-
Notifications
You must be signed in to change notification settings - Fork 0
/
go-fresh.html
527 lines (468 loc) · 18.8 KB
/
go-fresh.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
<!-- <!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN"> -->
<html>
<!-- ======================================================================= -->
<script src="http://www.google.com/jsapi" type="text/javascript"></script>
<script type="text/javascript">google.load("jquery", "1.3.2");</script>
<style type="text/css">
body {
font-family: "Titillium Web", "HelveticaNeue-Light", "Helvetica Neue Light", "Helvetica Neue", Helvetica, Arial, "Lucida Grande", sans-serif;
font-weight: 300;
font-size: 18px;
margin-left: auto;
margin-right: auto;
width: 100%;
}
h1 {
font-weight: 300;
}
div {
max-width: 95%;
margin: auto;
padding: 10px;
}
.table-like {
display: flex;
flex-wrap: wrap;
flex-flow: row wrap;
justify-content: center;
}
.disclaimerbox {
background-color: #eee;
border: 1px solid #eeeeee;
border-radius: 10px;
-moz-border-radius: 10px;
-webkit-border-radius: 10px;
padding: 20px;
}
video.header-vid {
height: 140px;
border: 1px solid black;
border-radius: 10px;
-moz-border-radius: 10px;
-webkit-border-radius: 10px;
}
img {
padding: 0;
display: block;
margin: 0 auto;
max-height: 100%;
max-width: 100%;
}
iframe {
max-width: 100%;
}
img.header-img {
height: 140px;
border: 1px solid black;
border-radius: 10px;
-moz-border-radius: 10px;
-webkit-border-radius: 10px;
}
img.rounded {
border: 1px solid #eeeeee;
border-radius: 10px;
-moz-border-radius: 10px;
-webkit-border-radius: 10px;
}
a:link,
a:visited {
color: #1367a7;
text-decoration: none;
}
a:hover {
color: #208799;
}
td.dl-link {
height: 160px;
text-align: center;
font-size: 22px;
}
.layered-paper-big {
/* modified from: http://css-tricks.com/snippets/css/layered-paper/ */
box-shadow:
0px 0px 1px 1px rgba(0, 0, 0, 0.35),
/* The top layer shadow */
5px 5px 0 0px #fff,
/* The second layer */
5px 5px 1px 1px rgba(0, 0, 0, 0.35),
/* The second layer shadow */
10px 10px 0 0px #fff,
/* The third layer */
10px 10px 1px 1px rgba(0, 0, 0, 0.35),
/* The third layer shadow */
15px 15px 0 0px #fff,
/* The fourth layer */
15px 15px 1px 1px rgba(0, 0, 0, 0.35),
/* The fourth layer shadow */
20px 20px 0 0px #fff,
/* The fifth layer */
20px 20px 1px 1px rgba(0, 0, 0, 0.35),
/* The fifth layer shadow */
25px 25px 0 0px #fff,
/* The fifth layer */
25px 25px 1px 1px rgba(0, 0, 0, 0.35);
/* The fifth layer shadow */
margin-left: 10px;
margin-right: 45px;
}
.layered-paper {
/* modified from: http://css-tricks.com/snippets/css/layered-paper/ */
box-shadow:
0px 0px 1px 1px rgba(0, 0, 0, 0.35),
/* The top layer shadow */
5px 5px 0 0px #fff,
/* The second layer */
5px 5px 1px 1px rgba(0, 0, 0, 0.35),
/* The second layer shadow */
10px 10px 0 0px #fff,
/* The third layer */
10px 10px 1px 1px rgba(0, 0, 0, 0.35);
/* The third layer shadow */
margin-top: 5px;
margin-left: 10px;
margin-right: 30px;
margin-bottom: 5px;
}
.vert-cent {
position: relative;
top: 50%;
transform: translateY(-50%);
}
hr {
border: 0;
height: 1px;
max-width: 1100px;
background-image: linear-gradient(to right, rgba(0, 0, 0, 0), rgba(0, 0, 0, 0.75), rgba(0, 0, 0, 0));
}
#authors td {
padding-bottom: 5px;
padding-top: 30px;
}
.authors {
font-size: 22px;
}
</style>
<!-- ======================================================================= -->
<script type="text/javascript" src="resources/hidebib.js"></script>
<link href='https://fonts.googleapis.com/css?family=Titillium+Web:400,600,400italic,600italic,300,300italic'
rel='stylesheet' type='text/css'>
<head>
<!-- Global site tag (gtag.js) - Google Analytics -->
<script async src="https://www.googletagmanager.com/gtag/js?id=G-9VC104QM02"></script>
<script>
window.dataLayer = window.dataLayer || [];
function gtag() { dataLayer.push(arguments); }
gtag('js', new Date());
gtag('config', 'G-9VC104QM02');
</script>
<div max-width=100%>
<meta charset="utf-8" />
<meta http-equiv="X-UA-Compatible" content="IE=edge" />
<link rel="icon" type="image/png" href="resources/seal_icon.png">
<title>Learning Goal-Conditioned Policies Offline with Self-Supervised Reward Shaping</title>
<meta name="HandheldFriendly" content="True" />
<meta name="viewport" content="width=device-width, initial-scale=1.0" />
<link rel="canonical" href="https://linamezghani.github.io/" />
<meta name="referrer" content="no-referrer-when-downgrade" />
</head>
<body>
<br>
<center><span style="font-size:44px;font-weight:bold;">Learning Goal-Conditioned Policies Offline with </br>
Self-Supervised Reward Shaping</span></center>
<div class="table-like authors" style="justify-content:space-evenly;max-width:650px;margin:auto;">
<div>
<center><a href="https://linamezghani.github.io" target="_blank">Lina Mezghani</a></center>
<center>Inria, Meta AI</center>
</div>
<div>
<center></center><a href="https://scholar.google.com/citations?user=ri1sE34AAAAJ&hl=en" target="_blank">Sainbayar
Sukhbaatar</a></center>
<center>Meta AI</center>
</div>
<div>
<center><a href="https://scholar.google.fr/citations?user=lJ_oh2EAAAAJ&hl=en" target="_blank">Piotr Bojanowski</a></center>
<center>Meta AI</center>
</div>
<div>
<center><a href="https://scholar.google.com/citations?user=6JZ3R6wAAAAJ&hl=en" target="_blank">Alessandro Lazaric</a></center>
<center>Meta AI</center>
</div>
<div>
<center><a href="https://lear.inrialpes.fr/people/alahari/" target="_blank">Karteek Alahari</a>
</center>
<center>Inria</center>
</div>
</div>
<center><span style="font-size:20px;"><a href='https:/corl2022.org/'> 6th Conference on Robot Learning (CoRL 2022)</a></span></center>
<div class="table-like" style="justify-content:space-evenly;max-width:700px;margin:auto;padding:5px">
<div>
<center><span style="font-size:28px"><a href="https://arxiv.org/abs/2301.02099">[Paper]</a></span></center>
</div>
<div>
<center><span style="font-size:28px"><a href='https://github.com/facebookresearch/go-fresh'>[Code]</a></span></center>
</div>
</div>
<br>
<center>
<h1>Abstract</h1>
</center>
<div style="width:800px; margin:0 auto; text-align:justify;">
Developing agents that can execute multiple skills by learning from pre-collected datasets is an important problem in robotics, where online interaction with the environment is extremely time-consuming.
Moreover, manually designing reward functions for every single desired skill is prohibitive.
Prior works targeted these challenges by learning goal-conditioned policies from offline datasets without manually specified rewards, through hindsight relabeling.
These methods suffer from the issue of sparsity of rewards, and fail at long-horizon tasks.
In this work, we propose a novel self-supervised learning phase on the pre-collected dataset to understand the structure and the dynamics of the model, and shape a dense reward function for learning policies offline.
We evaluate our method on three continuous control tasks, and show that our model significantly outperforms existing approaches, especially on tasks that involve long-term planning.
</div>
<br>
<hr>
<!-- <center>
<h1>Latent Trajectory Optimization</h1>
<a href="resources/teaser.mp4"><video src="resources/teaser.mp4" width="800px" autoplay muted loop></video></a>
</center>
<div style="width:800px; margin:0 auto; text-align:right">
Latent collocation (LatCo) on a tool use task, where the thermos needs to be pushed with the stick. Each image shows
a full plan at that optimization step, visualized via a diagnostic network. LatCo optimizes a latent state sequence
and is able to temporarily violate the dynamics during planning, such as the stick flying in the air without an
apparent cause. This allows it to rapidly discover the high-reward regions, while the subsequent refinement of the
planned trajectory focuses on feasibly achieving it. In contrast, shooting optimizes an action sequence directly and
is unable to discover picking the stick as the actions that lead to that are unlikely to be sampled.
</div>
<center id="sourceCode">
<h2>More LatCo Examples</h2>
</center>
<div class="table-like" style="justify-content:space-evenly;max-width:1100px;margin:auto;padding:0px">
<div style="max-width:250px">
<center>
<h3> Thermos </h3>
</center>
<a href="resources/mod2_opt_mug.mp4"><video src="resources/mod2_opt_mug.mp4" width="100%" autoplay muted
loop></video></a>
</div>
<div style="max-width:250px">
<center>
<h3> Hammer </h3>
</center>
<a href="resources/mod2_hammer_opt.mp4"><video src="resources/mod2_hammer_opt.mp4" width="100%" autoplay muted
loop></video></a>
</div>
<div style="max-width:250px">
<center>
<h3> Push </h3>
</center>
<a href="resources/opt_push.mp4"><video src="resources/opt_push.mp4" width="100%" autoplay muted loop></video></a>
</div>
<div style="max-width:250px">
<center>
<h3> Push </h3>
</center>
<a href="resources/opt_push2.mp4"><video src="resources/opt_push2.mp4" width="100%" autoplay muted
loop></video></a>
</div>
</div>
<hr>
<div style="width:800px; margin:0 auto; text-align:center">
<center>
<h1>Overview of the Algorithm</h1>
</center>
</div>
<div style="width:800px; margin:0 auto; text-align:right">
We propose a visual model-based reinforcement method that optimizes sequences of latent states using a constrained
optimization objective. First, a latent dynamics model is learned from images using an approximate maximum
likelihood objective (left). While shooting (bottom) optimizes actions directly, LatCo (top) instead optimizes a
sequence of latent states to maximize the reward. The latent states are constrained to be dynamically feasible by
recovering the corresponding sequence of actions and enforcing the dynamics. Dual descent on the Lagrangian is used
to solve the constrained optimization problem.
</div>
<br />
<center><a href="resources/method.png"><img src="resources/method.png" width="1100px"></img></a><br></center>
<br />
<hr>
<center id="sourceCode">
<h1>LatCo Results</h1>
</center>
<div class="table-like" style="justify-content:space-evenly;max-width:1100px;margin:auto;padding:0px">
<div style="width:70px"> Execution </div>
<div style="max-width:970px">
<a href="resources/results/modcombined_tasks_latco_exec_0.mp4"><video
src="resources/results/modcombined_tasks_latco_exec_0.mp4" width="100%" autoplay muted loop></video></a>
</div>
<div style="width:70px"> Plan </div>
<div style="max-width:970px">
<a href="resources/results/modcombined_tasks_latco_plan_0.mp4"><video
src="resources/results/modcombined_tasks_latco_plan_0.mp4" width="100%" autoplay muted loop></video></a>
</div>
</div>
</br><br>
<div class="table-like" style="justify-content:space-evenly;max-width:1100px;margin:auto;padding:0px">
<div style="width:70px"> Execution </div>
<div style="max-width:970px">
<a href="resources/results/modcombined_tasks_latco_exec_1.mp4"><video
src="resources/results/modcombined_tasks_latco_exec_1.mp4" width="100%" autoplay muted loop></video></a>
</div>
<div style="width:70px"> Plan </div>
<div style="max-width:970px">
<a href="resources/results/modcombined_tasks_latco_plan_1.mp4"><video
src="resources/results/modcombined_tasks_latco_plan_1.mp4" width="100%" autoplay muted loop></video></a>
</div>
</div>
<div style="width:800px; margin:0 auto; text-align:right">
Our method performs these seven challenging sparse reward and long-horizon tasks directly from image input. While
prior work on model-based reinforcement learning struggles with long-horizon tasks, latent collocation (LatCo) plans
sequences of latent states using a constrained optimization objective, which enables is to escape local minima and
make effective visual plans even for complex multiple-stage tasks like Thermos and Hammer.
</div></br>
<hr>
<center id="sourceCode">
<h1>Comparisons - Hammer</h1>
</center>
<div class="table-like" style="justify-content:space-evenly;max-width:1100px;margin:auto;padding:0px">
<div>
<center>
<h3> LatCo </h3>
</center>
<div class="table-like" style="justify-content:space-evenly;max-width:300px;margin:auto;padding:0px">
<div style="width:150px;padding:0px">Execution</div>
<div style="width:150px;padding:0px">Plan</div>
</div>
<a href="resources/results/modcombined_baselines_hammer_.mp4"><video
src="resources/results/modcombined_baselines_hammer_.mp4" width="300px" autoplay muted loop></video></a>
</div>
<div>
<center>
<h3> Shooting CEM (PlaNet, Hafner'19) </h3>
</center>
<div class="table-like" style="justify-content:space-evenly;max-width:300px;margin:auto;padding:0px">
<div style="width:150px;padding:0px">Execution</div>
<div style="width:150px;padding:0px">Plan</div>
</div>
<a href="resources/results/modcombined_baselines_hammer_cem.mp4"><video
src="resources/results/modcombined_baselines_hammer_cem.mp4" width="300px" autoplay muted loop></video></a>
</div>
<div>
<center>
<h3> Shooting GD </h3>
</center>
<div class="table-like" style="justify-content:space-evenly;max-width:300px;margin:auto;padding:0px">
<div style="width:150px;padding:0px">Execution</div>
<div style="width:150px;padding:0px">Plan</div>
</div>
<a href="resources/results/modcombined_baselines_hammer_gd.mp4"><video
src="resources/results/modcombined_baselines_hammer_gd.mp4" width="300px" autoplay muted loop></video></a>
</div>
</div>
<center id="sourceCode">
<h1>Comparisons - Thermos</h1>
</center>
<div class="table-like" style="justify-content:space-evenly;max-width:1100px;margin:auto;padding:0px">
<div>
<center>
<h3> LatCo </h3>
</center>
<div class="table-like" style="justify-content:space-evenly;max-width:300px;margin:auto;padding:0px">
<div style="width:150px;padding:0px">Execution</div>
<div style="width:150px;padding:0px">Plan</div>
</div>
<a href="resources/results/modcombined_baselines_thermos_.mp4"><video
src="resources/results/modcombined_baselines_thermos_.mp4" width="300px" autoplay muted loop></video></a>
</div>
<div>
<center>
<h3> Shooting CEM (PlaNet, Hafner'19) </h3>
</center>
<div class="table-like" style="justify-content:space-evenly;max-width:300px;margin:auto;padding:0px">
<div style="width:150px;padding:0px">Execution</div>
<div style="width:150px;padding:0px">Plan</div>
</div>
<a href="resources/results/modcombined_baselines_thermos_cem.mp4"><video
src="resources/results/modcombined_baselines_thermos_cem.mp4" width="300px" autoplay muted loop></video></a>
</div>
<div>
<center>
<h3> Shooting GD </h3>
</center>
<div class="table-like" style="justify-content:space-evenly;max-width:300px;margin:auto;padding:0px">
<div style="width:150px;padding:0px">Execution</div>
<div style="width:150px;padding:0px">Plan</div>
</div>
<a href="resources/results/modcombined_baselines_thermos_gd.mp4"><video
src="resources/results/modcombined_baselines_thermos_gd.mp4" width="300px" autoplay muted loop></video></a>
</div>
</div>
<div style="width:800px; margin:0 auto; text-align:right">
The Thermos and Hammer tasks require picking up a tool and using it to manipulate another object. The sparse reward
is only given after completing the full task, and the planner needs to infer all stages required to solve the task.
While LatCo can solve both of these challenging tasks, prior shooting methods fail to find a good trajectory.
</div></br>
<hr>
<center id="sourceCode">
<h1>Source Code</h1>
</center>
<div style="width:800px; margin:0 auto; text-align:right">
Try our implementation of LatCo in TensorFlow 2 as well as the Sparse MetaWorld environment below!
</div>
<div class="table-like" style="justify-content:space-evenly;max-width:700px;margin:auto;padding:5px">
<div>
<center><span style="font-size:28px"><a href='https://github.com/zchuning/latco'>[Method Code]</a></span></center>
</div>
<div>
<center><span style="font-size:28px"><a href='https://github.com/zchuning/metaworld'>[Env Code]</a></span>
</center>
</div>
</div>
<br>
<hr>
<table align=center width=850px>
<center>
<h1>Paper and Bibtex</h1>
</center>
<tr>
<td width=350px align=left>
<a href="https://arxiv.org/pdf/2106.13229.pdf"><img style="height:150px" src="resources/thumbnail.jpg" /></a>
<center>
<span style="font-size:20pt"><a href="https://arxiv.org/pdf/2106.13229.pdf">[Paper]</a>
<span style="font-size:20pt"><a href="https://arxiv.org/abs/2106.13229">[ArXiv]</a>
</center>
</td>
<td width=50px align=center>
</td>
<td width=550px align=left>
<p style="text-align:left;"><b><span style="font-size:20pt">Citation</span></b><br /><span
style="font-size:6px;"> <br /></span> <span style="font-size:15pt">Oleh Rybkin, Chuning Zhu, Anusha
Nagabandi, Kostas Daniilidis, Igor Mordatch, Sergey Levine. <b>Model-Based Reinforcement Learning via
Latent-Space Collocation. ICML, 2020.<br /></b></span></p>
<span style="font-size:20pt"><a shape="rect" href="javascript:togglebib('gcp2020_bib')"
class="togglebib">[Bibtex]</a></span>
</td>
</tr>
<tr>
<td width=250px align=left>
</td>
<td width=50px align=center>
</td>
<td width=550px align=left>
<div class="paper" id="gcp2020_bib">
<pre xml:space="preserve">
@misc{rybkin2020model,
title={Model-Based Reinforcement Learning
via Latent-Space Collocation},
author={Rybkin, Oleh and Zhu, Chuning and
Nagabandi, Anusha and Daniilidis, Kostas
and Mordatch, Igor and Levine, Sergey},
booktitle={International Conference on
Machine Learning (ICML)},
year={2021}
}</pre>
</div>
</td>
</tr>
</table>
<br>
<hr>
</table> -->
<script xml:space="preserve" language="JavaScript">
hideallbibs();
</script>
</div>
</body>
</html>