-ours {
background: #fff3ea;
font-weight: 600;
}
.onevl-table .best {
font-weight: 700;
}
.onevl-table .underline {
text-decoration: underline;
}
.onevl-table-caption {
text-align: center;
font-size: 0.88rem;
color: #6b7280;
margin: 0 0 0.5rem !important;
font-family: "Rubik", sans-serif;
}
</style>
<div class="onevl-table-wrap">
<p class="onevl-table-caption"><b>Performance comparisons on the NAVSIM benchmark.</b> PDM-score (higher is better) and average inference latency (lower is better). * indicates the result is derived from the corresponding paper.</p>
<table class="onevl-table">
<thead>
<tr>
<th>Method</th>
<th>Model Size</th>
<th>PDM-score ↑</th>
<th>Latency (s) ↓</th>
<th>Interpretability</th>
</tr>
</thead>
<tbody>
<tr class="group-header"><td colspan="5">Previous State-of-the-Art</td></tr>
<tr>
<td>AdaThinkDrive</td>
<td>8B</td>
<td>86.20*</td>
<td>—</td>
<td>Language</td>
</tr>
<tr>
<td>LaST-VLA</td>
<td>8B</td>
<td>87.30*</td>
<td>—</td>
<td>—</td>
</tr>
<tr class="group-header"><td colspan="5">AR-based Baselines (4B, Qwen3-VL)</td></tr>
<tr>
<td>AR Answer</td>
<td>4B</td>
<td>87.47</td>
<td class="underline">4.49</td>
<td>—</td>
</tr>
<tr>
<td>AR CoT+Answer</td>
<td>4B</td>
<td class="underline">88.29</td>
<td>6.58</td>
<td>Language</td>
</tr>
<tr class="group-header"><td colspan="5">Latent CoT Baselines (4B, Qwen3-VL)</td></tr>
<tr>
<td>COCONUT</td>
<td>4B</td>
<td>84.84</td>
<td>5.93</td>
<td>—</td>
</tr>
<tr>
<td>CODI</td>
<td>4B</td>
<td>83.92</td>
<td>8.62</td>
<td>—</td>
</tr>
<tr>
<td>SIM-CoT</td>
<td>4B</td>
<td>84.21</td>
<td>10.86</td>
<td>Language</td>
</tr>
<tr class="row-ours">
<td><b>OneVL</b></td>
<td>4B</td>
<td class="best">88.84</td>
<td class="best">4.46</td>
<td>Vision + Language</td>
</tr>
</tbody>
</table>
</div>
<br>
<p class="wb-lead">
On ROADWork, OneVL achieves <b>12.49</b> ADE and <b>28.80</b> FDE (pixels), significantly outperforming the previous SOTA YNet (22.68 / 80.78) and all latent CoT baselines. Inference latency is <b>4.71s</b> — faster than answer-only prediction and over 2x faster than explicit AR CoT (10.74s).
</p>
<br>
<div class="onevl-table-wrap">
<p class="onevl-table-caption"><b>Performance comparisons on the ROADWork benchmark.</b> ADE and FDE (pixels; lower is better), latency (lower is better). * indicates the result is derived from the corresponding paper.</p>
<table class="onevl-table">
<thead>
<tr>
<th>Method</th>
<th>ADE (pixel) ↓</th>
<th>FDE (pixel) ↓</th>
<th>Latency (s) ↓</th>
<th>Interpretability</th>
</tr>
</thead>
<tbody>
<tr class="group-header"><td colspan="5">Previous State-of-the-Art</td></tr>
<tr>
<td>YNet</td>
<td>22.68*</td>
<td>80.78*</td>
<td>—</td>
<td>—</td>
</tr>
<tr class="group-header"><td colspan="5">AR-based Baselines (4B, Qwen3-VL)</td></tr>
<tr>
<td>AR Answer</td>
<td>15.98</td>
<td>40.29</td>
<td class="underline">4.74</td>
<td>—</td>
</tr>
<tr>
<td>AR CoT+Answer</td>
<td class="underline">13.18</td>
<td class="underline">29.98</td>
<td>10.74</td>
<td>Language</td>
</tr>
<tr class="group-header"><td colspan="5">Latent CoT Baselines (4B, Qwen3-VL)</td></tr>
<tr>
<td>COCONUT</td>
<td>15.44</td>
<td>38.60</td>
<td class="underline">6.06</td>
<td>—</td>
</tr>
<tr>
<td>CODI</td>
<td>16.45</td>
<td>44.28</td>
<td>6.73</td>
<td>—</td>
</tr>
<tr>
<td>SIM-CoT</td>
<td>16.49</td>
<td>44.32</td>
<td>6.19</td>
<td>Language</td>
</tr>
<tr class="row-ours">
<td><b>OneVL</b></td>
<td class="best">12.49</td>
<td class="best">28.80</td>
<td class="best">4.71</td>
<td>Vision + Language</td>
</tr>
</tbody>
</table>
</div>
<br>
<p class="wb-lead">
On Impromptu, OneVL achieves <b>1.34</b> ADE and <b>3.70</b> FDE (meters), outperforming both Impromptu VLA (1.60 / 4.28) and explicit AR CoT (1.42 / 3.96). Latency is <b>4.02s</b> — faster than answer-only prediction and 41% faster than AR CoT (6.84s).
</p>
<br>
<div class="onevl-table-wrap">
<p class="onevl-table-caption"><b>Performance comparisons on the Impromptu benchmark.</b> ADE and FDE (meters; lower is better), latency (lower is better).</p>
<table class="onevl-table">
<thead>
<tr>
<th>Method</th>
<th>ADE (m) ↓</th>
<th>FDE (m) ↓</th>
<th>Latency (s) ↓</th>
<th>Interpretability</th>
</tr>
</thead>
<tbody>
<tr class="group-header"><td colspan="5">Previous State-of-the-Art</td></tr>
<tr>
<td>Impromptu VLA</td>
<td>1.60</td>
<td>4.28</td>
<td>6.10</td>
<td>—</td>
</tr>
<tr class="group-header"><td colspan="5">AR-based Baselines (4B, Qwen3-VL)</td></tr>
<tr>
<td>AR Answer</td>
<td>1.46</td>
<td>4.03</td>
<td class="underline">4.24</td>
<td>—</td>
</tr>
<tr>
<td>AR CoT+Answer</td>
<td class="underline">1.42</td>
<td class="underline">3.96</td>
<td>6.84</td>
<td>Language</td>
</tr>
<tr class="group-header"><td colspan="5">Latent CoT Baselines (4B, Qwen3-VL)</td></tr>
<tr>
<td>COCONUT</td>
<td>1.49</td>
<td>4.07</td>
<td>5.27</td>
<td>—</td>
</tr>
<tr>
<td>CODI</td>
<td>1.86</td>
<td>5.18</td>
<td>5.24</td>
<td>—</td>
</tr>
<tr>
<td>SIM-CoT</td>
<td>2.43</td>
<td>6.10</td>
<td>5.09</td>
<td>Language</td>
</tr>
<tr class="row-ours">
<td><b>OneVL</b></td>
<td class="best">1.34</td>
<td class="best">3.70</td>
<td class="best">4.02</td>
<td>Vision + Language</td>
</tr>
</tbody>
</table>
</div>
<br>
<p class="wb-lead">
On APR1, OneVL achieves <b>2.62</b> ADE (meters), the best among all methods, and <b>7.53</b> FDE, competitive with Cosmos-Reason (7.42) which uses RL-based fine-tuning. Latency is <b>3.26s</b>, faster than all latent CoT baselines.
</p>
<br>
<div class="onevl-table-wrap">
<p class="onevl-table-caption"><b>Performance comparisons on the APR1 benchmark.</b> ADE and FDE (meters; lower is better), latency (lower is better).</p>
<table class="onevl-table">
<thead>
<tr>
<th>Method</th>
<th>ADE (m) ↓</th>
<th>FDE (m) ↓</th>
<th>Latency (s) ↓</th>
<th>Interpretability</th>
</tr>
</thead>
<tbody>
<tr class="group-header"><td colspan="5">Previous State-of-the-Art</td></tr>
<tr>
<td>Cosmos-Reason</td>
<td class="underline">2.86</td>
<td class="best">7.42</td>
<td>—</td>
<td>Language</td>
</tr>
<tr class="group-header"><td colspan="5">AR-based Baselines (4B, Qwen3-VL)</td></tr>
<tr>
<td>AR Answer</td>
<td>3.27</td>
<td>9.59</td>
<td>3.06</td>
<td>—</td>
</tr>
<tr>
<td>AR CoT+Answer</td>
<td>2.99</td>
<td>8.54</td>
<td>3.51</td>
<td>Language</td>
</tr>
<tr class="group-header"><td colspan="5">Latent CoT Baselines (4B, Qwen3-VL)</td></tr>
<tr>
<td>COCONUT</td>
<td>3.29</td>
<td>9.48</td>
<td>3.76</td>
<td>—</td>
</tr>
<tr>
<td>CODI</td>
<td>3.22</td>
<td>9.25</td>
<td>3.85</td>
<td>—</td>
</tr>
<tr>
<td>SIM-CoT</td>
<td>3.40</td>
<td>9.85</td>
<td>3.78</td>
<td>Language</td>
</tr>
<tr class="row-ours">
<td><b>OneVL</b></td>
<td class="best">2.62</td>
<td class="underline">7.53</td>
<td class="best">3.26</td>
<td>Vision + Language</td>
</tr>
</tbody>
</table>
</div>
<br><br><br><br>
</div>
</section>
<!-- ==================== Detailed Examples ==================== -->
<section id="examples">
<div class="wb-section">
<div class="wb-kicker">
Trajectory, Future Frames & Reasoning
</div>
<div class="onevl-hr"></div>
<p class="wb-lead">
Each panel compares the AR baseline and OneVL side by side: front-view trajectory overlays, bird's-eye-view (BEV) plans, predicted future frames (T+1, T+2), and the decoded chain-of-thought reasoning.
</p>
<style>
.qual-panel {
max-width: 960px;
margin: 1.5rem auto;
border: 1px solid #e5e7eb;
border-radius: 14px;
overflow: hidden;
background: #fff;
box-shadow: none;
}
.qual-row-header {
display: grid;
gap: 0;
background: #fde0c8;
font-family: "Rubik", sans-serif;
font-weight: 600;
font-size: 0.92rem;
color: #111827;
text-align: center;
border-bottom: 1px solid #e5e7eb;
}
.qual-row-header.two-col {
grid-template-columns: 3fr 2fr;
}
.qual-row-header.three-col {
grid-template-columns: 1fr 1fr 1fr;
}
.qual-row-header span {
padding: 0.4rem 0.5rem;
}
.qual-row-header span:not(:last-child) {
border-right: 1px solid rgba(0,0,0,0.08);
}
.qual-row {
display: grid;
gap: 0;
border-bottom: 1px solid #e5e7eb;
}
.qual-row.two-col {
grid-template-columns: 3fr 2fr;
}
.qual-row.three-col {
grid-template-columns: 1fr 1fr 1fr;
}
.qual-row:last-child {
border-bottom: none;
}
.qual-cell {
padding: 0.5rem;
display: flex;
align-items: center;
justify-content: center;
}
.qual-cell:not(:last-child) {
border-right: 1px solid #e5e7eb;
}
.qual-cell img {
width: 100%;
height: auto;
display: block;
border-radius: 6px;
}
.qual-cell .qual-placeholder {
width: 100%;
aspect-ratio: 16 / 10;
background: #f3f4f6;
border: 2px dashed #d1d5db;
border-radius: 6px;
display: flex;
align-items: center;
justify-content: center;
color: #9ca3af;
font-size: 0.82rem;
font-family: "Rubik", sans-serif;
}
.qual-cell .qual-reasoning {
width: 100%;
padding: 0.6rem 0.8rem;
font-size: 0.82rem;
line-height: 1.55;
color: #374151;
font-family: "Rubik", sans-serif;
}
.qual-cell .qual-reasoning b {
color: #111827;
}
@media (max-width: 640px) {
.qual-row.two-col { grid-template-columns: 1fr; }
.qual-row.three-col { grid-template-columns: 1fr; }
.qual-row-header.two-col { grid-template-columns: 1fr; }
.qual-row-header.three-col { grid-template-columns: 1fr; }
.qual-cell:not(:last-child) { border-right: none; border-bottom: 1px solid #e5e7eb; }
.qual-row-header span:not(:last-child) { border-right: none; border-bottom: 1px solid rgba(0,0,0,0.08); }
}
</style>
<div class="qual-panel">
<!-- Row 1: Baseline -->
<div class="qual-row-header two-col">
<span>Baseline Front View</span>
<span>Baseline BEV</span>
</div>
<div class="qual-row two-col">
<div class="qual-cell">
<img src="assets/onevl/figures/example1/baseline_front_view.jpg" alt="Baseline Front View">
</div>
<div class="qual-cell">
<img src="assets/onevl/figures/example1/baseline_bev.png" alt="Baseline BEV">
</div>
</div>
<!-- Row 2: OneVL -->
<div class="qual-row-header two-col">
<span>OneVL Front View</span>
<span>OneVL BEV</span>
</div>
<div class="qual-row two-col">
<div class="qual-cell">
<img src="assets/onevl/figures/example1/onevl_front_view.jpg" alt="OneVL Front View">
</div>
<div class="qual-cell">
<img src="assets/onevl/figures/example1/onevl_bev.png" alt="OneVL BEV">
</div>
</div>
<!-- Row 3: Future frames + Reasoning -->
<div class="qual-row-header" style="display:grid; grid-template-columns: 1fr 1fr;">
<span>OneVL T+1</span>
<span>OneVL T+2</span>
</div>
<div class="qual-row" style="display:grid; grid-template-columns: 1fr 1fr;">
<div class="qual-cell">
<img src="assets/onevl/figures/example1/onevl_t+1.jpg" alt="OneVL T+1 Future Frame">
</div>
<div class="qual-cell">
<img src="assets/onevl/figures/example1/onevl_t+2.jpg" alt="OneVL T+2 Future Frame">
</div>
</div>
<!-- Row 4: Reasoning -->
<div class="qual-row-header" style="display:block; text-align:center;">
<span>OneVL Reasoning</span>
</div>
<div class="qual-row" style="display:block;">
<div class="qual-cell" style="border-right:none;">
<div class="qual-reasoning">
<b>CoT:</b> The right side of the lane where ego vehicle is located is close to the undrivable area, so I need to drive slightly to the left. There are no objects in the current scene that I need to pay attention to. Based on the understanding of the scene and the navigation information, the ego should maintain speed and turn left.
</div>
</div>
</div>
</div>
<br><br><br><br>
</div>
</section>
<!-- ==================== Contributors ==================== -->
<section id="contributors">
<div class="wb-section">
<h2>Contributors</h2>
<p class="wb-lead">
<span class="onevl-name"><span class="o1">O</span><span class="n">n</span><span class="e">e</span><span class="v">V</span><span class="l">L</span></span> is developed by the <img src="assets/onevl/figures/xiaomi.png" style="height:1.4em; vertical-align:text-bottom; margin: 0 0.15em;"> <b>Xiaomi Embodied Intelligence Team</b>.
</p>
<div class="wb-kicker" style="font-size:1.1rem;">Core Contributors</div>
<div class="onevl-hr"></div>
<div class="contributors-name-list">
<span>Jinghui Lu</span>
<span>Jiayi Guan</span>
<span>Zhijian Huang</span>
<span>Jinlong Li</span>
<span>Guang Li</span>
<span>Lingdong Kong</span>
<span>Yingyan Li</span>
<span>Han Wang</span>
<span>Shaoqing Xu</span>
<span>Yuechen Luo</span>
<span>Fang Li</span>
<span>Chenxu Dang</span>
<span>Junli Wang</span>
<span>Tao Xu</span>
<span>Jing Wu</span>
<span>Jianhua Wu</span>
<span>Xiaoshuai Hao</span>
<span>Wen Zhang</span>
<span>Tianyi Jiang</span>
<span>Kuiyuan Yang</span>
<span>Hangjun Ye</span>
<span>Long Chen<sup>†</sup></span>
</div>
<p style="font-size:0.82rem; color:#9ca3af; margin-top:0.3rem; font-family:'Rubik',sans-serif;">Note: <sup>†</sup> Corresponding Author</p>
<br>
<div class="wb-kicker" style="font-size:1.1rem;">Contributors</div>
<div class="onevl-hr"></div>
<div class="contributors-name-list">