-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathindex.html
882 lines (824 loc) · 48.5 KB
/
index.html
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
<!DOCTYPE html>
<html>
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>On the Data-Efficiency with Contrastive Image Transformation in Reinforcement Learning</title>
<script id="MathJax-script" async src="https://cdn.jsdelivr.net/npm/mathjax@3/es5/tex-mml-chtml.js">
var task_map = {
"simple-object-manipulation": "simple_object_manipulation",
"visual-goal-reaching": "visual_goal_reaching",
"novel-concept-grounding": "novel_concept_grounding",
"one-shot-video-imitation": "one_shot_video_imitation",
"visual-constraint-satisfaction": "visual_constraint_satisfaction",
"visual-reasoning": "visual_reasoning"
};
function updateDemoVideo(category) {
// var demo = document.getElementById("single-menu-demos").value;
var task = document.getElementById(category + "-menu-tasks").value;
var inst = document.getElementById(category + "-menu-instances").value;
console.log(task_map[category], task, inst)
var video = document.getElementById(category + "-single-task-video");
video.src = "assets/videos/demos/" +
task_map[category] +
"/" +
task +
"/" +
inst +
".mp4";
video.playbackRate = 2.0;
video.play();
}
</script>
<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="./static/css/academicons.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">
<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>
<script src="./static/js/index.js"></script>
</head>
<body>
<header style="background-color:hsl(210, 27%, 96%);width:100%;margin:auto;">
<div class="container">
<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">On the Data-Efficiency with Contrastive Image Transformation in
Reinforcement Learning</h1>
<div class="is-size-4 publication-authors">
<!-- <span class="author-block"> -->
<span class="author-block">Sicong Liu</a><sup>1 2 3 *</sup>, </span>
<span class="author-block">Xi Sheryl Zhang</a><sup>2 3 5†</sup>, </span>
<span class="author-block">Yushuo Li</a><sup>2</sup>, </span>
<span class="author-block">Yifan Zhang</a><sup>2 3 5</sup>, </span>
<span class="author-block">Jian Cheng</a><sup>2 3 4</sup>, </span>
<!-- <a target="_blank">Sicong Liu</a><sup>1 2 3 *</sup>,
<a target="_blank">Xi Sheryl Zhang</a><sup>2 3 5†</sup>,
<a target="_blank">Yushuo Li</a><sup>2</sup>,
<a target="_blank">Yifan Zhang</a><sup>2 3 5</sup>,
<a target="_blank">Jian Cheng</a><sup>2 3 4</sup>, -->
<!-- <a target="_blank" href="">Jian Cheng</a><sup>2 3 4</sup>, -->
<!-- <br> -->
<!-- † 十字架符号 -->
<!-- </span> -->
</div>
<div class="is-size-5 publication-authors">
<span class="author-block"><sup>1</sup>NJUST, </span>
<span class="author-block"><sup>2</sup>
Institute of Automation, Chinese Academy of Sciences, </span>
<span class="author-block"><sup>3</sup>AIRIA, </span>
<span class="author-block"><sup>4</sup>School of Future Technology, University of Chinese Academy of
Sciences, </span>
<span class="author-block"><sup>5</sup>University of Chinese Academy of Sciences, Nanjing
, </span>
</div>
<div class="is-size-5 publication-authors">
<span class="author-block"><sup>*</sup>Work done during CASIA internship</span>
<!-- <br> -->
<span class="author-block"><sup>†</sup>Corresponding Author </span>
</div>
<div class="column has-text-centered">
<div class="is-size-5 publication-links">
<!-- TODO PDF Link. -->
<span class="link-block">
<a target="_blank" href="https://openreview.net/forum?id=-nm-rHXi5ga">
<!-- class="external-link button is-normal is-rounded is-dark"> -->
<!-- <span class="icon"> -->
<!-- <i class="ai ai-arxiv"></i> -->
<!-- <i class="ai ai-openreview"></i> -->
<!-- </span> -->
<span>OpenReview</span>
</a>
</span>
<span>/</span>
<span class="link-block">
<a target="_blank" href="assets/coit_paper.pdf">
<!-- class="external-link button is-normal is-rounded is-dark"> -->
<!-- <span class="icon">
<i class="fas fa-file-pdf"></i>
</span> -->
<span>PDF</span>
</a>
</span>
<span>/</span>
<!-- Code Link. -->
<span class="link-block">
<a target="_blank" href="https://github.com/mooricAnna/CoIT">
<!-- class="external-link button is-normal is-rounded is-dark"> -->
<!-- <span class="icon">
<i class="fab fa-github"></i>
</span> -->
<span>Code</span>
</a>
<!-- <a target="_blank" href="https://github.com/vimalabs/VIMA#pretrained-models"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fa fa-network-wired"></i>
</span>
<span>Models</span>
</a>
<a target="_blank" href="https://github.com/vimalabs/VimaBench"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fas fa-robot"></i>
</span>
<span>Benchmark</span>
</a> -->
<!-- <a target="_blank" href="https://huggingface.co/datasets/VIMA/VIMA-Data"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fas fa-database"></i>
</span>
<span>Dataset</span> -->
<!-- </a> -->
</span>
</div>
</div>
</div>
</div>
</div>
</div>
</section>
</div>
</header>
<!-- <section class="hero is-light is-small">
<div class="hero-body">
<div class="container">
<div id="results-carousel" class="carousel results-carousel">
<div class="item item-sweep_without_exceeding">
<video poster="" id="sweep_without_exceeding" autoplay controls muted loop height="100%"
playbackRate=2.0>
<source src="assets/videos/demos/visual_constraint_satisfaction/sweep_without_exceeding/1.mp4"
type="video/mp4">
</video>
</div>
<div class="item item-sweep_without_touching">
<video poster="" id="sweep_without_touching" autoplay controls muted loop height="100%"
playbackRate=2.0>
<source src="assets/videos/demos/visual_goal_reaching/rearrange_then_restore/1.mp4"
type="video/mp4">
</video>
</div>
<div class="item item-rotate">
<video poster="" id="rotate" autoplay controls muted loop height="100%">
<source src="assets/videos/demos/one_shot_video_imitation/follow_motion/1.mp4"
type="video/mp4">
</video>
</div>
<div class="item item-scene_understanding">
<video poster="" id="scene_understanding" autoplay controls muted loop height="100%">
<source src="assets/videos/demos/novel_concept_grounding/twist/1.mp4"
type="video/mp4">
</video>
</div>
<div class="item item-simple_manipulation">
<video poster="" id="simple_manipulation" autoplay controls muted loop height="100%">
<source src="assets/videos/demos/visual_reasoning/manipulate_old_neighbor/1.mp4"
type="video/mp4">
</video>
</div>
<div class="item item-novel_adj">
<video poster="" id="novel_adj" autoplay controls muted loop height="100%">
<source src="assets/videos/demos/one_shot_video_imitation/follow_order/1.mp4"
type="video/mp4">
</video>
</div>
<div class="item item-novel_adj_and_noun">
<video poster="" id="novel_adj_and_noun" autoplay controls muted loop height="100%">
<source src="assets/videos/demos/simple_object_manipulation/scene_understanding/1.mp4"
type="video/mp4">
</video>
</div>
<div class="item item-novel_noun">
<video poster="" id="novel_noun" autoplay controls muted loop height="100%">
<source src="assets/videos/demos/visual_constraint_satisfaction/sweep_without_touching/1.mp4"
type="video/mp4">
</video>
</div>
<div class="item item-twist">
<video poster="" id="twist" autoplay controls muted loop height="100%">
<source src="assets/videos/demos/novel_concept_grounding/novel_adj/1.mp4"
type="video/mp4">
</video>
</div>
<div class="item item-follow_motion">
<video poster="" id="follow_motion" autoplay controls muted loop height="100%">
<source src="assets/videos/demos/simple_object_manipulation/simple_manipulation/1.mp4"
type="video/mp4">
</video>
</div>
<div class="item item-follow_order">
<video poster="" id="follow_order" autoplay controls muted loop height="100%">
<source src="assets/videos/demos/visual_goal_reaching/rearrange/1.mp4"
type="video/mp4">
</video>
</div>
<div class="item item-rearrange">
<video poster="" id="rearrange" autoplay controls muted loop height="100%">
<source src="assets/videos/demos/simple_object_manipulation/rotate/1.mp4"
type="video/mp4">
</video>
</div>
<div class="item item-manipulate_old_neighbor">
<video poster="" id="manipulate_old_neighbor" autoplay controls muted loop height="100%">
<source src="assets/videos/demos/visual_reasoning/same_color/1.mp4"
type="video/mp4">
</video>
</div>
<div class="item item-pick_in_order_then_restore">
<video poster="" id="pick_in_order_then_restore" autoplay controls muted loop height="100%">
<source src="assets/videos/demos/visual_reasoning/pick_in_order_then_restore/1.mp4"
type="video/mp4">
</video>
</div>
<div class="item item-same_color">
<video poster="" id="same_color" autoplay controls muted loop height="100%">
<source src="assets/videos/demos/novel_concept_grounding/novel_noun/1.mp4"
type="video/mp4">
</video>
</div>
<div class="item item-rearrange_then_restore">
<video poster="" id="rearrange_then_restore" autoplay controls muted loop height="100%">
<source src="assets/videos/demos/visual_reasoning/same_profile/1.mp4"
type="video/mp4">
</video>
</div>
<div class="item item-same_profile">
<video poster="" id="same_profile" autoplay controls muted loop height="100%">
<source src="assets/videos/demos/novel_concept_grounding/novel_adj_and_noun/1.mp4"
type="video/mp4">
</video>
</div>
</div>
</div>
</div>
</section> -->
<!-- <div style="border:1px groove #CCC"> -->
<!-- <hr width="1000" style="margin-left: auto; margin-right: auto"> -->
<section class="section">
<div class="container is-max-desktop">
<!-- Gifs. -->
<div class="rows is-centered has-text-centered">
<div class="row">
<div class="content has-text-justified">
<img src="assets/gifs/cartpole_swingup.gif" style="width: 19%" float="center"/>
<img src="assets/gifs/pendulum_swingup.gif" style="width: 19%" float="center"/>
<img src="assets/gifs/cup_catch.gif" style="width: 19%" float="center"/>
<img src="assets/gifs/quadruped_walk.gif" style="width: 19%" float="center"/>
<img src="assets/gifs/finger_spin.gif" style="width: 19%" float="center"/>
<br>
<img src="assets/gifs/acrobot_swingup.gif" style="width: 19%" float="center"/>
<img src="assets/gifs/cheetah_run.gif" style="width: 19%" float="center"/>
<img src="assets/gifs/finger_turn_hard.gif" style="width: 19%" float="center"/>
<img src="assets/gifs/reacher_hard.gif" style="width: 19%" float="center"/>
<img src="assets/gifs/walker_run.gif" style="width: 19%" float="center"/>
</div>
</div>
</div>
</section>
<!-- <hr width="1200" style="margin-left: auto; margin-right: auto"> -->
<section class="section">
<div class="container is-max-desktop">
<!-- Abstract. -->
<div class="columns is-centered has-text-centered">
<div class="column">
<h2 class="title is-3">Abstract</h2>
<div class="content has-text-justified">
<p style="font-size: 125%; width: 83%; height: auto; margin-left: auto; margin-right: auto">
Data-efficiency has always been an essential issue in pixel-based reinforcement learning (RL). As the agent
not only learns decision-making but also meaningful representations from images. The line of reinforcement
learning with data augmentation shows significant improvements in sample-efficiency. However, it is
challenging to guarantee the optimality invariant transformation, that is, the augmented data are readily
recognized as a completely different state by the agent. In the end, we propose a contrastive invariant
transformation (CoIT), a simple yet promising learnable data augmentation combined with standard model-free
algorithms to improve sample-efficiency. Concretely, the differentiable CoIT leverages original samples with
augmented samples and hastens the state encoder for a contrastive invariant embedding. We evaluate our
approach on DeepMind Control Suite and Atari100K. Empirical results verify advances using CoIT, enabling
it to outperform the new state-of-the-art on various tasks. Source code is available at
<a href="https://github.com/mooricAnna/CoIT"><span>https://github.com/mooricAnna/CoIT</span></a>.
</p>
</div>
</div>
</div>
</div>
</section>
<!-- <div style="border:1px solid #CCC; width: 70%; height: auto; margin-left: auto; margin-right: auto"> -->
<hr width="1200" style="margin-left: auto; margin-right: auto">
<!-- Method -->
<section class="section">
<div class="container is-max-widescreen">
<div class="row">
<!-- <div class="rows is-centered has-text-centered"> -->
<div class="rows is-centered">
<h2 class="title is-3">
<span class="dvima">
<!--            -->
Method
</span>
</h2>
<h3 class="title is-4">
<span class="dvima">
<!--               -->
Learnable Invariant Transformation
</span>
</h3>
<div class="content has-text-justified">
<p style="font-size: 110%; margin-left: auto; margin-right: auto">
Our method, <b>CoIT</b> introduces a learnable image augmentation to achieve an invariant image transformation.
We propose a theoretical analysis of how a learnable invariant image transformation can approximate
a stationary distribution over the transformed data by the optimal invariant metric, thus learning
better representations.
</p>
<img src="assets/images/lemma.png" style="width: 49%" float="left"/>
<img src="assets/images/theo_coit.png" style=" width: 49%" float="left"/>
<!-- <span style="font-size: 110%">
<span style="font-weight: bold">Scaling model and data</span>. <i>Top:</i> We compare performance of different methods with model sizes ranging from 2M to 200M parameters. Across all model sizes and generalization levels VIMA outperforms prior works. <i>Bottom:</i> For a fixed model size of 92M parameters we compare the effect of imitation learning dataset size of 0.1%, 1%, 10%, and full imitation data. VIMA is extremely sample efficient and can achieve performance comparable to other methods with 10× less data.
</span> -->
<!-- <br> -->
<h3 class="title is-4">
<span class="dvima">
<!--               -->
Stabilizing Reward Function
</span>
</h3>
<p style="font-size: 110%; margin-left: auto; margin-right: auto">
To further stabilize the reward function, we propose a mixed CoIT that samples multiple
transformed data from the learned distribution, and mix up them together for model training.
</p>
<div class="content has-text-justified">
<img src="assets/images/mixed_coit.png" class="interpolation-image" alt=""
style="display: block; width: 75%; margin-left: auto; margin-right: auto"/>
<h3 class="title is-4">
<span class="dvima">
Framework Design
</span>
</h3>
<p style="font-size: 110%; margin-left: auto; margin-right: auto">
We present a new framework with normalization variants to ensure above discussed learning guarantees by optimizing
paraneters. Below are key points of our method.
<ul style="font-size: 110%">
<li> We borrow a similar idea from <a href="https://arxiv.org/abs/1912.08795"><span>Yin et al.</span></a> to
regularize and smooth the distribution shift between
the transformed and overall data. We use the statistical data stored in the BN layers to approximate
the distribution of the overall data,
<br>
<!-- <div class="row">
<div class="rows is-centered has-text-centered"> -->
<!-- <div class="rows is-centered">
\(
\mathcal{K}_\omega(\mathbf{x}_t')= \sum_{l} \left\|\tilde{\mu}(\mathbf{x}_t')-\mathbb{E}(\tilde{\mu}_l(\mathbf{x})|\mathcal{O}) \right\|_2 +
\sum_{l} \left\|\tilde{\sigma}^2(\mathbf{x}_t')-\mathbb{E}(\tilde{\sigma}_l^2(\mathbf{x})|\mathcal{O}) \right\|_2
\)
</div> -->
<!-- </div> -->
<img src="assets/images/bn.png" class="interpolation-image"
alt="" style="display: block; margin-left: auto; margin-right: auto;
width: 60%; height: auto"/>
where \(\tilde{\mu}(\mathbf{x}_t')\) and \(\tilde{\sigma}^2(\mathbf{x}_t')\) are the <i>mean</i> and
<i>variance</i> of the transformed data and \(\omega\) represents the parameters of the learnable image
transformation. The expectation terms \(\mathbb{E}(\cdot)\) denotes the statistical estimation of the
batch-wise data stored in the \(l\)-th conv layer, and \(\mathcal{O}\) is the given observations.
</li>
<li> We utilize the similarity metric proposed by <a href="https://arxiv.org/abs/2002.05709"><span>Chen et al.</span></a> for learning the encoder that encodes the
observations into latent space to meet the invariant transformation,
<img src="assets/images/byol.png" class="interpolation-image"
alt="" style="display: block; margin-left: auto; margin-right: auto;
width: 70%; height: auto"/>
Here \(\bar{\xi}\) denotes the momentum version of parameters \(\xi\), and \(\mathcal{D}\) indicates the replay buffer.
</li>
<li> We update the critic network with transformed data \(\mathbf{x}_t'\) and \(\mathbf{x}_{t+n}'\) to
minimize the TD error for <i>n</i>-steps returns,
<img src="assets/images/critic.png" class="interpolation-image"
alt="" style="display: block; margin-left: auto; margin-right: auto;
width: 70%; height: auto"/>
</li>
</ul>
</p>
<p style="font-size: 110%; margin-left: auto; margin-right: auto">
Eventually, we give the unified objective function of the CoIT,
where \(\alpha \) and \(\lambda\) are hyper-parameters and the overall architecture is presented below.
<br>
<br>
<img src="assets/images/overall.png" class="interpolation-image"
alt="" style="display: block; margin-left: auto; margin-right: auto;
width: 40%; height: auto"/>
<!-- where \(\alpha \) and \(\lambda\) are hyper-parameters and the overall architecture is presented below. -->
</p>
<img src="assets/images/pull.png" class="interpolation-image"
alt="" style="display: block; margin-left: auto; margin-right: auto;
width: 70%; height: auto"/>
<p style="font-size: 100%; display: block; margin-left: auto; margin-right: auto;
width: 77%; height: auto">
<b>Overall architecture of CoIT.</b> The observations are transformed following a Gaussian distribution
\(\mathcal{G}(\mu, \sigma)\) and encoded by the state encoder \(g_\xi\).
The observation encoder \(g_{\bar{\xi}}\) and projection \(f_{\bar{\xi}}\) are the exponentially moving average
version of the state encoder and projection.
</p>
</div>
</div>
</div>
</div>
</section>
<!-- Model
<section class="section">
<div class="container is-max-widescreen">
<div class="rows">
<div class="rows is-centered ">
<div class="row is-full-width">
<h2 class="title is-3"><span
class="dvima">VIMA: Visuomotor Attention Model</span></h2>
<video poster="" autoplay controls muted loop height="100%">
<source src="assets/videos/vima_arch_animation.mp4"
type="video/mp4">
</video>
<span style="font-size: 110%">
<span style="font-weight: bold">VIMA architecture.</span> We encode the multimodal prompts with a pre-trained T5 model, and condition the robot controller on the prompt through cross-attention layers. The controller is a causal transformer decoder consisting of alternating self and cross attention layers that predicts motor commands conditioned on prompts and interaction history.</span>
</div>
</div>
</div>
</div>
</section> -->
<!-- <hr width="1150" style="margin-left: auto; margin-right: auto">
<section class="section">
<div class="container is-max-widescreen">
<div class="row">
<div class="rows is-centered has-text-centered">
<div class="row is-full-width">
<h2 class="title is-3"><span
class="dvima">VIMA-Bench: Benchmark for Multimodal Robot Learning</span></h2>
<div class="content has-text-justified">
<span style="font-size: 125%">
We provide 17 representative meta-tasks with multimodal prompt templates, which can be procedurally instantiated into thousands of individual tasks by various combinations of textures and tabletop objects.
</span>
<br>
<br>
<br>
<div class="columns">
Simple Object Manipulation
<div class="column has-text-left">
<h3 class="title is-5">Simple Object Manipulation</h3>
<div class="select is-medium">
<select id="simple-object-manipulation-menu-tasks"
onchange="updateDemoVideo('simple-object-manipulation')">
<option value="simple_manipulation" selected="selected">Visual Manipulation</option>
<option value="rotate">Rotate</option>
<option value="scene_understanding">Scene Understanding</option>
</select>
</div>
<div class="select is-medium">
<select id="simple-object-manipulation-menu-instances"
onchange="updateDemoVideo('simple-object-manipulation')">
<option value="1" selected="selected">1</option>
<option value="2">2</option>
<option value="3">3</option>
<option value="4">4</option>
<option value="5">5</option>
</select>
</div>
<video id="simple-object-manipulation-single-task-video"
controls
muted
autoplay
loop
width="100%">
<source src="assets/videos/demos/simple_object_manipulation/simple_manipulation/1.mp4"
type="video/mp4">
</video>
</div>
Visual Goal Reaching
<div class="column has-text-left">
<h3 class="title is-5">Visual Goal Reaching</h3>
<div class="select is-medium">
<select id="visual-goal-reaching-menu-tasks"
onchange="updateDemoVideo('visual-goal-reaching')">
<option value="rearrange" selected="selected">Rearrange</option>
<option value="rearrange_then_restore">Rearrange Then Restore</option>
</select>
</div>
<div class="select is-medium">
<select id="visual-goal-reaching-menu-instances"
onchange="updateDemoVideo('visual-goal-reaching')">
<option value="1" selected="selected">1</option>
<option value="2">2</option>
<option value="3">3</option>
<option value="4">4</option>
<option value="5">5</option>
</select>
</div>
<video id="visual-goal-reaching-single-task-video"
controls
muted
autoplay
loop
width="100%">
<source src="assets/videos/demos/visual_goal_reaching/rearrange/1.mp4"
type="video/mp4">
</video>
</div>
</div>
<br>
<div class="columns">
Novel Concept Grounding
<div class="column has-text-left">
<h3 class="title is-5">Novel Concept Grounding</h3>
<div class="select is-medium">
<select id="novel-concept-grounding-menu-tasks"
onchange="updateDemoVideo('novel-concept-grounding')">
<option value="novel_adj_and_noun" selected="selected">Novel Adjective and Noun
</option>
<option value="novel_adj">Novel Adjective</option>
<option value="novel_noun">Novel Noun</option>
<option value="twist">Twist</option>
</select>
</div>
<div class="select is-medium">
<select id="novel-concept-grounding-menu-instances"
onchange="updateDemoVideo('novel-concept-grounding')">
<option value="1" selected="selected">1</option>
<option value="2">2</option>
<option value="3">3</option>
<option value="4">4</option>
<option value="5">5</option>
</select>
</div>
<video id="novel-concept-grounding-single-task-video"
controls
muted
autoplay
loop
width="100%">
<source src="assets/videos/demos/novel_concept_grounding/novel_adj_and_noun/1.mp4"
type="video/mp4">
</video>
</div>
One Shot Video Imitation
<div class="column has-text-left">
<h3 class="title is-5">One-shot Video Imitation</h3>
<div class="select is-medium">
<select id="one-shot-video-imitation-menu-tasks"
onchange="updateDemoVideo('one-shot-video-imitation')">
<option value="follow_motion" selected="selected">Follow Motion</option>
<option value="follow_order">Follow Order</option>
</select>
</div>
<div class="select is-medium">
<select id="one-shot-video-imitation-menu-instances"
onchange="updateDemoVideo('one-shot-video-imitation')">
<option value="1" selected="selected">1</option>
<option value="2">2</option>
<option value="3">3</option>
<option value="4">4</option>
<option value="5">5</option>
</select>
</div>
<video id="one-shot-video-imitation-single-task-video"
controls
muted
autoplay
loop
width="100%">
<source src="assets/videos/demos/one_shot_video_imitation/follow_motion/1.mp4"
type="video/mp4">
</video>
</div>
</div>
<br>
<div class="columns">
Visual Constraint Satisfaction
<div class="column has-text-left">
<h3 class="title is-5">Visual Constraint Satisfaction</h3>
<div class="select is-medium">
<select id="visual-constraint-satisfaction-menu-tasks"
onchange="updateDemoVideo('visual-constraint-satisfaction')">
<option value="sweep_without_exceeding" selected="selected">Sweep without
Exceeding
</option>
<option value="sweep_without_touching">Sweep without Touching</option>
</select>
</div>
<div class="select is-medium">
<select id="visual-constraint-satisfaction-menu-instances"
onchange="updateDemoVideo('visual-constraint-satisfaction')">
<option value="1" selected="selected">1</option>
<option value="2">2</option>
<option value="3">3</option>
<option value="4">4</option>
<option value="5">5</option>
</select>
</div>
<video id="visual-constraint-satisfaction-single-task-video"
controls
muted
autoplay
loop
playbackRate="2.0"
width="100%">
<source src="assets/videos/demos/visual_constraint_satisfaction/sweep_without_exceeding/1.mp4"
type="video/mp4">
</video>
</div>
Visual Reasoning
<div class="column has-text-left">
<h3 class="title is-5">Visual Reasoning</h3>
<div class="select is-medium">
<select id="visual-reasoning-menu-tasks" onchange="updateDemoVideo('visual-reasoning')">
<option value="manipulate_old_neighbor" selected="selected">Manipulate Old
Neighbor
</option>
<option value="pick_in_order_then_restore">Pick in Order Then Restore</option>
<option value="same_color">Same Texture</option>
<option value="same_profile">Same Shape</option>
</select>
</div>
<div class="select is-medium">
<select id="visual-reasoning-menu-instances"
onchange="updateDemoVideo('visual-reasoning')">
<option value="1" selected="selected">1</option>
<option value="2">2</option>
<option value="3">3</option>
<option value="4">4</option>
<option value="5">5</option>
</select>
</div>
<video id="visual-reasoning-single-task-video"
controls
muted
autoplay
loop
width="100%">
<source src="assets/videos/demos/visual_reasoning/manipulate_old_neighbor/1.mp4"
type="video/mp4">
</video>
</div>
</div>
</div>
</div>
</div>
</div>
</section> -->
<hr width="1200" style="margin-left: auto; margin-right: auto">
<!--Experiments-->
<section class="section">
<div class="container is-max-widescreen">
<div class="rows">
<div class="rows is-centered">
<div class="row is-full-width">
<h2 class="title is-3">
<span class="dvima">
<!--            -->
Experiments
</span>
</h2>
<div class="content has-text-justified">
<p style="font-size: 125%">
We benchmark our method on the DeepMind control suite and compare CoIT with prior model-free methods which
use data augmentation and contrastive learning to improve data-efficiency:
<ul style="font-size: 125%">
<li> <a href="https://arxiv.org/abs/2107.09645"><span>DrQ-v2</span></a>: introducing a simple data augmentation called <i>random shift</i> into a
fine-tuned DDPG.
</li>
<li> <a href="https://arxiv.org/abs/2004.04136"><span>CURL</span></a>: combining SAC with a self-supervised learning framework.
</li>
<li> Vanilla SAC and DDPG that directly trained from the image input
</li>
</ul>
</p>
<p style="font-size: 125%">
We also present ablation studies to show the details of our method.
</p>
<!-- <br> -->
<!-- <br> -->
<h3 class="title is-4"><span
class="dvima">
<!--            -->
Evaluation Results</span></h3>
<div class="content has-text-justified">
<img src="assets/images/dmc_results.png" class="interpolation-image"
alt="" style="display: block; margin-left: auto; margin-right: auto"/>
<br>
<span style="font-size: 110%">
<b>Results of 8 complex tasks in DMControl.</b> These tasks are chosen to offer multiple degrees of challenges,
including complex dynamics, sparse rewards, hard exploration, and more. Below are key findings:
<!-- (i) CoIT outperforms
vanilla DDPG and SAC in a wide range. -->
(i) Although DrQ-v2 has already performed remarkably for continuous control, CoIT is
more data-efficient on multiple tasks.
(ii) From general trends of the learning curves,
CoIT improves or keeps the data-efficiency in a more stable manner which is not trivial on DMControl tasks.
</span>
<br>
<!-- <br> -->
<!-- <br> -->
<h3 class="title is-4"><span
class="dvima">
<!--            -->
Ablation Studies </span></h3>
<div class="content has-text-justified">
<br>
<img src="assets/images/heatmap.png" class="interpolation-image"
alt="" style="display: block; margin-left: auto; margin-right: auto"/>
<br>
<span style="font-size: 110%">
<span style="font-weight: bold">Saliency maps for CoIT on 6 tasks:</span>
(a) Cartpole Swingup Sparse,
(b) Hopper Stand,
(c) Walker Run,
(d) Cheetah Run,
(e) Quadruped Walk, and
(f) Finger Turn Hard.
These saliency maps demonstrate that CoIT is beneficial for the agent to focus on task-relevant elements
like the whole robot body and ignore the task-irrelevant information like the floor and background.
Especially in <i>Finger Turn Hard</i>, the lightest part in the saliency map is a red ball in the
observation, which is highly related to the reward.
</span>
<br>
<br>
<br>
<!-- <br> -->
<img src="assets/images/gaussian_dis.png" class="interpolation-image"
alt="" style="display: block; margin-left: auto; margin-right: auto"/>
<br>
<span style="font-size: 110%">
<span style="font-weight: bold">Visualization of the parameters of the Gaussian distribution for image transformation. </span>
These curves demonstrate that the gaussian distribution proposed in CoIT
could automatically find an appropriate transformation to smooth the distribution shift between the different views
of the same observation, therefore being beneficial to the representation learning.
</div>
</div>
</div>
</div>
</section>
<hr width="1200" style="margin-left: auto; margin-right: auto">
<!--Conclusion-->
<section class="section">
<div class="container is-max-widescreen">
<div class="row">
<div class="rows is-centered">
<div class="row is-full-width">
<h2 class="title is-3">
<span class="dvima">
<!--            -->
Conclusion
</span></h2>
<div class="content has-text-justified">
<p style="font-size: 125%">
A novel pixel transformation CoIT under model-free RL algorithms that significantly improves the
data-efficiency and stability for visual tasks is introduced in this work. We theoretically analyze how
the learnable transformation constrains the distribution of the abstracted data, and dissect its benefits
to representation learning. CoIT is no need for any additional modifications to the backbone RL
algorithm and is easy to implement. We compare CoIT to SOTA methods on popular benchmarks and
certify that it gains promising performance with advanced stability. Hopefully, contrastive invariant
transformation can lead to a new branch for representation learning in RL.
</p>
</div>
</div>
</div>
</div>
</section>
<hr width="1200" style="margin-left: auto; margin-right: auto">
<section class="section" id="BibTeX">
<div class="row">
<div class="rows is-centered">
<div class="row is-full-width">
<div class="container is-max-widescreen content">
<h2 class="title">BibTeX</h2>
<pre>
<code style="font-size: 100%; margin-left: auto; margin-right: auto;
width: auto; height: auto">
@inproceedings{liudata,
title={On the Data-Efficiency with Contrastive Image Transformation in Reinforcement Learning},
author={Liu, Sicong and Zhang, Xi Sheryl and Li, Yushuo and Zhang, Yifan and Cheng, Jian},
booktitle={The Eleventh International Conference on Learning Representations}
}
</code>
</pre>
</div>
</section>
<footer class="footer">
<div class="container">
<div class="columns is-centered has-text-centered">
<div class="column">
<div class="content">
<p>
Website template borrowed from <a
href="https://github.com/nerfies/nerfies.github.io">NeRFies</a> and <a
href="https://github.com/cliport/cliport.github.io">CLIPort</a>.
</p>
</div>
</div>
</div>
</div>
</footer>
</body>
</html>