-
Notifications
You must be signed in to change notification settings - Fork 3
Expand file tree
/
Copy pathindex.html
More file actions
499 lines (439 loc) · 29.6 KB
/
index.html
File metadata and controls
499 lines (439 loc) · 29.6 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
<!DOCTYPE html>
<html>
<head>
<title>DigiRL</title>
<link rel="icon" href="website/images/icon/icon.png" type="image/icon type">
<meta name="viewport" content="width=device-width, initial-scale=1">
<script src="https://cdn.jsdelivr.net/npm/chart.js"></script>
<script src="https://cdn.jsdelivr.net/npm/chartjs-plugin-datalabels@2.0.0"></script>
<script
src="https://cdn.jsdelivr.net/npm/chartjs-plugin-annotation@3.0.1/dist/chartjs-plugin-annotation.min.js"></script>
<link href="https://fonts.googleapis.com/css?family=Google+Sans|Noto+Sans|Castoro" rel="stylesheet">
<link rel="stylesheet" href="website/css/bulma.min.css">
<link rel="stylesheet" href="website/css/bulma-carousel.min.css">
<link rel="stylesheet" href="website/css/bulma-slider.min.css">
<link rel="stylesheet" href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css">
<script src="https://ajax.googleapis.com/ajax/libs/jquery/3.5.1/jquery.min.js"></script>
<script src="./website/javascript/bulma-carousel.min.js"></script>
<script src="./website/javascript/bulma-slider.min.js"></script>
<link href="https://cdn.jsdelivr.net/npm/bootstrap@5.1.3/dist/css/bootstrap.min.css" rel="stylesheet"
integrity="sha384-1BmE4kWBq78iYhFldvKuhfTAU6auU8tT94WrHftjDbrCEXSU1oBoqyl2QvZ6jIW3" crossorigin="anonymous">
<script src="https://cdn.jsdelivr.net/npm/bootstrap@5.1.3/dist/js/bootstrap.bundle.min.js"
integrity="sha384-ka7Sk0Gln4gmtz2MlQnikT1wXgYsOg+OMhuP+IlRH9sENBO0LRn5q+8nbTov4+1p"
crossorigin="anonymous"></script>
<link href="https://unpkg.com/tabulator-tables@5.5.2/dist/css/tabulator_bootstrap4.min.css" rel="stylesheet">
<script type="text/javascript" src="https://unpkg.com/tabulator-tables@5.5.2/dist/js/tabulator.min.js"></script>
<!-- <script src="website/javascript/peity-vanilla.js"></script> -->
<script src="website/javascript/benchmark_table.js" type="module"></script>
<script src="website/javascript/success_rate_vs_k_vis.js" type="module"></script>
<script src="website/javascript/success_rate_vs_k_vis2.js" type="module"></script>
<script src="website/javascript/feedback_success_rate_vis.js" type="module"></script>
<script src="website/javascript/feedback_provider_efficacy.js" type="module"></script>
<script src="website/javascript/demos.js" type="module"></script>
<link rel="stylesheet" href="website/css/index.css">
<!-- Google tag (gtag.js) -->
<script async src="https://www.googletagmanager.com/gtag/js?id=G-C7GJ4FYMY9"></script>
<script>
window.dataLayer = window.dataLayer || [];
function gtag() { dataLayer.push(arguments); }
gtag('js', new Date());
gtag('config', 'G-C7GJ4FYMY9');
</script>
<noscript>
<p><img alt="Clicky" width="1" height="1" src="//in.getclicky.com/101339888ns.gif" /></p>
</noscript>
</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 publication-title">
<img src="website/images/icon/icon.png" alt="logo" width="40" height="40" />
DigiRL: Training In-The-Wild Device-Control Agents with Autonomous Reinforcement Learning
</h1>
<div class="is-size-5 publication-authors">
<span class="author-block">
<a href="https://www.jackgethome.com/">Hao Bai</a><sup>1,2*</sup>,
</span>
<span class="author-block">
<a href="https://yifeizhou02.github.io">Yifei Zhou</a><sup>1*</sup>,
</span>
<span class="author-block">
<a href="https://scholar.google.com/citations?user=sMEFwf8AAAAJ&hl=en">Mert Cemri</a><sup>1</sup>,
</span>
<span class="author-block">
<a href="https://www.jiayipan.me/">Jiayi Pan</a><sup>1</sup>,
</span>
<br>
<span class="author-block">
<a href="https://www.alanesuhr.com/">Alane Suhr</a><sup>1</sup>,
</span>
<span class="author-block">
<a href="https://people.eecs.berkeley.edu/~svlevine/">Sergey Levine</a><sup>1</sup>,
</span>
<span class="author-block">
<a href="https://aviralkumar2907.github.io/">Aviral Kumar</a><sup>3</sup>
</span>
</div>
<div class="is-size-5 publication-authors">
<span class="author-block"><sup>1</sup> UC Berkeley,</span>
<span class="author-block"><sup>2</sup> UIUC,</span>
<span class="author-block"><sup>3</sup> Google DeepMind</span>
</div>
<div class="is-size-5 publication-authors">
<span class="author-block"><small>*Equal contribution, alphabetic order; work done at UC Berkeley</small></span>
<div class="column has-text-centered">
<div class="publication-links">
<!-- PDF Link. -->
<span class="link-block">
<a href="https://arxiv.org/abs/2406.11896" class="btn btn-outline-dark"
role="button">📝
Paper</a>
</span>
<!-- Code Link. -->
<span class="link-block">
<a href="https://github.com/DigiRL-agent/digirl" class="btn btn-outline-dark"
role="button">💻
Code</a>
</span>
<!-- Dataset Link. -->
<span class="link-block">
<a href="https://drive.google.com/drive/folders/14Iu6lAHePQ2qG0ghYkVG1RG6RUu7e2Hz?usp=sharing"
class="btn btn-outline-dark" role="button">📂
Data</a>
</div>
</div>
<!-- <h2 class="subtitle" style="text-align: left;">
<b>MINT benchmark</b> measures LLMs' ability to solve tasks with multi-turn interactions
by
(1) using tools and (2) leveraging natural language feedback.
</h2> -->
</div>
</div>
</div>
</div>
</section>
<section class="section" id="propaganda">
<div class="container is-max-desktop">
<div class="is-centered has-text-centered">
<div class="custom-column-large">
<div class="is-centered custom-column-large">
<h2 class="title is-3">Demo</h2>
<div class="text-center">
<div class="btn-group btn-group-toggle text-center task-selector" data-toggle="buttons">
<button type="button" class="btn btn-outline-secondary btn-sm inline-vis-button" disabled>Speed:</button>
<button type="button" class="btn btn-outline-secondary btn-sm active" id="fast-button">Fast</button>
<button type="button" class="btn btn-outline-secondary btn-sm" id="slow-button">Slow</button>
</div>
</div>
</div>
<br>
<!-- All video -->
<div id="fast-container">
<video controls autoplay muted loop>
<source src="website/videos/demo2_5x.mp4" type="video/mp4">
</video>
</div>
<div id="slow-container" style="display:none;">
<video controls autoplay muted loop>
<source src="website/videos/demo2.mp4" type="video/mp4">
</video>
</div>
<small>Success rate and corresponding trajectories. A <font color="#007800">green</font> final screen indicates a successful trajectory; a <font color="#ff1900">red</font> final screen indicates a failed trajectory.</s>
</div>
</div>
</div>
</section>
<section class="section" id="abstract">
<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>
Training corpuses for vision language models typically lack sufficient amounts of decision-centric data. This renders off-the-shelf VLMs sub-optimal for decision-making tasks such as in-the-wild device control through graphical user interfaces (GUIs). While training with static demonstrations has shown some promise, we show that such methods fall short when controlling <b>real</b> GUIs due to their failure to deal with real world <b>stochasticity</b> not captured in static observational data.
This paper introduces a novel autonomous RL approach, called <b>DigiRL</b>, for training in-the-wild device control agents through fine-tuning a pre-trained VLM in two stages: offline RL to initialize the model, followed by offline-to-online RL. To do this, we build a <b>scalable</b> and <b>parallelizable</b> Android learning environment equipped with a VLM-based evaluator and develop a simple yet effective RL approach for learning in this domain. Our approach runs advantage-weighted RL with advantage estimators enhanced to account for stochasticity along with an automatic curriculum for deriving maximal learning signal.
We demonstrate the effectiveness of <b>DigiRL</b> using the Android-in-the-Wild (AitW) dataset, where our 1.5B VLM trained with RL achieves a <b>49.5% absolute improvement</b> -- from 17.7% to 67.2% success rate -- over supervised fine-tuning with static human demonstration data. These results significantly surpass not only the prior best agents, including AppAgent with GPT-4V (8.3% success rate) and the 17B CogAgent trained with AitW data (14.4%), but also the prior best autonomous RL approach based on filtered behavior cloning (57.8%), thereby <b>establishing a new state-of-the-art</b> for digital agents for in-the-wild device control.
</p>
</div>
</div>
</div>
<!--/ Abstract. -->
</div>
<div class="container is-max-desktop">
<div class="is-centered has-text-centered">
<div class="custom-column-large">
<div class="is-centered custom-column-large">
<h4><b>AitW General</b><br>
Search for some good Italian restaurants</h4>
<div class="text-center">
<div class="btn-group btn-group-toggle text-center task-selector" data-toggle="buttons">
<button type="button" class="btn btn-outline-secondary btn-sm inline-vis-button" disabled>Visualization:</button>
<button type="button" class="btn btn-outline-secondary btn-sm active" id="all-methods">All Methods</button>
<button type="button" class="btn btn-outline-secondary btn-sm" id="each-method">Each Method (pause at wish)</button>
</div>
</div>
</div>
<br>
<!-- All video -->
<div id="all-video-container">
<video controls autoplay muted loop>
<source src="website/videos/general.mp4" type="video/mp4">
</video>
</div>
<!-- Each video -->
<div id="each-video-container" style="display:none;">
<video controls autoplay muted loop>
<source src="website/videos/general/slice1.mp4" type="video/mp4">
</video>
<video controls autoplay muted loop>
<source src="website/videos/general/slice2.mp4" type="video/mp4">
</video>
<video controls autoplay muted loop>
<source src="website/videos/general/slice3.mp4" type="video/mp4">
</video>
<video controls autoplay muted loop>
<source src="website/videos/general/slice4.mp4" type="video/mp4">
</video>
</div>
</div>
</div>
</div>
<br>
<div class="container is-max-desktop">
<div class="is-centered has-text-centered">
<div class="custom-column-large">
<div class="custom-column-large">
<h4><b>AitW Web Shopping</b><br>
Go to newegg.com, and search for "Alienware Aurora"</h4>
<div class="text-center">
<div class="btn-group btn-group-toggle text-center task-selector" data-toggle="buttons">
<button type="button" class="btn btn-outline-secondary btn-sm inline-vis-button" disabled>Visualization:</button>
<button type="button" class="btn btn-outline-secondary btn-sm active" id="webshop-all-methods">All Methods</button>
<button type="button" class="btn btn-outline-secondary btn-sm" id="webshop-each-method">Each Method (pause at wish)</button>
</div>
</div>
</div>
<br>
<!-- All video -->
<div id="all-webshop-container">
<video controls autoplay muted loop>
<source src="website/videos/webshop.mp4" type="video/mp4">
</video>
</div>
<!-- Each video -->
<div id="each-webshop-container" style="display:none;">
<video controls autoplay muted loop>
<source src="website/videos/webshop/slice1.mp4" type="video/mp4">
</video>
<video controls autoplay muted loop>
<source src="website/videos/webshop/slice2.mp4" type="video/mp4">
</video>
<video controls autoplay muted loop>
<source src="website/videos/webshop/slice3.mp4" type="video/mp4">
</video>
<video controls autoplay muted loop>
<source src="website/videos/webshop/slice4.mp4" type="video/mp4">
</video>
</div>
</div>
</div>
</div>
</section>
<section class="hero teaser">
<div class="container is-max-desktop">
<div class="hero-body">
<h2 class="subtitle">
<b>DigiRL</b> solves open-ended realistic Android tasks with an novel online reinforcement learning algorithm with autonomous VLM evaluator.
</h2>
<div class="tab-content" id="myTabContent">
<div class="tab-pane fade show active" id="benchmark-table-content" role="tabpanel"
aria-labelledby="benchmark-table-content">
<div id="benchmark-table"></div>
</div>
<div class="tab-pane fade" id="eurus-code-table-content" role="tabpanel"
aria-labelledby="eurus-code-table-content">
<p class="mt-2 px-2">
This code subset follows the <a href="https://arxiv.org/abs/2404.02078">Eurus
paper</a> and contains MBPP and HumanEval.
</p>
</div>
<p class="mt-2 px-2">
This table contains the success rate across all approaches measured in the DigiRL paper. It includes performance on
two subsets: AitW General and AitW Web Shopping. The codename of GPT-4V we use is <i>gpt-4-vision-preview</i>
and the codename of Gemini-1.5-Pro is <i>gemini-1.5-pro-latest</i>.
</p>
</div>
</div>
</div>
</section>
<section class="section" id="interaction-framework">
<div class="container is-max-desktop">
<div class="columns is-full-width">
<!-- Visual Effects. -->
<div class="column">
<div class="text-justified">
<h2 class="title is-3">DigiRL: Autonomous RL for Building a Strong Device-Control Agent</h2>
<h3> Why RL over the alternatives? </h3>
<ul class="custom-bullets">
<li>LLM Agent data such as device-control actions is poorly represented in the pre-training corpus of <b>Off-the-shelf proprietary VLMs</b> such as GPT4V and Gemini-1.5-Pro.</li>
<li><b>Supervised Fine-Tuning</b> 1) requires a large amount of human demonstration data and 2) cannot recover from degrading model performance when real websites/applications have changed. As shown in the plot below, a frozen good policy trained with prior data experiences a gradual drop in performance as the websites change over time, while the DigiRL policy constantly updates with fresh autonomous data can maintain a stable performance.</li>
</ul>
</div>
<div class="text-center">
<!-- button for visualizing offline or not -->
<div class="btn-group btn-group-toggle text-center task-selector" data-toggle="buttons">
<button type="button" class="btn btn-outline-secondary btn-sm inline-vis-button" disabled
id="visualize-sr-vs-k-scale-with-model-size-llama2-base-disabled">Visualization:</button>
<button type="button" class="btn btn-outline-secondary btn-sm active"
id="exc-offline">Exclude Offline Results (click to visualize)</button>
<button type="button" class="btn btn-outline-secondary btn-sm"
id="inc-offline">Include Offline Results</button>
</div>
</div>
<div class="text-justified">
<div class="chart-container" id="chart-k2" style="display:block;margin:0 auto;">
<canvas id="chart-sr-vs-k2"></canvas>
</div>
<h3> What are we using RL for?</h3>
DigiRL consists of two steps:
<ul class="custom-bullets">
<li>First, we use <b>Offline RL</b> to make the most out of a potentially sub-optimal existing offline dataset.</li>
<li>Then, we use <b>Offline-to-Online RL</b> to encourage the agent to learn from its own trials and errors.</li>
</ul>
DigiRL identifies the most simple yet effective RL design choices for device-control agent problems. Our RL algorithmic framework automatically achieves the following advantages compared to state-of-the-art alternatives such as rejection sampling (or Filtered Behavior Cloning):
<ul class="custom-bullets">
<li>We makes use of an <b>instruction-level value function</b> to implicitly construct an automatic curriculum that prioritizes on the tasks most informative to the agent.</li>
<li>We makes use of a <b>step-level value function</b> to pick out the advantageous actions (actions that mark progress towards the goal) in a trajectory while leaving the noisy actions (actions that do not contribute to the goal).</li>
</ul>
Please check out our paper for more details of our algorithm!
<div style="text-align:center;">
<img src="website/images/algo.png" alt="illustrative-example"
style="margin: 0 auto; display: block; max-width: 1000px; width: 100%; height: auto;" />
<br>
</div>
<h3>Learning Curves</h3>
<div class="text-justified" id="tool-augmented">
In addition to the convergence performance reported in the paper, we also present the sample complexity comparison of DigiRL against the state-of-the-art alternative Filtered Behavior Cloning (or rejection sampling). We find that DigiRL not only converges to a superior performance, but also learns more efficiently.
</div>
<div class="text-center">
<br>
<div class="btn-group btn-group-toggle text-center task-selector" data-toggle="buttons">
<button type="button" class="btn btn-outline-secondary btn-sm inline-vis-button" disabled
id="visualize-sr-vs-k-scale-with-model-size-llama2-base-disabled">Subset:</button>
<button type="button" class="btn btn-outline-secondary btn-sm active"
id="visualize-sr-vs-k-scale-with-model-size-llama2-base">AitW General (click to visualize)</button>
<button type="button" class="btn btn-outline-secondary btn-sm"
id="visualize-sr-vs-k-scale-with-model-size-llama2-rlhf">AitW Web Shopping</button>
</div>
</div>
<div class="chart-container" id="chart-k" style="display:block;margin:0 auto;">
<canvas id="chart-sr-vs-k"></canvas>
</div>
</div>
</div>
<!--/ Visual Effects. -->
</div>
</section>
<section class="section" id="evaluation">
<div class="container is-max-desktop">
<div class="columns is-full-width">
<!-- Visual Effects. -->
<div class="column">
<div class="content">
<h3 class="title is-3">Autonomous Evaluation</h3>
<p>
Our main results are autonomously evaluated with Gemini-1.5-Pro. We also manually evaluate on some subsets and finds that the autonomous evaluation results highly align with manual evaluations with an average difference less than 3%:
</p>
<div class="text-center">
<div class="btn-group btn-group-toggle text-center feedback-provider-sort-by-selector"
data-toggle="buttons">
<button type="button" class="btn btn-outline-secondary btn-sm" disabled>Subset:</button>
<button type="button" class="btn btn-outline-secondary btn-sm active"
id="sort-by-feedback-gain">
AitW General
</button>
<button type="button" class="btn btn-outline-secondary btn-sm"
id="sort-by-feedback-provider-perf">
AitW Web Shopping
</button>
</div>
</div>
<div class="chart-container" id="chart-feedback-p" style="display:block;margin:0 auto;">
<canvas id="chart-feedback-provider"></canvas>
</div>
<h3>Failure Mode Analysis</h3>
<p>
While all the types of failure modes benefit from offline and offline-to-online RL training, the most consistent and significant reduction is for the failure mode of failing to recover from mistakes. By training on autonomously-collected rollouts, our agent DigiRL is able to learn from its own mistakes and reduces failures to recover over training.
</p>
<div class="text-center">
<div class="btn-group btn-group-toggle text-center task-selector" data-toggle="buttons">
<button type="button" class="btn btn-outline-secondary btn-sm" disabled>Subset:</button>
<button type="button" class="btn btn-outline-secondary btn-sm active"
id="avg_micro">AitW General</button>
<button type="button" class="btn btn-outline-secondary btn-sm"
id="reasoning">AitW Web Shopping</button>
</div>
<!-- <div class="btn-group btn-group-toggle text-center sort-by-selector" data-toggle="buttons">
<button type="button" class="btn btn-outline-secondary btn-sm" disabled>Failure:</button>
<button type="button" class="btn btn-outline-secondary btn-sm active"
id="sort-by-feedbacksr">Fail to recover from mistakes</button>
<button type="button" class="btn btn-outline-secondary btn-sm"
id="sort-by-nofeedbacksr">Get stuck midway</button>
<button type="button" class="btn btn-outline-secondary btn-sm"
id="sort-by-feedbackdelta">Arrive at wrong goal</button>
</div> -->
</div>
<div class="chart-container" id="chart-feedback" style="position:relative;margin:0 auto;">
<canvas id="chart-sr-w-feedback" style="max-height: 100%;"></canvas>
</div>
</div>
</div>
</div>
</section>
<section class="section" id="evaluation">
<div class="container is-max-desktop">
<div class="columns is-full-width">
<div class="column">
<div class="content">
<h3>Misc</h3>
<h4>Re the Charts 📈 </h4>
Try clicking on the legend of the charts!
</div>
<div class="content">
<h4>Re the Icon <img src="website/images/icon/icon.png" alt="logo" width="30" height="30" /></h4>
<p>
<b>Infinity:</b> Our environment is open-ended, which can be easily generalized to infinite open-ended tasks sets with our open-ended evaluator.
<br>
<b>Loop:</b> We use online reinforcement learning, which is closed-loop: the agent interacts with the environment and learns from its own trials and errors.
</p>
</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{bai2024digirl,
title={DigiRL: Training In-The-Wild Device-Control Agents with Autonomous Reinforcement Learning},
author={Bai, Hao and Zhou, Yifei and Cemri, Mert and Pan, Jiayi and Suhr, Alane and Levine, Sergey and Kumar, Aviral},
journal={arXiv preprint arXiv:2406.11896},
year={2024}
}</code></pre>
</div>
</section>
<footer class="footer">
<div align="center" class="container">
<div class="columns is-centered">
<div class="content is-small">
This website templated is borrowed from <a href="https://github.com/nerfies/nerfies.github.io">here</a> and <a href="https://xwang.dev/mint-bench/">here</a>.
</div>
</div>
</div>
</footer>
</body>
</html>