-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathindex.html
More file actions
716 lines (651 loc) · 32.4 KB
/
index.html
File metadata and controls
716 lines (651 loc) · 32.4 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
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
<!DOCTYPE html>
<html>
<style type="text/css">
.layered-paper-big { /* modified from: http://css-tricks.com/snippets/css/layered-paper/ */
box-shadow: 0px 0px 1px 1px rgba(0, 0, 0, 0.35), /* The top layer shadow */ 5px 5px 0 0px #fff, /* The second layer */ 5px 5px 1px 1px rgba(0, 0, 0, 0.35), /* The second layer shadow */ 10px 10px 0 0px #fff, /* The third layer */ 10px 10px 1px 1px rgba(0, 0, 0, 0.35), /* The third layer shadow */ 15px 15px 0 0px #fff, /* The fourth layer */ 15px 15px 1px 1px rgba(0, 0, 0, 0.35), /* The fourth layer shadow */ 20px 20px 0 0px #fff, /* The fifth layer */ 20px 20px 1px 1px rgba(0, 0, 0, 0.35), /* The fifth layer shadow */ 25px 25px 0 0px #fff, /* The fifth layer */ 25px 25px 1px 1px rgba(0, 0, 0, 0.35); /* The fifth layer shadow */
margin-left: 10px;
margin-right: 45px;
}
vl {
border-left: 2px solid black;
height: 300px;
position: absolute;
left: 50%;
margin-left: -3px;
top: 50;
}
.centerTable { margin: 0px auto; }
.center {
display: block;
margin-left: auto;
margin-right: auto;
width: 50%;
}
hr {
border: 0;
height: 1px;
background-image: linear-gradient(to right, rgba(0, 0, 0, 0), rgba(0, 0, 0, 0.15), rgba(0, 0, 0, 0));
}
table caption {
padding: 10px;
caption-side: bottom;
}
figure {
display: inline-block;
margin: 20px; /* adjust as needed */
}
figure img {
vertical-align: top;
}
figure figcaption {
text-align: center;
}
</style>
<style>
.videoWrapper {
position: relative;
top: 0;
left: 0;
width: 100%;
height: 100%;
}
.videoWrapper iframe {
position: absolute;
top: 0;
left: 0;
width: 100%;
height: 100%;
}
</style>
<head>
<meta charset="utf-8">
<meta property="og:title"
content="SoftCTRL: Soft conservative KL-control of Transformer Reinforcement Learning for Autonomous Driving"/>
<meta property="og:description" content="Drive&Segment is a novel method for cross-modal unsupervised learning of semantic image segmentation
by leveraging synchronized LiDAR and image data"/>
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>SoftCtrl page</title>
<script>
function CopyToClipboard(id) {
var r = document.createRange();
r.selectNode(document.getElementById(id));
window.getSelection().removeAllRanges();
window.getSelection().addRange(r);
document.execCommand('copy');
window.getSelection().removeAllRanges();
}
</script>
<!-- Global site tag (gtag.js) - Google Analytics -->
<script async src="https://www.googletagmanager.com/gtag/js?id=G-1P8DXFXFKM"></script>
<script>
window.dataLayer = window.dataLayer || [];
function gtag() {
dataLayer.push(arguments);
}
gtag('js', new Date());
gtag('config', 'G-1P8DXFXFKM');
</script>
<!-- <link rel="stylesheet" href="https://gradio.s3-us-west-2.amazonaws.com/2.6.2/static/bundle.css"> -->
<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/bulma-carousel.min.css">
<link rel="stylesheet" href="./static/css/bulma-slider.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="icon" href="./static/images/favicon.svg">
<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>
<script src="./static/js/bulma-carousel.min.js"></script>
<script src="./static/js/bulma-slider.min.js"></script>
</head>
<!-- Google tag (gtag.js) -->
<script async src="https://www.googletagmanager.com/gtag/js?id=G-C2VYDTHPJZ"></script>
<script>
window.dataLayer = window.dataLayer || [];
function gtag(){dataLayer.push(arguments);}
gtag('js', new Date());
gtag('config', 'G-C2VYDTHPJZ');
</script>
<body>
<nav class="navbar" role="navigation" aria-label="main navigation">
<div class="navbar-brand">
<a role="button" class="navbar-burger" aria-label="menu" aria-expanded="false">
<span aria-hidden="true"></span>
<span aria-hidden="true"></span>
<span aria-hidden="true"></span>
</a>
</div>
</div>
</nav>
<section class="hero">
<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-1 publication-title">SoftCTRL: Soft conservative KL-control of Transformer Reinforcement Learning
<br> for Autonomous Driving</h1>
<div class="is-size-5 publication-authors">
<div class="column has-text-centered">
<!-- <h1 class="title is-3 publication-title">ECCV 2022, oral presentation</h1> -->
<div class="is-size-5 publication-authors">
<span class="author-block">
<a href="https://pronton2001.github.io/" target="_blank"
rel="noopener noreferrer">Tri Minh Huynh</a>,</span>
<span class="author-block">
<a href="https://pronton2001.github.io/" target="_blank"
rel="noopener noreferrer">Duc Dung Nguyen</a></span>
<span class="author-block">
</span>
</div>
<div class="is-size-5 publication-authors">
<span class="author-block">VNU HCMUT</span>
</div>
<div class="column has-text-centered">
<div class="publication-links">
<!-- arXiv Link. -->
<span class="link-block">
<a href="https://arxiv.org/abs/2410.22752" target="_blank" rel="noopener noreferrer"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="ai ai-arxiv"></i>
</span>
<span>arXiv</span>
</a>
</span>
<!-- Code Link. -->
<span class="link-block">
<a href="https://github.com/Pronton2001/softCTRL" target="_blank" rel="noopener noreferrer"
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>
<!-- Code Link. -->
<!-- <span class="link-block">
<a href="https://huggingface.co/spaces/vobecant/DaS" target="_blank" rel="noopener noreferrer"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fa fa-gamepad"></i>
</span>
<span>Gradio</span>
</a>
</span> -->
<!-- Code Link. -->
<!-- <span class="link-block">
<a href="https://colab.research.google.com/drive/126tBVYbt1s0STyv8DKhmLoHKpvWcv33H?usp=sharing"
target="_blank" rel="noopener noreferrer"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fab fa-google"></i>
</span>
<span>Colab</span>
</a>
</span> -->
<!-- Video Link. -->
<!-- <span class="link-block">
<a href="https://www.youtube.com/watch?v=B9LK-Fxu7ao" target="_blank" rel="noopener noreferrer"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fab fa-youtube"></i>
</span>
<span>Video</span>
</a>
</span> -->
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
<!-- Abstract. -->
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-3">Abstract</h2>
<div class="content has-text-justified">
<p>
In recent years, motion planning for urban self-driving cars (SDV) has
become a popular problem due to its complex interaction of road components.
To tackle this, many methods have relied on large-scale, human-sampled data
processed through Imitation learning (IL). Although effective, IL alone
cannot adequately handle safety and reliability concerns. Combining IL with
Reinforcement learning (RL) by adding KL divergence between RL and IL policy
to the RL loss can alleviate IL's weakness but suffer from over-conservation
caused by covariate shift of IL. To address this limitation, we introduce
a method that combines IL with RL using an implicit entropy-KL control that
offers a simple way to reduce the over-conservation characteristic. In
particular, we validate different challenging simulated urban scenarios
from the unseen dataset, indicating that although IL can perform well in
imitation tasks, our proposed method significantly improves robustness
(over 17% reduction in failures) and generates human-like driving behavior.
</p>
<img class="round" style="width:1000px" src="./sources/teaser/transformerRL_pipeline.svg"/>
</div>
</div>
</div>
<!-- <div>
<img class="round" style="width:400px" src="./sources/teaser/transformerRL-pipeline.svg"/>
<vl>
<img class="round" style="width:800px" src="./sources/teaser/rl_finetuning.svg"/>
</div> -->
</div>
</section>
<hr>
<section class="section">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-3">Sample Outputs</h2>
</div>
</div>
<div class="columns is-centered">
<img class="round" style="width:200px" src="./sources/qualitative/videos/SAC_ImKL/4.gif"/>
<img class="round" style="width:200px" src="./sources/qualitative/videos/SAC_ImKL/5.gif"/>
<img class="round" style="width:200px" src="./sources/qualitative/videos/SAC_ImKL/6.gif"/>
<img class="round" style="width:200px" src="./sources/qualitative/videos/SAC_ImKL/7.gif"/>
<img class="round" style="width:200px" src="./sources/qualitative/videos/SAC_ImKL/9.gif"/>
</div>
<br>
<div class="columns is-centered">
<img class="round" style="width:200px" src="./sources/qualitative/videos/SAC_ImKL/10.gif"/>
<img class="round" style="width:200px" src="./sources/qualitative/videos/SAC_ImKL/43.gif"/>
<img class="round" style="width:200px" src="./sources/qualitative/videos/SAC_ImKL/57.gif"/>
<img class="round" style="width:200px" src="./sources/qualitative/videos/SAC_ImKL/65.gif"/>
<img class="round" style="width:200px" src="./sources/qualitative/videos/SAC_ImKL/78.gif"/>
</div>
<!-- <div class="content has-text-justified"> -->
<div class="columns is-centered">
<p>
<b>Qualitative results.</b>
Demostrations of SoftCTRL in the unseen test set. Each video is a 25-second scenario,
SDV in red, other agents in blue, crosswalks in yellow, SDV trajectory
in the green line, SDV ground-truth trajectory in yellow line.
</p>
</div>
</div>
</section>
<hr>
<section class="section">
<div class="container is-max-desktop">
<!-- Abstract. -->
<div class="columns is-centered has-text-centered">
<!-- <div class="column is-four-fifths"> -->
<h2 class="title is-3">Overview</h2>
<!-- </div> -->
</div>
<div class="content has-text-justified">
<p>
Combining IL and RL is crucial to enhance realism with large-scale human
driving datasets and safety with causal relationship modeling.
Traditional KL-loss methods use IL (teacher) policy to constrain RL
(student) policy. However, these approach may lead to the over-conservation
problem caused by covariate shift of estimated expert policies (IL),
preventing the RL model from searching for a global solution.
</p>
<p>
To address these problems, we propose a <b>S</b>oft conservative <b>KL</b>-<b>C</b>on<b>T</b>rol
of <b>RL</b> model (<b>SoftCTRL</b>) for autonomous driving, a novel approach for
off-policy RL. First, we use a pretrained Transformer IL model to
constrain RL updates with simple rewards. Regarding RL network, we leverage
a learnable Transformer encoder to embed vectorized information and a MLP
decoder to output Q value and action distribution. During the training
step, the RL policy is implicitly regularized by the KL divergence from this
prior model and the entropy of its actions. The entropy term
encourages the model to generate more diverse trajectories.
This combats the over-conservation issues when retaining the
proximity to the distribution of the realistic IL model by KL
control. Our experiments show that a simple reward combining
imitation with collision is sufficient for our proposed method.
Also, our new approach beat the pretrained IL model by a
large margin in overall tasks, demonstrating that our approach
is not affected by the covariate shift of IL.
</p>
</div>
<!-- <div>
<img class="round" style="width:400px" src="./sources/teaser/transformerRL-pipeline.svg"/>
<vl>
<img class="round" style="width:800px" src="./sources/teaser/rl_finetuning.svg"/>
</div> -->
</div>
</section>
<hr>
<section class="section">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-2">Method</h2>
</div>
</div>
<div class="content has-text-justified">
<p>
We propose SoftCTRL - a novel implicit entropy-KL control approach
for RL fine-tuning. By leverage IL and RL advantages while mitigating
the over-conservation issue inherent in traditional KL-regularized methods,
SoftCTRL surpasses the performance of its IL baseline.
</p>
</div>
<div class="columns is-centered">
<img class="round" style="width:850px" src="./sources/method/klsac.svg"/>
</div>
<div class="content has-text-justified">
<p>
SoftCTRL is built on SAC, we simply modify the regression
target of the policy evaluation step in SAC by adding the scaled log policy
of reference behavioral method to the reward in any TD scheme. What happens underneath this
modification is that it implicitly performs maximum entropy and minimum KL regularization from the
reference policy π0, with KL scaled by ατ and entropy scaled
by (1 − α)τ
</p>
</div>
</div>
</section>
<hr>
<section class="section">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<!-- <div class="column is-four-fifths"> -->
<h2 class="title is-2">Result Highlights</h2>
<!-- </div> -->
</div>
<br>
<div class="columns is-centered">
<p>
<!-- <b>Qualitative results.</b> -->
In the table below, we find that our SAC-ImKL (SoftCTRL) outperforms IL-only,
RL-only, and KL-regularized RL by reducing 17% failure cases compared to BC-perturb
and SAC-ExKL baselines. Moreover, our approach achieves a smooth driving performance on
par with IL-only policy and 3.55x less jerker motion than RL-only.
</p>
</div>
<!-- <div class="content has-text-justified">
<p>
</p>
</div> -->
<!-- <div class="columns is-centered"> -->
<figure>
<img class="center" style="width:1000px" src="./sources/method/quantitative.png"/>
<figcaption class = 'round'>
<b>Quantitative results.</b>
Evaluation of different methods using closed-loop evaluation in DriverGym
on 100 sampled scenarios. Lower is better. The values in the table are
expressed as mean ± standard deviation. Our SAC-ImKL (or SoftCTRL) method improves safety
over baselines while driving as smoothly as BC-perturb
</figcaption>
</figure>
<!-- </div> -->
<br>
<div class="content has-text-justified">
<p>
We further study the complementarity role of entropy and KL divergence to balance
between the unnatural (diversity) and conservation (realism) of the model.
Our experiment shows that a high enough value of KL (=0.7) helps to align the
smoothness close to the reference policy (3.84), but too large
KL causes high failures and jerkier motion. Notable, increasing the entropy coefficient (e.g. 4, 7),
is beneficial in dropping both discomfort rate and failure cases.
</p>
</div>
<!-- <figure>
<img class="round" style="width:400px" src="./sources/experiments/entropy_effect_of_SAC_ImKL.svg"/>
<figcaption> Entropy coefficient (KL coefficient = 0.5) </figcaption>
</figure>
<figure>
<img class="round" style="width:400px" src="./sources/experiments/KL_effect_of_SAC_ImKL.svg"/>
<figcaption> KL coefficient (entropy coefficient= 0.7) </figcaption>
</figure> -->
<TABLE BORDER="0" class="centerTable" id="cssTable">
<caption>
<b>Entropy and KL trade-off.</b> The effect of KL and entropy terms of SAC-ImKL (SoftCTRL) that
are evaluated on the validation set (Failure and Discomfort).
The blue and red dash lines indicate the Failure and Discomfort of BC-perturb respectively</caption>
</div>
<TR>
<!-- <TD> <img class="round" style="width:400px" src="./sources/experiments/entropy_effect_of_SAC_ImKL.svg"/> </TD> -->
<TD>
<figure>
<img class="round" style="width:600px" src="./sources/experiments/entropy_effect_of_SAC_ImKL.svg"/>
<figcaption> Entropy coefficient (KL coefficient = 0.5) </figcaption>
</figure>
</TD>
<!-- <TD> <img class="round" style="width:400px" src="./sources/experiments/KL_effect_of_SAC_ImKL.svg"/> </TD> -->
<TD>
<figure>
<img class="round" style="width:600px" src="./sources/experiments/KL_effect_of_SAC_ImKL.svg"/>
<figcaption> KL coefficient (entropy coefficient= 0.7) </figcaption>
</figure>
</TD>
</TR>
</TABLE>
</div>
</section>
<hr>
<section class="section">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-2">Visualizations</h2>
</div>
</div>
<div>
<TABLE BORDER="0" class="centerTable" id="cssTable">
<caption><b>Scenario 1: Red traffic light.</b> BC-perturb collides with a front vehicle stopped at a red light, while RL methods
stop and leave enough clearance</caption>
<TR>
<TD><b>BC</b></TD> <TD ><b>SAC</b></TD> <TD> <b>SAC-ExKL</b> </TD> <TD> <b>SAC-ImKL</b> </TD>
</TR>
<TR>
<TD> <img class="round" style="width:250px" src="./sources/qualitative/videos/BC_perturb/2.gif"/> </TD>
<TD> <img class="round" style="width:250px" src="./sources/qualitative/videos/SAC/2.gif"/> </TD>
<TD> <img class="round" style="width:250px" src="./sources/qualitative/videos/SAC_ExKL/2.gif"/> </TD>
<TD> <img class="round" style="width:250px" src="./sources/qualitative/videos/SAC_ImKL/2.gif"/> </TD>
</TR>
<TR>
<TD> <img class="round" style="width:250px" src="./sources/qualitative/actions/BC_perturb/2.png"/> </TD>
<TD> <img class="round" style="width:250px" src="./sources/qualitative/actions/SAC/2.png"/> </TD>
<TD> <img class="round" style="width:250px" src="./sources/qualitative/actions/SAC_ExKL/2.png"/> </TD>
<TD> <img class="round" style="width:250px" src="./sources/qualitative/actions/SAC_ImKL/2.png"/> </TD>
</TR>
</TABLE>
</div>
<br>
<div>
<TABLE BORDER="0" class="centerTable" id="cssTable">
<CAPTION><b>Scenario 2: Dense traffic.</b> SAC changes lane and gets stuck in the middle of an intersection,
in contrast, BC and SAC-ExKL and SAC-ImKL (SoftCTRL) drives smoothly through the intersection without any collision.
</CAPTION>
<TR>
<TD><b>BC</b></TD> <TD ><b>SAC</b></TD> <TD> <b>SAC-ExKL</b> </TD> <TD> <b>SAC-ImKL</b> </TD>
</TR>
<TR>
<TD> <img class="round" style="width:250px" src="./sources/qualitative/videos/BC_perturb/65.gif"/> </TD>
<TD> <img class="round" style="width:250px" src="./sources/qualitative/videos/SAC/65.gif"/> </TD>
<TD> <img class="round" style="width:250px" src="./sources/qualitative/videos/SAC_ExKL/65.gif"/> </TD>
<TD> <img class="round" style="width:250px" src="./sources/qualitative/videos/SAC_ImKL/65.gif"/> </TD>
</TR>
<TR>
<TD> <img class="round" style="width:250px" src="./sources/qualitative/actions/BC_perturb/65.png"/> </TD>
<TD> <img class="round" style="width:250px" src="./sources/qualitative/actions/SAC/65.png"/> </TD>
<TD> <img class="round" style="width:250px" src="./sources/qualitative/actions/SAC_ExKL/65.png"/> </TD>
<TD> <img class="round" style="width:250px" src="./sources/qualitative/actions/SAC_ImKL/65.png"/> </TD>
</TR>
</TABLE>
<br>
<div>
<TABLE BORDER="0" class="centerTable" id="cssTable">
<CAPTION><b>Scenario 3: T-junction. </b>All RL approaches diverge from human logs, whereas IL accurately predicts the correct turns.
Therefore, IL achieves significantly lower IL error compared to RL agents. SAC and SAC-ExKL exhibit erratic behavior and drive off-road.
Notably, SAC-ImKL (SoftCTRL) successfully executes a smooth left turn without colliding with other objects or driving off-road.
</CAPTION>
<TR>
<TD><b>BC</b></TD> <TD ><b>SAC</b></TD> <TD> <b>SAC-ExKL</b> </TD> <TD> <b>SAC-ImKL</b> </TD>
</TR>
<TR>
<TD> <img class="round" style="width:250px" src="./sources/qualitative/videos/BC_perturb/57.gif"/> </TD>
<TD> <img class="round" style="width:250px" src="./sources/qualitative/videos/SAC/57.gif"/> </TD>
<TD> <img class="round" style="width:250px" src="./sources/qualitative/videos/SAC_ExKL/57.gif"/> </TD>
<TD> <img class="round" style="width:250px" src="./sources/qualitative/videos/SAC_ImKL/57.gif"/> </TD>
</TR>
<TR>
<TD> <img class="round" style="width:250px" src="./sources/qualitative/actions/BC_perturb/57.png"/> </TD>
<TD> <img class="round" style="width:250px" src="./sources/qualitative/actions/SAC/57.png"/> </TD>
<TD> <img class="round" style="width:250px" src="./sources/qualitative/actions/SAC_ExKL/57.png"/> </TD>
<TD> <img class="round" style="width:250px" src="./sources/qualitative/actions/SAC_ImKL/57.png"/> </TD>
</TR>
</TABLE>
<div class="content has-text-justified">
<b>Qualitative results. </b>Demostrations of 3 win case against baselines method in
the unseen test set (SDV in red, other agents in blue, crosswalks in yellow, SDV trajectory
in the green line, SDV ground-truth trajectory in yellow line). The bottom side plots are the associated acceleration action
(red dashed lines represent the boundary value of smoothness). Every plot depicts 250 time
steps in 1 scene, which is a 25-second scene (0.1 sec interval)
</div>
</div>
<!-- <p>BC</p>
</div>
<div>
<img class="round" style="width:200px" src="./sources/qualitative/videos/BC_perturb/2.gif"/>
<img class="round" style="width:250px" src="./sources/qualitative/actions/BC_perturb/2.png"/>
<img class="round" style="width:200px" src="./sources/qualitative/videos/BC_perturb/2.gif"/>
<img class="round" style="width:250px" src="./sources/qualitative/actions/BC_perturb/2.png"/>
</div> -->
</div>
</section>
<!-- <hr> -->
<!-- <section class="section">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths"> -->
<!-- <h2 class="title is-2">Code</h2>
<p> We provide inference code of our method on GitHub. Furthermore, to show our method, we provide
a Colab notebook.</p> -->
<!-- Code Link. -->
<!-- <span class="link-block">
<a href="https://github.com/vobecant/DriveAndSegment" target="_blank" rel="noopener noreferrer"
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> -->
<!-- Code Link. -->
<!-- <span class="link-block">
<a href="https://huggingface.co/spaces/vobecant/DaS" target="_blank" rel="noopener noreferrer"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fa fa-gamepad"></i>
</span>
<span>Gradio</span>
</a>
</span>
-->
<!-- Code Link. -->
<!-- <span class="link-block">
<a href="https://colab.research.google.com/drive/126tBVYbt1s0STyv8DKhmLoHKpvWcv33H?usp=sharing"
target="_blank" rel="noopener noreferrer"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fab fa-google"></i>
</span>
<span>Colab</span>
</a>
</span> -->
<!-- </div>
</div>
</div>
</section> -->
<!-- <section class="section">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h3 class="title is-3">Live Demo</h3>
<p> If the Huggingface demo below is not loading properly for you, please visit this <a
href="https://huggingface.co/spaces/vobecant/DaS" target="_blank" rel="noopener noreferrer">link
<i class="fas fa-external-link-alt"></i></a>. </p>
</div>
</div>
</div>
</section> -->
<section class="section">
<div id="target"></div>
<script src="https://gradio.s3-us-west-2.amazonaws.com/2.6.2/static/bundle.js"></script>
<script>
launchGradioFromSpaces("vobecant/DaS", "#target")
</script>
</section>
<hr>
<section class="section">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-2">Paper and Supplementary Material</h2>
</div>
</div>
<br>
<div class="columns is-centered">
<a href="https://arxiv.org/abs/2203.11160" target="_blank" rel="noopener noreferrer"><img
class="layered-paper-big" style="height:175px"
src="./sources/paper.png"/></a>
<span style="font-size:14pt">Tri, Huynh and Dung, Nguyen.<br>
<b>SoftCTRL: Soft conservative KL-control of Transformer Reinforcement Learning
<br> for Autonomous Driving</b><br>
ArXiv preprint, 2024.<br>
(hosted on <a href="https://arxiv.org/abs/2410.22752" target="_blank"
rel="noopener noreferrer">ArXiv</a>)<br>
</span>
</div>
<br>
<div class="column is-four-fifths">
<h2 class="title is-3">BibTeX <a href="#" onclick="CopyToClipboard('bibtex');return false;"><span
class="icon"><i class="fa fa-copy"></i></span></a></h2>
</div>
<div class="columns is-centered">
<pre><code id="bibtex">@misc{huynh2024softctrl,
title={SoftCTRL: Soft conservative KL-control of Transformer Reinforcement Learning for Autonomous Driving},
author={Minh Tri Huynh and Duc Dung Nguyen},
year={2024}
eprint={2410.22752},
archivePrefix={arXiv},
primaryClass={cs.RO},
url={https://arxiv.org/abs/2410.22752},
}</code></pre>
</div>
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
<!-- Ack. -->
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h4 class="title is-3">Acknowledgements</h4>
<div class="content has-text-justified">
<p>
The computational resource for this work is supported by Innovation FabLab,
Ho Chi Minh City University of Technology (HCMUT).
</p>
</div>
</div>
</div>
</div>
</section>
<footer class="footer">
<div class="container">
<div class="columns is-centered">
<div class="column is-8">
<div class="content">
<p>
This website is forked from source of <a href="https://vobecant.github.io/DriveAndSegment/"
target="_blank" rel="noopener noreferrer">
this page</a>.
</p>
</div>
</div>
</div>
</div>
</footer>
</body>
</html>