-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathindex.html
849 lines (801 loc) · 49.2 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
<!DOCTYPE html>
<html>
<head>
<meta charset="utf-8">
<title>OpenVLA: An Open-Source Vision-Language-Action Model</title>
<meta name="description" content="OpenVLA: An Open-Source Vision-Language-Action Model">
<meta name="keywords" content="OpenVLA">
<meta name="viewport" content="width=device-width, initial-scale=1">
<meta property="og:title" content="OpenVLA: An Open-Source Vision-Language-Action Model">
<meta property="og:type" content="website">
<meta property="og:site_name" content="OpenVLA: An Open-Source Vision-Language-Action Model">
<meta property="og:image" content="https://openvla.github.io/static/images/teaser.png" />
<meta property="og:image:type" content="image/png" />
<meta property="og:image:width" content="1939" />
<meta property="og:image:height" content="772" />
<meta property="og:url" content="https://openvla.github.io/" />
<meta property="og:description" content="OpenVLA: An Open-Source Vision-Language-Action Model" />
<meta name="twitter:title" content="OpenVLA: An Open-Source Vision-Language-Action Model" />
<meta name="twitter:description" content="OpenVLA: An Open-Source Vision-Language-Action Model" />
<meta name="twitter:image" content="https://openvla.github.io/static/images/teaser.png" />
<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/icon.png">
<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>
<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">OpenVLA:<br><span style="font-size:2.4rem;">An Open-Source Vision-Language-Action Model</span></h1>
<div class="is-size-5 publication-authors">
<span class="author-block">
<a href="https://moojink.com/">Moo Jin Kim</a><sup>*,1</sup>,
</span>
<span class="author-block">
<a href="https://kpertsch.github.io/">Karl Pertsch</a><sup>*,1,2</sup>,
</span>
<span class="author-block">
<a href="https://www.siddkaramcheti.com/">Siddharth Karamcheti</a><sup>*,1,3</sup>,
</span>
<br>
<span class="author-block">
<a href="https://tedxiao.me/">Ted Xiao</a><sup>4</sup>,
</span>
<span class="author-block">
<a href="https://abalakrishna123.github.io/">Ashwin Balakrishna</a><sup>3</sup>,
</span>
<span class="author-block">
<a href="https://suraj-nair-1.github.io/">Suraj Nair</a><sup>3</sup>,
</span>
<span class="author-block">
<a href="https://rmrafailov.github.io/">Rafael Rafailov</a><sup>1</sup>,
</span>
<span class="author-block">
<a href="https://www.linkedin.com/in/ethan-paul-foster">Ethan Foster</a><sup>1</sup>,
</span>
<span class="author-block">
<a href="https://www.linkedin.com/in/grace-lam/">Grace Lam</a>,
</span>
<br>
<span class="author-block">
<a href="https://www.linkedin.com/in/pannag-sanketi-bb91142">Pannag Sanketi</a><sup>4</sup>,
</span>
<span class="author-block">
<a href="https://quanvuong.github.io/">Quan Vuong</a><sup>5</sup>,
</span>
<span class="author-block">
<a href="https://aicenter.stanford.edu/people/thomas-kollar">Thomas Kollar</a><sup>3</sup>,
</span>
<span class="author-block">
<a href="https://www.benburchfiel.com/">Benjamin Burchfiel</a><sup>3</sup>,
</span>
<span class="author-block">
<a href="https://groups.csail.mit.edu/locomotion/russt.html">Russ Tedrake</a><sup>3,6</sup>,
</span>
<span class="author-block">
<a href="https://dorsa.fyi/">Dorsa Sadigh</a><sup>1</sup>,
</span>
<span class="author-block">
<a href="https://people.eecs.berkeley.edu/~svlevine/">Sergey Levine</a><sup>2</sup>,
</span>
<span class="author-block">
<a href="https://cs.stanford.edu/~pliang/">Percy Liang</a><sup>1</sup>,
</span>
<span class="author-block">
<a href="https://ai.stanford.edu/~cbfinn/">Chelsea Finn</a><sup>1</sup>,
</span>
</div>
<div class="is-size-5 publication-authors">
<span class="author-block"><sup><font size="-0.4">*</sup>Equal contribution</font></span><br>
<span class="author-block"><sup>1</sup>Stanford University,</span>
<span class="author-block"><sup>2</sup>UC Berkeley,</span>
<span class="author-block"><sup>3</sup>Toyota Research Institute,</span>
<span class="author-block"><sup>4</sup>Google DeepMind,</span>
<span class="author-block"><sup>5</sup>Physical Intelligence,</span>
<span class="author-block"><sup>6</sup>MIT,</span>
</div>
<div class="column has-text-centered">
<div class="publication-links">
<!-- PDF Link. -->
<span class="link-block">
<a href="https://arxiv.org/abs/2406.09246"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fas fa-file-pdf"></i>
</span>
<span>Paper</span>
</a>
</span>
<!-- Video Link. -->
<!-- <span class="link-block">
<a href="https://www.youtube.com/watch?v=SIcPxapIgBI" 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> -->
<!-- Code Link. -->
<span class="link-block">
<a href="https://github.com/openvla/openvla"
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>
<!-- Model Link. -->
<span class="link-block">
<a href="https://huggingface.co/openvla"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<img src="static/images/hf_icon.svg" />
</span>
<span>Models</span>
</a>
</span>
<!-- CoLab Link. -->
<!-- <span class="link-block">
<a href="https://colab.research.google.com/github/openvla/openvla/blob/main/examples/openvla.ipynb"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<img src="static/images/colab_icon.png" />
</span>
<span>Let's try it!</span>
</a>
</span> -->
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<section class="hero teaser teaser-video">
<div class="container is-max-desktop has-text-centered">
<div class="hero-body">
<video id="teaser" autoplay muted loop playsinline width="80%">
<source src="static/videos/openvla_teaser_video.mp4" type="video/mp4">
</video>
</div>
</div>
</section>
<section class="hero teaser">
<div class="container is-max-desktop">
<div class="hero-body">
<!-- <video id="teaser" autoplay muted loop playsinline height="100%">
<source src=""
type="video/mp4">
</video> -->
<img src="static/images/openvla_teaser.jpg" />
<h2 class="subtitle has-text-centered">
We introduce OpenVLA, a 7B parameter open-source vision-language-action model (VLA), pretrained on 970k robot episodes from the Open X-Embodiment dataset.
OpenVLA sets a new state of the art for generalist robot manipulation policies. It supports controlling multiple robots out of the box and
can be quickly adapted to new robot setups via parameter-efficient fine-tuning. The OpenVLA checkpoints and PyTorch training pipeline are fully open-source
and models can be downloaded and fine-tuned from HuggingFace.
</h2>
</div>
</div>
</section>
<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-steve has-text-centered">
<video poster="" id="steve video" autoplay controls muted loop playsinline height="100%">
<source src="static/videos/carousel/bridge_pick_clutter_2.mp4" type="video/mp4">
</video>
<p id="overlay">Bridge Put Corn on Plate</p>
</div>
<div class="item item-steve has-text-centered">
<video poster="" id="steve video" autoplay controls muted loop playsinline height="100%">
<source src="static/videos/carousel/rt1_robot_coke_upright.mp4" type="video/mp4">
</video>
<p id="overlay">Google Place Coke Upright</p>
</div>
<div class="item item-chair-tp has-text-centered">
<video poster="" id="chair-tp video" autoplay controls muted loop playsinline height="100%">
<source src="static/videos/carousel/franka_pour_corn.mp4" type="video/mp4">
</video>
<p id="overlay">Franka Pour Corn in Pot (4x)</p>
</div>
<div class="item item-chair-tp has-text-centered">
<video poster="" id="chair-tp video" autoplay controls muted loop playsinline height="100%">
<source src="static/videos/carousel/wipe_ood_4x.mp4" type="video/mp4">
</video>
<p id="overlay">Franka Wipe Table (4x)</p>
</div>
<div class="item item-fullbody has-text-centered">
<video poster="" id="fullbody video" autoplay controls muted loop playsinline height="100%">
<source src="static/videos/carousel/bridge_pick_clutter.mp4" type="video/mp4">
</video>
<p id="overlay">Bridge Put Eggplant in Bowl</p>
</div>
<div class="item item-shiba has-text-centered">
<video poster="" id="shiba video" autoplay controls muted loop playsinline height="100%">
<source src="static/videos/carousel/franka_cover.mp4" type="video/mp4">
</video>
<p id="overlay">Franka Cover Pink Bowl (4x)</p>
</div>
<div class="item item-steve has-text-centered">
<video poster="" id="steve video" autoplay controls muted loop playsinline height="100%">
<source src="static/videos/carousel/rt1_robot_orange_near_coke.mp4" type="video/mp4">
</video>
<p id="overlay">Google Move Orange near Coke</p>
</div>
<div class="item item-fullbody has-text-centered">
<video poster="" id="fullbody video" autoplay controls muted loop playsinline height="100%">
<source src="static/videos/carousel/wipe_ood_2_4x.mp4" type="video/mp4">
</video>
<p id="overlay">Franka Wipe Table (4x)</p>
</div>
<div class="item item-fullbody has-text-centered">
<video poster="" id="fullbody video" autoplay controls muted loop playsinline height="100%">
<source src="static/videos/carousel/bridge_stack.mp4" type="video/mp4">
</video>
<p id="overlay">Bridge Stack Cups (2x)</p>
</div>
<div class="item item-fullbody has-text-centered">
<video poster="" id="fullbody video" autoplay controls muted loop playsinline height="100%">
<source src="static/videos/carousel/franka_flip_pot.mp4" type="video/mp4">
</video>
<p id="overlay">Franka Flip Pot (4x)</p>
</div>
<div class="item item-fullbody has-text-centered">
<video poster="" id="fullbody video" autoplay controls muted loop playsinline height="100%">
<source src="static/videos/carousel/franka_knock.mp4" type="video/mp4">
</video>
<p id="overlay">Franka Knock over Yellow Pony (2x)</p>
</div>
</div>
<br>
<p class="has-text-centered">WidowX & Google robot videos show real <b>"zero-shot"</b> rollouts with the OpenVLA model<br>Franka Panda robot videos depict <b>fine-tuned</b> OpenVLA policies</p>
</div>
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column is-full-width">
<h2 class="title is-3">The OpenVLA Model</h2>
<div class="content has-text-justified has-text-centered">
<img src="static/images/openvla_model.jpg" />
<p>
We train OpenVLA by fine-tuning a pretrained <a href="https://arxiv.org/abs/2402.07865">Prismatic-7B VLM</a>. Our model consists of three key elements:
(1) a <strong>fused visual encoder</strong>, consisting of a SigLIP and a DinoV2 backbone, that maps image inputs to a number of ``image patch embeddings'',
(2) a <strong>projector</strong> that takes the output embeddings of the visual encoder and maps them into the input space of a large language model, and
(3) a Llama 2 7B <strong>language model backbone</strong> that predicts tokenized output actions. These tokens get decoded into continuous output actions that
can be directly executed on the robot.
</p>
<p>
To train OpenVLA, we curate a dataset of 970k robot manipulation trajectories from the <a href="https://robotics-transformer-x.github.io/">Open X-Embodiment (OpenX) dataset</a>.
Our dataset spans a wide range of tasks, scenes and robot embodiments. We train OpenVLA on a cluster of 64 A100 GPUs for 15 days. The trained model checkpoints
can be <a href="https://huggingface.co/openvla">downloaded from HuggingFace</a> and used with a few lines of code.
</p>
<!-- <h3 class="title is-4">Metrics for Real-to-Sim Evaluation</h3>
<img src="static/images/metrics.png" />
<p>
An effective & useful simulation-based evaluation should demonstrate <strong>good correlations in policy ranking & performance</strong> with real evaluations.
</p>
<p>
To measure such correlations, one can apply the traditional <strong>Pearson correlation metric ("r")</strong>, but it has the following limitations: (1) Pearson correlation only assess
the linear fit between real-and-sim performances, while for simulated evaluation we don't necessarily need linear correlations,
as long as sim eval reflects real-world performance improvements between different policies (middle-right); (2) Pearson correlation does not reflect the range of values it is computed over.
For policy sets that perform closely in real (far-right), Pearson r may change drastically based on small real-world performance differences,
which can often be attributed to the inherent noise in real-world evaluations.
</p>
<p>
Thus, we introduce the <strong>Mean Maximum Rank Violation (MMRV)</strong> metric (lower the better)
to better assess the real-and-sim policy ranking consistency.
The key underlying quantity is the rank violation between two policies, which weighs the significance of the
simulator incorrectly ranking the policies by the corresponding margin in real-world performance.
MMRV then aggregates the N^2 rank violations by averaging the worst-case rank violation for each policy.
</p>
<h3 class="title is-4">Visual Matching Mitigates the Real-to-Sim Visual Gap</h3>
<img src="static/images/visual_matching.png" style="width: 70%; height: auto; display: block; margin: 0 auto;"/>
<p>
Visual discrepancies between real-world and simulated environments can comprise a distribution shift that adversely
affects a learned policy’s behavior, rendering simulated evaluation unreliable. Our goal is to match the simulator
visuals to those of the real-world environment with only a modest amount of manual effort. Our proposed Visual Matching
consists of (1) <strong>green screening</strong>, i.e. segmenting out interactive simulated assets and overlaying them onto real-world
backgrounds; and (2) <strong>texture matching</strong>, which involves projecting real object textures onto simulation assets and tuning
robot arm colors using real videos.
</p>
<h3 class="title is-4">System Identification Mitigates the Real-to-Sim Control Gap</h3>
<img src="static/images/control_gap.png" />
<p>
The goal of mitigating the control gap between simulated and real-world environments is to ensure that policy actions
executed in simulation yields comparable effects on the robot’s end-effector as those observed when executed on the real
robot. We perform system identification (SysID) for closing the control gap between real and simulated environments on a small sample of trajectories from the real world dataset.
</p>
<div class="columns is-vcentered interpolation-panel">
<div class="column has-text-centered">
<video autoplay controls muted loop playsinline height="100%">
<source src="static/videos/sysid/real_rollout.mp4" type="video/mp4">
</video>
<p >Real World Rollout</p>
</div>
<div class="column has-text-centered">
<video autoplay controls muted loop playsinline height="100%">
<source src="static/videos/sysid/bad_control.mp4" type="video/mp4">
</video>
<p >Control without SysID</p>
</div>
<div class="column has-text-centered">
<video autoplay controls muted loop playsinline height="100%">
<source src="static/videos/sysid/good_control.mp4" type="video/mp4">
</video>
<p >Control with SysID</p>
</div>
</div> -->
</div>
</div>
</div>
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column is-full-width">
<h2 class="title is-3">Experiments</h2>
<h3 class="title is-4">Direct Evaluations on Multiple Robot Platforms</h3>
<img src="static/images/bridge_results.jpg" />
<div class="content has-text-justified">
<table width="100%" align="center" border="0" cellspacing="0" cellpadding="0">
<tr>
<p>
We evaluate OpenVLA's ability to control multiple robot platforms ``out-of-the-box'' across two setups: the WidowX setup
from Bridge V2 and the Google Robot from the RT-series of papers. Our results show that OpenVLA sets a new state of the art,
outperforming prior generalist policies RT-1-X and Octo. Notably, as a product of the added data diversity and new model
components, it also outperforms RT-2-X, a 55B parameter closed VLA.
</p>
</tr>
<tr>
<td width="50%">
<p>
We test OpenVLA across a wide range of generalization tasks, such as
<strong>visual</strong> (unseen backgrounds, distractor objects, colors/appearances of objects);
<strong>motion</strong> (unseen object positions/orientations);
<strong>physical</strong> (unseen object sizes/shapes);
and <strong>semantic</strong> (unseen target objects, instructions, and concepts from the Internet) generalization.
Qualitatively, we find that both RT-2-X and OpenVLA exhibit markedly more robust behaviors than the other tested model,
such as approaching the correct object when distractor objects are present, properly orienting the robot's end-effector
to align with the orientation of the target object, and even recovering from mistakes such as insecurely grasping objects
</p>
</td>
<td width="50%">
<img src="static/images/rt1_results.jpg">
</td>
</tr>
</table>
</div>
<br>
<h3 class="title is-4">Data-Efficient Adaptation to New Robot Setups</h3>
<div class="content has-text-justified">
<img src="static/images/finetune_results.jpg" />
<table width="100%" align="center" border="0" cellspacing="0" cellpadding="0">
<tr>
<p>
Effective fine-tuning of VLA models to new tasks and robot setups is largely unexplored,
yet is key for their widespread adoption. We investigate OpenVLA’s ability to be quickly adapted to a new robot setup
in two domains: Franka-Tabletop, a stationary, table-mounted Franka Emika Panda 7-DoF robot arm, controlled
at a frequency of 5 Hz; and Franka-DROID, the Franka robot arm setup from the recently released <a href="https://droid-dataset.github.io/">DROID dataset</a>, controlled at 15 Hz.
</p>
</tr>
<tr>
<p>
We compare to Diffusion Policy, a state of the art data-efficient imitation learning approach, trained from scratch.
Additionally, we evaluate Octo fine-tuned on the target dataset. OpenVLA clearly outperforms Octo across most tasks.
Diffusion policy is strongest on narrower, more precise tasks, while OpenVLA shows better performance on tasks that
require grounding language to behavior in multi-task, multi-object settings.
OpenVLA is the only approach that achieves at least 50% success rate across all tested
tasks, suggesting that it can be a strong default option for imitation learning tasks, particularly if they
involve a diverse set of language instructions.
</p>
</tr>
</table>
</div>
<br>
<h3 class="title is-4">Parameter-Efficient Fine-Tuning</h3>
<div class="content has-text-justified">
<img src="static/images/lora_table.png">
<table width="100%" align="center" border="0" cellspacing="0" cellpadding="0">
<tr>
<p>
We test various approaches for parameter-efficient fine-tuning of OpenVLA policies across multiple Franka-Tabletop tasks.
We find that only fine-tuning the network’s last layer or freezing the vision encoder leads to poor performance.
LoRA achieves the best trade-off between performance and training memory consumption, matching
full fine-tuning performance while fine-tuning only 1.4% of the parameters.
</p>
</tr>
</table>
</div>
<br>
<br>
<h3 class="title is-4">Sample OpenVLA Rollout Videos</h2>
<div class="content has-text-justified">
<p>
Below are videos of the OpenVLA generalist policy demonstrating various robust behaviors. (Videos are sped up by 1.5x.)
</p>
</div>
<h4 class="title is-6">Real-World Bridge V2 WidowX Robot Rollouts</h4>
<div class="content has-text-justified">
<p>In a scene with many distractor objects, OpenVLA is able to approach and manipulate the correct target object.</p>
</div>
<div class="columns is-vcentered interpolation-panel">
<div class="column has-text-centered">
<h5>Put Eggplant into Pot</h5>
<video autoplay controls muted loop playsinline width="49%">
<source src="static/videos/qualitative_results/correct_target/openvla--put_eggplant_into_pot--clutter.mp4" type="video/mp4">
</video>
<video autoplay controls muted loop playsinline width="49%">
<source src="static/videos/qualitative_results/correct_target/openvla--put_eggplant_into_pot--clutter--2.mp4" type="video/mp4">
</video>
</div>
<div class="column has-text-centered">
<h5>Put Yellow Corn on Pink Plate</h5>
<video autoplay controls muted loop playsinline width="49%">
<source src="static/videos/qualitative_results/correct_target/openvla--put_corn_on_plate--clutter.mp4" type="video/mp4">
</video>
<video autoplay controls muted loop playsinline width="49%">
<source src="static/videos/qualitative_results/correct_target/openvla--put_corn_on_plate--clutter--2.mp4" type="video/mp4">
</video>
</div>
</div>
<div class="content has-text-justified">
Similarly, we assess the policy's language grounding by prompting it to manipulate different target objects given the same initial states. We find that OpenVLA reliably targets the correct object in most cases.</p>
</div>
<div class="columns is-vcentered interpolation-panel">
<div class="column has-text-centered">
<h5>Lift Red Chili Pepper</h5>
<video autoplay controls muted loop playsinline width="100%">
<source src="static/videos/qualitative_results/good_lang_cond/openvla--lift_red_chili_pepper.mp4" type="video/mp4">
</video>
</div>
<div class="column has-text-centered">
<h5>Lift Cheese</h5>
<video autoplay controls muted loop playsinline width="100%">
<source src="static/videos/qualitative_results/good_lang_cond/openvla--lift_cheese.mp4" type="video/mp4">
</video>
</div>
<div class="column has-text-centered">
<h5>Put Pink Cup on Plate</h5>
<video autoplay controls muted loop playsinline width="100%">
<source src="static/videos/qualitative_results/good_lang_cond/openvla--put_pink_cup_on_plate.mp4" type="video/mp4">
</video>
</div>
<div class="column has-text-centered">
<h5>Put Blue Cup on Plate</h5>
<video autoplay controls muted loop playsinline width="100%">
<source src="static/videos/qualitative_results/good_lang_cond/openvla--put_blue_cup_on_plate.mp4" type="video/mp4">
</video>
</div>
</div>
<div class="content has-text-justified">
<p>OpenVLA properly orients the robot's end-effector to align with the orientation of the target object before grasping it.</p>
</div>
<div class="columns is-vcentered interpolation-panel">
<div class="column has-text-centered">
<h5>Lift AAA Battery</h5>
<video autoplay controls muted loop playsinline width="80%">
<source src="static/videos/qualitative_results/good_orient/openvla--lift_AAA_battery.mp4" type="video/mp4">
</video>
</div>
<div class="column has-text-centered">
<h5>Put Carrot on Plate</h5>
<video autoplay controls muted loop playsinline width="80%">
<source src="static/videos/qualitative_results/good_orient/openvla--put_corn_on_plate--motion.mp4" type="video/mp4">
</video>
</div>
</div>
<div class="content has-text-justified">
<p>In some cases, after an initial mistake, OpenVLA can even recover and successfully complete the task.</p>
</div>
<div class="columns is-vcentered interpolation-panel">
<div class="column has-text-centered">
<h5>Put Blue Cup on Plate</h5>
<video autoplay controls muted loop playsinline width="80%">
<source src="static/videos/qualitative_results/reattempt/openvla--put_blue_cup_on_plate--reattempt.mp4" type="video/mp4">
</video>
</div>
<div class="column has-text-centered">
<h5>Put Eggplant into Pot</h5>
<video autoplay controls muted loop playsinline width="80%">
<source src="static/videos/qualitative_results/reattempt/openvla--put_eggplant_into_pot--clutter--reattempt.mp4" type="video/mp4">
</video>
</div>
</div>
<br><br>
<h3 class="title is-4">Comparisons with State-of-the-Art Models</h2>
<div class="content has-text-justified">
<p>
Here we show how OpenVLA compares with other baseline methods in various evaluation tasks. (Videos are sped up by 2x.)
</p>
</div>
<h4 class="title is-6">Real-World Google Robot Rollouts</h4>
<div class="content has-text-justified">
<p>Both RT-2-X (closed-source 55B-parameter model) and OpenVLA perform reliably on in-distribution and basic out-of-distribution (OOD) generalization tasks.</p>
</div>
<div class="columns is-vcentered interpolation-panel">
<div class="column has-text-centered">
<h5><b>RT-2-X:</b><br>Pick Coke Can<br>(In-Distribution)</h5>
<h5 style="font-size:30px;">✅</h5>
<video autoplay controls muted loop playsinline width="80%">
<source src="static/videos/comparisons_with_baselines/rt1_robot/rt2x--pick_coke_can.mp4" type="video/mp4">
</video>
</div>
<div class="column has-text-centered">
<h5><b>OpenVLA:</b><br>Pick Coke Can<br>(In-Distribution)</h5>
<h5 style="font-size:30px;">✅</h5>
<video autoplay controls muted loop playsinline width="80%">
<source src="static/videos/comparisons_with_baselines/rt1_robot/openvla--pick_coke_can.mp4" type="video/mp4">
</video>
</div>
</div>
<div class="columns is-vcentered interpolation-panel">
<div class="column has-text-centered">
<h5><b>RT-2-X:</b><br>Move Banana near Plate<br>(OOD: unseen target object)</h5>
<h5 style="font-size:30px;">✅</h5>
<video autoplay controls muted loop playsinline width="80%">
<source src="static/videos/comparisons_with_baselines/rt1_robot/rt2x--move_banana_near_plate.mp4" type="video/mp4">
</video>
</div>
<div class="column has-text-centered">
<h5><b>OpenVLA:</b><br>Move Banana near Plate<br>(OOD: unseen target object)</h5>
<h5 style="font-size:30px;">✅</h5>
<video autoplay controls muted loop playsinline width="80%">
<source src="static/videos/comparisons_with_baselines/rt1_robot/openvla--move_banana_near_plate.mp4" type="video/mp4">
</video>
</div>
</div>
<div class="columns is-vcentered interpolation-panel">
<div class="column has-text-centered">
<h5><b>RT-2-X:</b><br>Place Banana on Plate<br>(OOD: unseen target object & instruction)</h5>
<h5 style="font-size:30px;">✅</h5>
<video autoplay controls muted loop playsinline width="80%">
<source src="static/videos/comparisons_with_baselines/rt1_robot/rt2x--place_banana_on_plate.mp4" type="video/mp4">
</video>
</div>
<div class="column has-text-centered">
<h5><b>OpenVLA:</b><br>Place Banana on Plate<br>(OOD: unseen target object & instruction)<br></h5>
<h5 style="font-size:30px;">✅</h5>
<video autoplay controls muted loop playsinline width="80%">
<source src="static/videos/comparisons_with_baselines/rt1_robot/openvla--place_banana_on_plate.mp4" type="video/mp4">
</video>
</div>
</div>
<br>
<div class="content has-text-justified">
<p>However, RT-2-X performs better than OpenVLA on difficult semantic generalization tasks, i.e., tasks that require knowledge of concepts from the Internet that do not appear in the robot action training data, such as Taylor Swift in the videos below.<br><br>This is expected given that RT-2-X uses larger-scale Internet pretraining data and is co-fine-tuned with both robot action data and Internet pretraining data to better preserve the pretraining knowledge (for OpenVLA, we fine-tune the pretrained vision-language model solely on robot action data for simplicity).</p>
</div>
<div class="columns is-vcentered interpolation-panel">
<div class="column has-text-centered">
<h5><b>RT-2-X:</b><br>Move Coke Can near Taylor Swift<br>(OOD: unseen concept from Internet)</h5>
<h5 style="font-size:30px;">✅</h5>
<video autoplay controls muted loop playsinline width="80%">
<source src="static/videos/comparisons_with_baselines/rt1_robot/rt2x--move_coke_can_near_taylor_swift.mp4" type="video/mp4">
</video>
</div>
<div class="column has-text-centered">
<h5><b>OpenVLA:</b><br>Move Coke Can near Taylor Swift<br>(OOD: unseen concept from Internet)</h5>
<h5 style="font-size:30px;">❌</h5>
<video autoplay controls muted loop playsinline width="80%">
<source src="static/videos/comparisons_with_baselines/rt1_robot/openvla--move_coke_can_near_taylor_swift.mp4" type="video/mp4">
</video>
</div>
</div>
<br><br>
<h4 class="title is-6">Real-World Franka Emika Panda Robot Rollouts: Fine-Tuning on Franka-Tabletop Datasets</h4>
<div class="content has-text-justified">
<p>In narrow single-instruction fine-tuning tasks, Diffusion Policy trained from scratch outperforms fine-tuned generalist policies, Octo and OpenVLA. (Videos are sped up by 5x.)</p>
</div>
<div class="columns is-vcentered interpolation-panel">
<div class="column has-text-centered">
<h5><b>Diffusion Policy:</b><br>Put Carrot in Bowl</h5>
<h5 style="font-size:30px;">✅</h5>
<video autoplay controls muted loop playsinline width="80%">
<source src="static/videos/comparisons_with_baselines/franka_tabletop/dp--put_carrot.mp4" type="video/mp4">
</video>
</div>
<div class="column has-text-centered">
<h5><b>Octo:</b><br>Put Carrot in Bowl</h5>
<h5 style="font-size:30px;">❌</h5>
<video autoplay controls muted loop playsinline width="80%">
<source src="static/videos/comparisons_with_baselines/franka_tabletop/octo--put_carrot.mp4" type="video/mp4">
</video>
</div>
<div class="column has-text-centered">
<h5><b>OpenVLA:</b><br>Put Carrot in Bowl</h5>
<h5 style="font-size:30px;">❌</h5>
<video autoplay controls muted loop playsinline width="80%">
<source src="static/videos/comparisons_with_baselines/franka_tabletop/openvla--put_carrot.mp4" type="video/mp4">
</video>
</div>
</div>
<div class="columns is-vcentered interpolation-panel">
<div class="column has-text-centered">
<h5><b>Diffusion Policy:</b><br>Pour Corn into Pot</h5>
<h5 style="font-size:30px;">✅</h5>
<video autoplay controls muted loop playsinline width="80%">
<source src="static/videos/comparisons_with_baselines/franka_tabletop/dp--pour_corn.mp4" type="video/mp4">
</video>
</div>
<div class="column has-text-centered">
<h5><b>Octo:</b><br>Pour Corn into Pot</h5>
<h5 style="font-size:30px;">❌</h5>
<video autoplay controls muted loop playsinline width="80%">
<source src="static/videos/comparisons_with_baselines/franka_tabletop/octo--pour_corn.mp4" type="video/mp4">
</video>
</div>
<div class="column has-text-centered">
<h5><b>OpenVLA:</b><br>Pour Corn into Pot</h5>
<h5 style="font-size:30px;">❌</h5>
<video autoplay controls muted loop playsinline width="80%">
<source src="static/videos/comparisons_with_baselines/franka_tabletop/openvla--pour_corn.mp4" type="video/mp4">
</video>
</div>
</div>
<br>
<div class="content has-text-justified">
<p>However, in fine-tuning tasks that involve multiple objects in the scene and require language conditioning, Octo and OpenVLA generally perform better, as their OpenX pretraining enables them to better adapt to these more diverse tasks.</p>
</div>
<div class="columns is-vcentered interpolation-panel">
<div class="column has-text-centered">
<h5><b>Diffusion Policy:</b><br>Move Yellow Corn onto Plate</h5>
<h5 style="font-size:30px;">❌</h5>
<video autoplay controls muted loop playsinline width="80%">
<source src="static/videos/comparisons_with_baselines/franka_tabletop/dp--move_corn_onto_plate.mp4" type="video/mp4">
</video>
</div>
<div class="column has-text-centered">
<h5><b>Octo:</b><br>Move Yellow Corn onto Plate</h5>
<h5 style="font-size:30px;">⚠️</h5>
<video autoplay controls muted loop playsinline width="80%">
<source src="static/videos/comparisons_with_baselines/franka_tabletop/octo--move_corn_onto_plate.mp4" type="video/mp4">
</video>
</div>
<div class="column has-text-centered">
<h5><b>OpenVLA:</b><br>Move Yellow Corn onto Plate</h5>
<h5 style="font-size:30px;">✅</h5>
<video autoplay controls muted loop playsinline width="80%">
<source src="static/videos/comparisons_with_baselines/franka_tabletop/openvla--move_corn_onto_plate.mp4" type="video/mp4">
</video>
</div>
</div>
<div class="columns is-vcentered interpolation-panel">
<div class="column has-text-centered">
<h5><b>Diffusion Policy:</b><br>Knock Brown Bear Over</h5>
<h5 style="font-size:30px;">❌</h5>
<video autoplay controls muted loop playsinline width="80%">
<source src="static/videos/comparisons_with_baselines/franka_tabletop/dp--knock_brown_bear_over.mp4" type="video/mp4">
</video>
</div>
<div class="column has-text-centered">
<h5><b>Octo:</b><br>Knock Brown Bear Over</h5>
<h5 style="font-size:30px;">⚠️</h5>
<video autoplay controls muted loop playsinline width="80%">
<source src="static/videos/comparisons_with_baselines/franka_tabletop/octo--knock_brown_bear_over.mp4" type="video/mp4">
</video>
</div>
<div class="column has-text-centered">
<h5><b>OpenVLA:</b><br>Knock Brown Bear Over</h5>
<h5 style="font-size:30px;">✅</h5>
<video autoplay controls muted loop playsinline width="80%">
<source src="static/videos/comparisons_with_baselines/franka_tabletop/openvla--knock_brown_bear_over.mp4" type="video/mp4">
</video>
</div>
</div>
<div class="columns is-vcentered interpolation-panel">
<div class="column has-text-centered">
<h5><b>Diffusion Policy:</b><br>Cover White Bowl with Towel</h5>
<h5 style="font-size:30px;">❌</h5>
<video autoplay controls muted loop playsinline width="80%">
<source src="static/videos/comparisons_with_baselines/franka_tabletop/dp--cover_white_bowl_with_towel.mp4" type="video/mp4">
</video>
</div>
<div class="column has-text-centered">
<h5><b>Octo:</b><br>Cover White Bowl with Towel</h5>
<h5 style="font-size:30px;">✅</h5>
<video autoplay controls muted loop playsinline width="80%">
<source src="static/videos/comparisons_with_baselines/franka_tabletop/octo--cover_white_bowl_with_towel.mp4" type="video/mp4">
</video>
</div>
<div class="column has-text-centered">
<h5><b>OpenVLA:</b><br>Cover White Bowl with Towel</h5>
<h5 style="font-size:30px;">⚠️</h5>
<video autoplay controls muted loop playsinline width="80%">
<source src="static/videos/comparisons_with_baselines/franka_tabletop/openvla--cover_white_bowl_with_towel.mp4" type="video/mp4">
</video>
</div>
</div>
<br>
<div class="content has-text-justified">
<p>We see evidence of the benefits of OpenX pretraining when comparing OpenVLA to OpenVLA (Scratch), which ablates OpenX pretraining and directly fine-tunes the base vision-language model on the Franka-Tabletop dataset. The full OpenVLA model exhibits much more reliable behaviors than OpenVLA (Scratch).</p>
</div>
<div class="columns is-vcentered interpolation-panel">
<div class="column has-text-centered">
<h5><b>OpenVLA (Scratch):</b><br>Flip Pot Upright</h5>
<h5 style="font-size:30px;">⚠️</h5>
<video autoplay controls muted loop playsinline width="80%">
<source src="static/videos/comparisons_with_baselines/franka_tabletop/base_vlm--flip_pot_upright.mp4" type="video/mp4">
</video>
</div>
<div class="column has-text-centered">
<h5><b>OpenVLA:</b><br>Flip Pot Upright</h5>
<h5 style="font-size:30px;">✅</h5>
<video autoplay controls muted loop playsinline width="80%">
<source src="static/videos/comparisons_with_baselines/franka_tabletop/openvla--flip_pot_upright.mp4" type="video/mp4">
</video>
</div>
</div>
<div class="columns is-vcentered interpolation-panel">
<div class="column has-text-centered">
<h5><b>OpenVLA (Scratch):</b><br>Move White Salt Shaker onto Plate</h5>
<h5 style="font-size:30px;">❌</h5>
<video autoplay controls muted loop playsinline width="80%">
<source src="static/videos/comparisons_with_baselines/franka_tabletop/base_vlm--move_salt_shaker_onto_plate.mp4" type="video/mp4">
</video>
</div>
<div class="column has-text-centered">
<h5><b>OpenVLA:</b><br>Move White Salt Shaker onto Plate</h5>
<h5 style="font-size:30px;">✅</h5>
<video autoplay controls muted loop playsinline width="80%">
<source src="static/videos/comparisons_with_baselines/franka_tabletop/openvla--move_salt_shaker_onto_plate.mp4" type="video/mp4">
</video>
</div>
</div>
</div>
</div>
</section>
<section class="section" id="BibTeX">
<div class="container is-max-desktop content">
<h2 class="title">BibTeX</h2>
<pre><code>@article{kim24openvla,
title={OpenVLA: An Open-Source Vision-Language-Action Model},
author={{Moo Jin} Kim and Karl Pertsch and Siddharth Karamcheti and Ted Xiao and Ashwin Balakrishna and Suraj Nair and Rafael Rafailov and Ethan Foster and Grace Lam and Pannag Sanketi and Quan Vuong and Thomas Kollar and Benjamin Burchfiel and Russ Tedrake and Dorsa Sadigh and Sergey Levine and Percy Liang and Chelsea Finn},
journal = {arXiv preprint arXiv:2406.09246},
year={2024},
} </code></pre>
</div>
</section>
<footer class="footer">
<div class="container">
<!-- <div class="content has-text-centered">
<a class="icon-link" href="https://arxiv.org/pdf/2210.05714.pdf">
<i class="fas fa-file-pdf"></i>
</a>
<a class="icon-link" href="" class="external-link" disabled>
<i class="fab fa-github"></i>
</a>
</div> -->
<div class="columns is-centered">
<div class="column is-8">
<div class="content">
<p> Website borrowed from <a href="https://github.com/nerfies/nerfies.github.io">NeRFies</a> under a <a
href="https://creativecommons.org/licenses/by-sa/4.0/">Creative Commons Attribution-ShareAlike 4.0
International</a>
</p>
</div>
</div>
</div>
</div>
</footer>
</body>
</html>