-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathindex.html
463 lines (430 loc) · 20.5 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
<!DOCTYPE html>
<html lang="en">
<head>
<title>6-DOF Grasp Pose Evaluation and Optimization via Transfer Learning from NeRFs</title>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<link rel="stylesheet" href="https://www.w3schools.com/w3css/4/w3.css">
<link rel="stylesheet" href="https://fonts.googleapis.com/css?family=Lato">
<link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/font-awesome/4.7.0/css/font-awesome.min.css">
<style>
body {
font-family: "Lato", sans-serif
}
/* The Modal (background) */
.modal {
display: none; /* Hidden by default */
position: fixed; /* Stay in place */
z-index: 1; /* Sit on top */
padding-top: 60px; /* Location of the box */
left: 0;
top: 0;
width: 100%; /* Full width */
height: 100%; /* Full height */
overflow: auto; /* Enable scroll if needed */
background-color: rgb(0, 0, 0); /* Fallback color */
background-color: rgba(0, 0, 0, 0.9); /* Black w/ opacity */
}
/* Modal Content (Image) */
.modal-content {
margin: auto;
display: block;
}
/* Caption of Modal Image */
#modal_caption {
margin: auto;
display: block;
width: 80%;
max-width: 700px;
text-align: center;
color: #ccc;
padding: 10px 0;
font-weight: bold;
}
/* Add Animation */
.modal-content, #modal_caption {
animation-name: zoom;
animation-duration: 0.6s;
}
@keyframes zoom {
from {
transform: scale(0)
}
to {
transform: scale(1)
}
}
/* The Close Button */
.close {
position: absolute;
top: 15px;
right: 35px;
color: #f1f1f1;
font-size: 40px;
font-weight: bold;
transition: 0.3s;
}
.close:hover,
.close:focus {
color: #bbb;
text-decoration: none;
cursor: pointer;
}
/* Responsive Columns */
.column {
width: 100%;
max-width: 400px;
}
/* Clear floats after the columns */
.row:after {
content: "";
display: table;
clear: both;
}
/* Responsive Layout */
@media screen and (max-width: 600px) {
.column {
width: 100%;
}
}
/* Style for the thumbnail images */
.thumbnail {
width: 350px; /* or your desired width */
height: 300px; /* or your desired height */
background-size: auto;
background-position: top center;
background-repeat: no-repeat;
cursor: pointer;
transition: opacity 0.3s, transform 0.3s; /* smooth transition for opacity and scale */
}
.thumbnail:hover {
opacity: 0.8; /* slightly fade the image on hover */
transform: scale(1.05); /* slightly enlarge the image on hover */
}
/* Optional: Add a "Click to Enlarge" text on hover */
.thumbnail::after {
content: "Click to Enlarge";
display: block;
position: absolute;
top: 50%;
left: 50%;
transform: translate(-50%, -50%);
color: #000;
font-size: 20px;
text-align: center;
opacity: 0;
transition: opacity 0.3s;
}
.thumbnail:hover::after {
opacity: 1;
}
.small-thumbnail {
width: 175px; /* half the width of the original thumbnail */
height: 150px; /* half the height of the original thumbnail */
background-size: auto;
background-position: top center;
background-repeat: no-repeat;
cursor: pointer;
transition: opacity 0.3s, transform 0.3s;
}
.small-thumbnail:hover {
opacity: 0.8;
transform: scale(1.05);
}
.small-thumbnail::after {
content: "Click to Enlarge";
display: block;
position: absolute;
top: 50%;
left: 50%;
transform: translate(-50%, -50%);
color: #000;
font-size: 15px; /* smaller font size for smaller thumbnails */
text-align: center;
opacity: 0;
transition: opacity 0.3s;
}
.small-thumbnail:hover::after {
opacity: 1;
}
</style>
</head>
<body>
<div class="w3-content" style="max-width:2000px;margin-top:46px">
<div class="w3-container w3-content w3-center w3-padding-64" style="max-width:800px" id="band">
<h2 class="w3-wide"><b>6-DOF Grasp Pose Evaluation and Optimization via Transfer Learning from NeRFs</b></h2>
<p class="w3-justify">We address the problem of robotic grasping of known and unknown objects using implicit
behavior cloning. We train a grasp evaluation model from a small number of demonstrations that outputs
higher values for grasp candidates that are more likely to succeed in grasping. This evaluation model serves
as an objective function, that we maximize to identify successful grasps. Key to our approach is the
utilization of learned implicit representations of visual and geometric features derived from a pre-trained
NeRF. Though trained exclusively in a simulated environment with simplified objects and 4-DoF top-down
grasps, our evaluation model and optimization procedure demonstrate generalization to 6-DoF grasps and novel
objects both in simulation and in real-world settings, without the need for additional data.</p>
</div>
<div class="w3-container w3-content w3-padding-32" style="max-width:800px">
<h2 class="w3-wide">Supplementary Material</h2>
<h3 class="w3-wide">Network Architectures</h3>
<div class="w3-row-padding" style="margin-top:32px">
<div class="w3-margin-bottom">
We implemented our models in Tensorflow 2.9.1. Blocks with rounded corners represent utilities,
activation functions or batch normalization and are not trainable. For these we either use available
implementations from tensorflow, keras or tensorflow-addons or we implement them ourselves.
Blocks with sharp corners represent inputs or trainable variables.
Ocher blocks are layers available in keras or tensorflow.
</div>
<div class="w3-margin-bottom">
We use two main pipelines in our work. The first one is the <b>rendering pipeline</b>. Its functional
purpose is to render novel views of the scene given some input observations with known camera poses. In
our work, we mainly use it to pre-train the NeRF model and to utilize it as a feature extractor for the
grasp evaluation in the <b>grasp evaluation pipeline</b>. The following figures show the basic
architecture of both pipelines.
</div>
<div class="w3-half w3-margin-bottom">
<div class="w3-white">
<p><b>Rendering pipeline</b></p>
</div>
<div class="thumbnail" style="background-image: url('images/rendering_pipeline.png');"
data-alt="Rendering pipeline" id="rendering_pipeline"></div>
</div>
<div class="w3-half w3-margin-bottom">
<div class="w3-white">
<p><b>Grasp evaluation pipeline</b></p>
</div>
<div class="thumbnail" style="background-image: url('images/grasp_evaluation_pipeline.png');"
data-alt="Grasp evaluation pipeline" id="grasp_evaluation_pipeline"></div>
</div>
<div class="w3-margin-bottom">
When training using the <b>rendering pipeline</b>, all trainable variables are trained. When training
using the <b>grasp evaluation pipeline</b>, only the trainable variables of the <b>GraspReadout</b>
are trained.
<br>
Both pipelines employ <b>VisualFeatures</b>, a module that extracts visual features from the input
images. The following figure shows its architecture.
</div>
<div style="padding-left: 8px">
<div class="w3-white">
<p><b>VisualFeatures</b></p>
</div>
<div class="thumbnail" style="background-image: url('images/visual_features.png');"
data-alt="Visual features" id="visual_features"></div>
</div>
<div class="w3-margin-bottom">
Its implementation is analogous to the feature extraction module in VisionNeRF from the paper "Vision
Transformer
for NeRF-Based View Synthesis from a Single Input Image" by Lin et al. (<a
href="https://cseweb.ucsd.edu/~viscomp/projects/VisionNeRF/">https://cseweb.ucsd.edu/~viscomp/projects/VisionNeRF</a>).
The model relies on a VisionTransformer (ViT), which is based on the
implementation in Ross Wightman's timm library (<a
href="https://github.com/rwightman/pytorch-image-models">https://github.com/rwightman/pytorch-image-models</a>)
The following figures show the components used in the VisualFeatures module.
</div>
<div class="w3-quarter w3-margin-bottom">
<div class="w3-white">
<p><b>ConvEncoder</b></p>
</div>
<div class="small-thumbnail" style="background-image: url('images/conv_encoder.png');"
data-alt="Convolutional encoder" id="conv_encoder"></div>
</div>
<div class="w3-quarter w3-margin-bottom">
<div class="w3-white">
<p><b>ViTEncoder</b></p>
</div>
<div class="small-thumbnail" style="background-image: url('images/vit_encoder.png');"
data-alt="ViTEncoder" id="vit_encoder"></div>
</div>
<div class="w3-quarter w3-margin-bottom">
<div class="w3-white">
<p><b>ResnetConvBlock</b></p>
</div>
<div class="small-thumbnail" style="background-image: url('images/resnet_conv_block.png');"
data-alt="Rendering pipeline" id="resnet_conv_block"></div>
</div>
<div class="w3-quarter w3-margin-bottom">
<div class="w3-white">
<p><b>ResnetConvBlock-DS</b></p>
</div>
<div class="small-thumbnail" style="background-image: url('images/resnet_conv_block_ds.png');"
data-alt="ResnetConvBlock-DS" id="resnet_conv_block_ds"></div>
</div>
<div class="w3-quarter w3-margin-bottom">
<div class="w3-white">
<p><b>ViT</b></p>
</div>
<div class="small-thumbnail" style="background-image: url('images/vit.png');"
data-alt="ViT" id="vit"></div>
</div>
<div class="w3-quarter w3-margin-bottom">
<div class="w3-white">
<p><b>PostProcess1</b></p>
</div>
<div class="small-thumbnail" style="background-image: url('images/postprocess_1.png');"
data-alt="PostProcess1" id="postprocess_1"></div>
</div>
<div class="w3-quarter w3-margin-bottom">
<div class="w3-white">
<p><b>PostProcess2</b></p>
</div>
<div class="small-thumbnail" style="background-image: url('images/postprocess_2.png');"
data-alt="PostProcess2" id="postprocess_2"></div>
</div>
<div class="w3-quarter w3-margin-bottom">
<div class="w3-white">
<p><b>PostProcess3</b></p>
</div>
<div class="small-thumbnail" style="background-image: url('images/postprocess_3.png');"
data-alt="PostProcess3" id="postprocess_3"></div>
</div>
<div class="w3-quarter w3-margin-bottom">
<div class="w3-white">
<p><b>PostProcess4</b></p>
</div>
<div class="small-thumbnail" style="background-image: url('images/postprocess_4.png');"
data-alt="PostProcess4" id="postprocess_4"></div>
</div>
<div class="w3-quarter w3-margin-bottom">
<div class="w3-white">
<p><b>PatchEmbed</b></p>
</div>
<div class="small-thumbnail" style="background-image: url('images/patch_embed.png');"
data-alt="PatchEmbed" id="patch_embed"></div>
</div>
<div class="w3-quarter w3-margin-bottom">
<div class="w3-white">
<p><b>TransformerBlock</b></p>
</div>
<div class="small-thumbnail" style="background-image: url('images/transformer_block.png');"
data-alt="TransformerBlock" id="transformer_block"></div>
</div>
<div style="clear: both;"></div>
<div class="w3-margin-bottom">
For novel view synthesis using the <b>rendering pipeline</b>, we need to process single 5-DoF poses
that are sampled along a camera-ray before we can apply volumetric rendering. Just like in VisionNeRF,
we use a two-stage approach consisting of a coarse and a fine stage. The following figure shows the
processing of a single 5-DoF pose.
</div>
<div style="padding-left: 8px">
<div class="w3-white">
<p><b>Process single 5-DoF pose - render</b></p>
</div>
<div class="thumbnail" style="background-image: url('images/process_single.png');"
data-alt="Process single 5-DoF pose - render" id="process_single"></div>
</div>
<div class="w3-margin-bottom">
The <b>MVResNetMLP</b> fuses the information from multiple views into a single feature vector, or into
the scene activations. For rendering, the model uses two <b>MVResNetMLP</b>s: one for the coarse stage
and one for the fine stage. For grasp evaluation we use a single <b>MVResNetMLP</b> that is initialized
with the weights of the fine stage. The following figure shows the architecture of the
<b>MVResNetMLP</b> and its main building block, the <b>ResNetMLPBlock</b>.
</div>
<div class="w3-half w3-margin-bottom">
<div class="w3-white">
<p><b>MVResNetMLP</b></p>
</div>
<div class="thumbnail" style="background-image: url('images/mv_resnet_mlp.png');"
data-alt="MVResNetMLP" id="mv_resnet_mlp"></div>
</div>
<div class="w3-quarter w3-margin-bottom">
<div class="w3-white">
<p><b>ResNetMLPBlock</b></p>
</div>
<div class="small-thumbnail" style="background-image: url('images/resnet_mlp_block.png');"
data-alt="ResNetMLPBlock" id="resnet_mlp_block"></div>
</div>
<div style="clear: both;"></div>
<div class="w3-margin-bottom">
The <b>MVResNetMLP</b> extracts the features from the input images and the poses. These are then further
processed by the readouts. The <b>RenderReadout</b> uses the output embeddings to render the scene,
and the <b>GraspReadout</b> uses the output scene activations to evaluate the grasp candidates. The
following figure shows the architecture of both readout modules and the <b>GraspReadout</b>'s main
building blocks.
</div>
<div class="w3-half w3-margin-bottom">
<div class="w3-white">
<p><b>RenderReadout</b></p>
</div>
<div class="thumbnail" style="background-image: url('images/render_readout.png');"
data-alt="RenderReadout" id="render_readout"></div>
</div>
<div class="w3-half w3-margin-bottom">
<div class="w3-white">
<p><b>GraspReadout</b></p>
</div>
<div class="thumbnail" style="background-image: url('images/grasp_readout.png');"
data-alt="GraspReadout" id="grasp_readout"></div>
</div>
<div class="w3-quarter w3-margin-bottom">
<div class="w3-white">
<p><b>ResNetMLPBlock2</b></p>
</div>
<div class="small-thumbnail" style="background-image: url('images/resnet_mlp_block_2.png');"
data-alt="ResNetMLPBlock2" id="resnet_mlp_block_2"></div>
</div>
<div class="w3-quarter w3-margin-bottom">
<div class="w3-white">
<p><b>ResNetMLPBlock3</b></p>
</div>
<div class="small-thumbnail" style="background-image: url('images/resnet_mlp_block_3.png');"
data-alt="ResNetMLPBlock3" id="resnet_mlp_block_3"></div>
</div>
</div>
</div>
<div class="w3-container w3-content w3-center w3-padding-64" style="max-width:800px">
<h3 class="w3-opacity">Additional content coming soon ... </h3>
</div>
<div class="w3-container w3-content w3-padding-64" style="max-width:800px" id="contact">
<h2 class="w3-wide">Contact</h2>
<div class="w3-row w3-padding-32">
<div class="w3-col m12 w3-large w3-margin-bottom">
<i class="fa fa-user" style="width:30px"></i> Gergely Sóti<br>
<i class="fa fa-envelope" style="width:30px"> </i> [email protected]<br>
<i class="fa fa-map-marker" style="width:30px"></i> Institute for Robotics and Autonomous Systems,
Karlsruhe University of Applied Sciences, 76133 Karlsruhe, Germany<br>
</div>
</div>
</div>
<footer class="w3-container w3-padding-32 w3-center w3-opacity w3-light-grey w3-xlarge">
<p class="w3-medium">Powered by <a href="https://www.w3schools.com/w3css/default.asp" target="_blank">w3.css</a>
</p>
</footer>
</div>
<!-- The Modal -->
<div id="myModal" class="modal">
<div id="modal_caption"></div>
<span class="close">×</span>
<img class="modal-content" id="modal_img">
</div>
<script>
// Get the modal
var modal = document.getElementById("myModal");
var modalImg = document.getElementById("modal_img");
var captionText = document.getElementById("modal_caption");
// Array of image IDs
var imageIds =
["rendering_pipeline", "grasp_evaluation_pipeline", "visual_features", "conv_encoder", "resnet_conv_block",
"resnet_conv_block_ds", "vit_encoder", "vit", "postprocess_1", "postprocess_2", "postprocess_3",
"postprocess_4", "patch_embed", "transformer_block", "process_single", "mv_resnet_mlp",
"resnet_mlp_block", "render_readout", "grasp_readout", "resnet_mlp_block_2", "resnet_mlp_block_3"];
// Add click event listener to each image
imageIds.forEach(function (id) {
var img = document.getElementById(id);
img.onclick = function () {
modal.style.display = "block";
modalImg.src = this.style.backgroundImage.slice(5, -2); // Extract the URL from the background-image property
captionText.innerHTML = this.getAttribute("data-alt");
}
});
// Get the <span> element that closes the modal
var span = document.getElementsByClassName("close")[0];
// When the user clicks on <span> (x), close the modal
span.onclick = function () {
modal.style.display = "none";
}
window.onclick = function (event) {
if (event.target === modal) {
modal.style.display = "none";
}
}
</script>
</body>
</html>