-
Notifications
You must be signed in to change notification settings - Fork 0
/
index.html
534 lines (460 loc) · 28.4 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
<!DOCTYPE html>
<html>
<head>
<meta charset="utf-8">
<meta name="description"
content="AugInsert: Learning Robust Visual-Force Policies via Data Augmentation for Object Assembly Tasks">
<meta name="keywords" content="Contact-Rich Manipulation, Imitation Learning, Data Augmentation">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>AugInsert</title>
<!-- <script async src="https://www.googletagmanager.com/gtag/js?id=G-FV4ZJ9PVSV"></script>
<script>
window.dataLayer = window.dataLayer || [];
function gtag(){dataLayer.push(arguments);}
gtag('js', new Date());
gtag('config', 'G-FV4ZJ9PVSV');
</script> -->
<script>
function updateRealWorldVideo() {
var variation = document.getElementById("task-variation").value;
var video = document.getElementById("real-world-canonical-video");
video.src = "media/results/real_world/" +
"real_world_" +
variation +
".mp4"
video.playbackRate = 1.0;
video.play();
}
function updateAttnVisVideo() {
var attnVisVariation = document.getElementById("task-variation-attnvis").value;
var video = document.getElementById("attnvis-video");
video.src = "media/results/attn_vis/" +
attnVisVariation +
'_attnvis' +
".mp4"
video.playbackRate = 1.0;
video.play();
}
</script>
<link href="https://fonts.googleapis.com/css?family=Google+Sans|Noto+Sans|Castoro"
rel="stylesheet">
<link rel="stylesheet" href="./static/css/bulma.min.css">
<link rel="stylesheet" href="./static/css/bulma-carousel.min.css">
<link rel="stylesheet" href="./static/css/bulma-slider.min.css">
<link rel="stylesheet" href="./static/css/fontawesome.all.min.css">
<link rel="stylesheet"
href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css">
<link rel="stylesheet" href="./static/css/index.css">
<script src="https://ajax.googleapis.com/ajax/libs/jquery/3.5.1/jquery.min.js"></script>
<script defer src="./static/js/fontawesome.all.min.js"></script>
<script src="./static/js/bulma-carousel.min.js"></script>
<script src="./static/js/bulma-slider.min.js"></script>
<script src="./static/js/index.js"></script>
</head>
<body onload="updateRealWorldVideo(); updateAttnVisVideo();">
<nav class="navbar" role="navigation" aria-label="main navigation">
<div class="navbar-brand">
<a role="button" class="navbar-burger" aria-label="menu" aria-expanded="false">
<span aria-hidden="true"></span>
<span aria-hidden="true"></span>
<span aria-hidden="true"></span>
</a>
</div>
<div class="navbar-menu">
<div class="navbar-start" style="flex-grow: 1; justify-content: center;">
<a class="navbar-item" target="_blank" href="https://ryangdiaz.github.io">
<span class="icon">
<i class="fas fa-home"></i>
</span>
</a>
<div class="navbar-item has-dropdown is-hoverable">
<a class="navbar-link">
More Research
</a>
<div class="navbar-dropdown">
<a class="navbar-item" target="_blank" href="https://sites.google.com/view/geometric-peg-in-hole">
Geometric Peg-in-Hole
</a>
</div>
</div>
</div>
</div>
</nav>
<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">AugInsert: Learning Robust Visual-Force Policies via Data Augmentation for Object Assembly Tasks </h1>
<!-- <h3 class="title is-4 conference-authors">In Review</h3> -->
<div class="is-size-5 publication-authors">
<span class="author-block">
<a target="_blank" href="https://ryangdiaz.github.io">Ryan Diaz</a><sup>1</sup>,</span>
<span class="author-block">
Adam Imdieke<sup>1</sup>,</span>
<!-- <a target="_blank">Adam Imdieke</a><sup>1</sup>,</span> -->
<span class="author-block">
<a target="_blank" href="https://uk.linkedin.com/in/vivkjv?challengeId=AQG2hYTlKNvSfgAAAZJn0G9kIMKEUJrV9seua7v50qrxagyTbcx-IKbgDMOOmSBOBCBRQJuNhDqwv0Ew9Gmf9_lax0aK2_Pu2Q&submissionId=3d87fe1f-9438-fc17-f9ad-445166baba51&challengeSource=AgGFQNWOIp7OrQAAAZJn0Hc5q4aBOEUdjS8KFr70hgFKqSaeGG5TSJQzTEWgXLY&challegeType=AgHrX4EpDIW4CgAAAZJn0Hc8sJx6fvobvF2x1gbfALFcqJ2QMq1TwLA&memberId=AgEZ9eO9oHVbzQAAAZJn0HdBhXx60YlsQVN3Ohy85LuNAYc&recognizeDevice=AgHHEpZk1f0w7QAAAZJn0HdGe0TWh6yAhQXJonFwXMUPascvv_v5">Vivek Veeriah</a><sup>2</sup>,</span>
<span class="author-block">
<a target="_blank" href="https://karthikdesingh.com">Karthik Desingh</a><sup>1</sup>
</span>
</div>
<div class="is-size-5 publication-authors">
<span class="author-block"><sup>1</sup>University of Minnesota - Twin Cities,</span>
<span class="author-block"><sup>2</sup>Google DeepMind</span>
</div>
<div class="column has-text-centered">
<div class="publication-links">
<!-- PDF Link. -->
<span class="link-block">
<a target="_blank" href="paper/auginsert_icra2025.pdf"
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>
<!-- Arxiv Link. -->
<span class="link-block">
<a target="_blank" href="https://arxiv.org/abs/2410.14968"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fas fa-file"></i>
</span>
<span>ArXiv</span>
</a>
</span>
<!-- Video Link. -->
<span class="link-block">
<a target="_blank" href="https://www.youtube.com/watch?v=UTA7sefgs2o"
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>
<!-- Talk Link. -->
<!-- <span class="link-block">
<a target="_blank" href="https://youtu.be/QcuXwmQgurE"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fas fa-chalkboard-teacher"></i>
</span>
<span>Talk</span>
</a>
</span> -->
<!-- Code Link. -->
<span class="link-block">
<a target="_blank" href="https://github.com/RyangDiaz/auginsert"
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>
</div>
</div>
<br>
<br>
</div>
</div>
</div>
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
<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>
This paper primarily focuses on learning robust visual-force policies in the context of high-precision object assembly
tasks. Specifically, we focus on the <i>contact phase</i> of the assembly task where both objects (peg and hole) have made
contact and the objective lies in maneuvering the objects to complete the assembly. Moreover, we aim to learn contact-rich
manipulation policies with multisensory inputs on limited expert data by expanding human demonstrations via online data
augmentation.
</p>
<p>
We develop a simulation environment with a dual-arm robot manipulator to evaluate the effect of augmented expert demonstration
data. Our focus is on evaluating the robustness of our model with respect to certain task variations: <strong>grasp pose, peg/hole
shape object body shape, scene appearance, camera pose</strong>, and <strong>force-torque/proprioception noise</strong>. We show that
our proposed data augmentation method helps in learning a multisensory manipulation policy that is robust to unseen instances of these
variations, particularly physical variations such as <strong>grasp pose</strong>. Additionally, our ablative studies show the
significant contribution of force-torque data to the robustness of our model.
</p>
<p>
For additional qualitative results and ablation studies, including experiments carried out in our real-world environment, refer to the
supplementary material on this webpage!
</p>
</div>
</div>
</div>
<br>
<br>
</div>
<div class="columns is-centered has-text-centered">
<div class="column is-two-thirds">
<h2 class="title is-3">Video</h2>
<div class="publication-video">
<iframe src="https://www.youtube.com/embed/UTA7sefgs2o?si=rb-pFjDM_QlrQyeM"
frameborder="0" allow="autoplay; encrypted-media" allowfullscreen></iframe>
</div>
</div>
</div>
</section>
<section class="section">
<div class="container is-max-widescreen">
<div class="rows">
<div class="rows is-centered ">
<div class="row is-full-width">
<h2 class="title is-3">Model Architecture</h2>
<img src="media/figures/model_architecture.png" class="interpolation-image"
alt="Interpolate start reference image." />
</br>
</br>
<p>
Our behavior cloning framework implementation is based off of <a target="_blank" href="https://robomimic.github.io">Robomimic</a>. To encode our observations, we draw upon the success of <a target="_blank" href="https://www.mmintlab.com/research/vtt/">visuotactile transformer</a> encoders and utilize a similar attention-based mechanism for RGB and tactile modality fusion. Rather than performing self-attention directly with the input tokens, we found that introducing a cross-attention step similar to the <a target="_blank" href="https://deepmind.google/discover/blog/building-architectures-that-can-handle-the-worlds-data/">PerceiverIO</a> architecture seemed to work best for our task. We tokenize our inputs by computing linear projections of both visual patches (as in <a target="_blank" href="https://arxiv.org/abs/2010.11929">vision transformers</a>) for RGB inputs and individual readings per timestep for the force-torque input, and then add modality-specific position encodings. We then cross-attend these input tokens with a set of 8 learned latent vectors that then travel through a series of self-attention layers before ultimately being compressed and projected (as in VTT) to an output latent embedding. We encode proprioception with a multilayer perceptron to get an output embedding and concatenate both output embeddings to get the input to the policy network. The policy network is then a multilayer perceptron that outputs 3-dimensional delta actions.
</p>
<br/>
<br/>
<h2 class="title is-3">Supplementary Experiments and Results</h2>
<h3 class="title is-4">Additional Simulation Experiments</h3>
<h4 class="title is-5">Attention Visualization</h4>
<p>
To gain further insight into the information being learned by our model, we visualize the attention weights in the latent vector cross-attention
step of the transformer visuotactile encoder. For each modality, we plot attention weights as proportions of total attention to tokens in that
specific modality averaged over the 8 learned latent vectors. These weights are visualized as heatmaps overlayed on left and right wristview images
for visual attention, and bars for each timestep under the force reading for tactile attention. We also plot the proportion of total attention for
each modality (visual and tactile) during the course of a rollout.
</p>
<br/>
<br/>
<div class="columns"></div>
<div class="column has-text-centered">
<h3 class="title is-5">Attention Visualization: Model Trained on <i>Grasp Pose</i>, <i>Peg/Hole Shape</i>, and <i>Object Body Shape</i></h3>
Model evaluated on
<div class="select is-small">
<select id="task-variation-attnvis" onchange="updateAttnVisVideo()">
<option value="canonical" selected="selected">Canonical (No Variations)</option>
<option value="grasp">Grasp Pose</option>
<option value="peg_hole_shape">Peg/Hole Shape</option>
<option value="body_shape">Object Body Shape</option>
<option value="visual">Scene Appearance</option>
<option value="camera_angle">Camera Pose</option>
<option value="sensor_noise">Sensor Noise</option>
<option value="all_vars">All Variations</option>
</select>
</div>
task variation (videos sped up x2)
<br/>
<br/>
<video id="attnvis-video"
muted
autoplay
loop
width="95%">
<source src="media/results/attn_vis/canonical_attnvis.mp4"
type="video/mp4">
</video>
<p><strong>NOTE:</strong> 5 rollouts per video (50 rollouts for experiment results).</p>
</div>
</div>
<p>
<strong>Takeaways:</strong> Despite our model taking in twice as many visual tokens (72 tokens, 36 per view) as
tactile ones (32 tokens), we observe that tactile attention takes up almost the entire proportion of attention across
the input (as seen in the right-most plot of the videos). This finding provides further evidence to the importance of tactile
information over visual information as discussed in our paper, where we found that removing visual information from our input had
little impact on the robustness of our model. Furthermore, we observe that the visual attention is mostly focused on semantically insignficant parts of the
scene, such as the gripper at the bottom of the view, suggesting that the model is not receiving much useful visual information.
</p>
<br/>
<br/>
<!-- Offline vs. online augmentation -->
<h4 class="title is-5">Comparing Data Augmentation Methods for Model Robustness</h4>
<p>
In an effort to evaluate the validity of the online augmentation method for increasing the robustness of our model, we construct a dataset
of human-generated trajectories with an <i>extended</i> set of visual variations and sensor noise, attempting to emulate a baseline data
augmentation method that applies augmentations independently to each sensory modality offline during training. More specifically, we generate
a dataset with training set variations of <i>Scene Appearance</i> (including object color, floor texture, and lighting), <i>Camera Pose</i>, and
<i>Sensor Noise</i> with 12 augmentations per demonstration, but rather than keep applied variations consistent through each augmented rollout, we apply new instances of <i>Scene Appearance</i>
and <i>Camera Pose</i> variations in each step of the demonstration. We also multiply the force and torque history reading by a random constant (from
0.1 to 2.0) independently determined each frame, following a similar data augmentation strategy used in <a href="https://arxiv.org/abs/2104.14223">InsertionNet</a>.
We denote this dataset as <strong>Expanded Visual+Noise</strong>.
</p>
<br/>
<div class="column has-text-centered">
<video id="expanded_augments_video"
muted
autoplay
loop
width="60%">
<source src="media/video/offline_aug_dataset.mp4" type="video/mp4">
</video>
<p>
Visualization of augmented observations collected for the <strong>Expanded Visual+Noise</strong> Dataset
</p>
</div>
<br/>
<p class="justify"></p>
<img src="media/figures/ablation_training_set_extended.png" class="interpolation-image" width="640" align="right" style="margin:0% 4% " alt="Interpolate start reference image." />
We report % success rate change from the canonical environment success rate on models trained on the <strong>Expanded Visual+Noise</strong> dataset and compare
it with the original training set models from our paper (namely <strong>Visual+Sensor Noise</strong> that does not apply new variation instances per frame and
<strong>Base</strong> that includes <i>Grasp Pose</i> variations).
</p>
<br/>
<p>
<strong>Takeaways:</strong> We observe that our dataset with an expanded set of augmentations independently applied to each sensory modality does not necessarily
improve robustness on most task variations (save for <i>Peg/Hole Shape</i>) compared to the original <strong>Visual+Sensor Noise</strong> dataset that was less extensive in terms of data augmentation. Most crucially, we
do not see a significant improvement on <i>Grasp Pose</i> variations, validating the effect of non-independent multisensory data augmentation via trajectory replay. Thus, we have shown that even
extensive independent augmentation of our multisensory input may not be enough to deal with certain task variations involved in our contact-rich task.
</p>
<br/>
<br/>
<!-- Canonical success rates for all experiments -->
<h4 class="title is-5">Success Rates in Canonical Environment</h4>
<p>
For full transparency for our experiments that involve reporting the % sucess rate change from the <i>Canonical</i> environment, we explicitly report the success rates in the no-variations
<i>Canonical</i> environment, which the % success rate change is based off of, for each trained model. Success rates are averaged across 6 training seeds, with error bars representing
one standard deviation from the mean. It is worth noting that the average % success rate change across the 6 training seeds was calculated by determining the % success rate change
for each individual seed and then calculating the average over those values, rather than calculating the average success rate across the 6 seeds first and then determining the difference of those averages.
</p>
<br/>
<br/>
<p class="justify"></p>
<img src="media/figures/ablation_training_set_wristviews_evals_canonical.png" class="interpolation-image" width="360" align="left" style="margin:2% 0% " alt="Interpolate start reference image." />
<img src="media/figures/ablation_modinput_wristview_canonical_evals_canonical.png" class="interpolation-image" width="360" align="center" style="margin:0% 0% " alt="Interpolate start reference image." />
<img src="media/figures/ablation_modinput_wristview_tclone_evals_canonical.png" class="interpolation-image" width="360" align="right" style="margin:0% 2% " alt="Interpolate start reference image." />
</p>
<br/>
<p>
<strong>Left Plot:</strong> Success rates on <i>Canonical</i> environment for models with different training set variations. This graph corresponds to the results reported in Figure 5 in our paper.
</p>
<br/>
<p>
<strong>Center Plot:</strong> Success rates on <i>Canonical</i> environment for models with different modality input combinations, trained on <i>No Variations</i>. This graph corresponds to the results reported in the top graph of Figure 7 in our paper.
</p>
<br/>
<p>
<strong>Right Plot:</strong> Success rates on <i>Canonical</i> environment for models with different modality input combinations, trained on <i>Grasp Pose</i>, <i>Peg/Hole Shape</i>, and <i>Object Body Shape</i>. This graph corresponds to the results reported in the bottom graph of Figure 7 in our paper. We especially note the instability of performance in the <strong>No Vision</strong> model, which provides context for its omission in the corresponding plot in our paper.
</p>
<br/>
<br/>
<h3 class="title is-4">Real World Experiments</h3>
<!-- Real world setup definition -->
<h4 class="title is-5">Real World Experiment Setup</h4>
<p>
Our real-world task setup is built to mirror our simulation setup as closely as possible. We designate one arm to be compliant,
applying a constant amount of force while the other arm moves according to the actions given to it by the policy. In contrast to
policies trained in simulation, our real-world policies predict 2-dimensional delta actions in the axes perpendicular to the
axis of insertion (rather than 3-dimensional actions that include the axis along the direction of insertion), in order to prevent potentially unsafe interactions that may occur as a result of a premature insertion attempt.
Once the peg and hole are aligned, the compliant arm automatically moves its held object forward to complete the insertion. We train
our real-world models with the same hyperparameters as those in simulation, although we only initiate 1 training seed per model
(rather than 6). Additionally, we evaluate each model at the end of the entire training process, rather than performing training set
rollouts during the training process to determine the best checkpoint. Successes and failures follow the same general criteria as in
simulation, though a human manually annotates successes and failures per trial.
</p>
<br/>
<br/>
<h4 class="title is-5">Real World Task Variation Difficulty</h4>
<!-- Canonical experiments graph -->
<p class="justify">
<img src="media/figures/ablation_real_world_canonical.png" class="interpolation-image" width="640" align="right" style="margin:0% 4% " alt="Interpolate start reference image." />
As a real-world analog to our simulation experiment aimed towards determining the difficulty of generalization for each of our task
variations, we evaluate a real-world policy trained on a dataset of human-generated demonstrations with no applied task variations
on real-world versions of a subset of our task variations. Reported success rates over 20 rollouts can be found in the accompanying figure.
</p>
<br/>
<p>
<strong>Takeaways:</strong> Like in simulation, we observe that <i>Grasp Pose</i> variations seem to be the most difficult to generalize
to, while the model is able to handle the mostly unisensory perturbations of <i>Object Body Shape</i> and <i>Scene Appearance</i> (Object Color and Lighting). We also
notice that our model struggles with <i>Grasp Pose</i> even when rotational grasp variations are removed; we hypothesize that this may be
because a translational offset disrupts the desired behavior of lining up end-effector positions (given from proprioceptive input) in order
to line up the peg and hole (i.e. solving the task can no longer be done by just matching the end-effector positions of the two arms).
Including <i>Grasp Pose</i> variations into the training dataset (as was done in simulation through online augmentation) may also improve
performance in the real world.
</p>
<br/>
<br/>
<!-- Canonical experiments visualization -->
<div class="columns">
<div class="column has-text-centered">
<h3 class="title is-5">Real World Rollouts: Model Trained on <i>No Variations</i></h3>
Model evaluated on
<div class="select is-small">
<select id="task-variation" onchange="updateRealWorldVideo()">
<option value="canonical" selected="selected">Canonical (No Variations)</option>
<option value="grasp">Grasp Pose (XT+ZR)</option>
<option value="body_shape">Object Body Shape</option>
<option value="visual">Visual (Lighting/Object Color)</option>
</select>
</div>
task variation (videos sped up x2)
<br/>
<br/>
<video id="real-world-canonical-video"
muted
autoplay
loop
width="80%">
<source src="media/results/real_world/real_world_canonical.mp4"
type="video/mp4">
</video>
<br/>
<br/>
<p><strong>NOTE:</strong> 5 rollouts per video (20 rollouts for experiment results).</p>
</div>
</div>
<br/>
<br/>
<!-- Modality input ablation graph -->
<h4 class="title is-5">Real World Sensory Input Ablation Study</h4>
<p class="justify">
<img src="media/figures/ablation_real_world_modinput.png" class="interpolation-image" width="640" align="right" style="margin:0% 4% " alt="Interpolate start reference image." />
We conduct a reduced real-world analog to the ablation study in our paper that investigated how much each sensory modality contributed
to overall model performance by training models that only accepted a subset of input modalities (vision, touch, and/or proprioception).
We train real-world policies on a dataset of only human demonstrations and evaluate them on a smaller subset of our real-world task variations.
Reported success rates over 20 rollouts can be found in the accompanying figure.
</p>
<br/>
<p>
<strong>Takeaways:</strong> Like in simulation, we observe that the removal of force-torque data as input (the <strong>No Touch</strong> model) leads to a significant drop in
success rate for all task variations compared to the <strong>Full Model</strong>, including the no-variations <i>Canonical</i> environment. We also see a small drop in performance
for the <strong>No Vision</strong> model, somewhat aligning with our findings in simulation of the insignificance of visual input for our task. Surprisingly, we see performance
increases in all task variations for the <strong>No Prop.</strong> model. We hypothesize that the small ranges of possible end-effector poses in our training dataset due to the
high precision required for our task may cause our models to not learn much useful information from the proprioceptive embedding, though this observation may of course also be
the result of a low sample size of trained models. Averaging the performance of models trained over multiple seeds (as was done in simulation), which was not able to be performed
due to time constraints, may give us some more robust results.
</p>
</div>
</div>
</div>
</section>
<section class="section" id="BibTeX">
<div class="container is-max-widescreen content">
<h2 class="title">BibTeX</h2>
<pre><code>@article{diaz2024auginsert,
title = {AugInsert: Learning Robust Visual-Force Policies via Data Augmentation for Object Assembly Tasks},
author = {Diaz, Ryan and Imdieke, Adam and Veeriah, Vivek and Desingh, Karthik},
booktitle = {arXiv:2410.14968},
year = {2024},
}</code></pre>
</div>
</section>
<footer class="footer">
<div class="container">
<div class="columns is-centered">
<div class="column">
<div class="content has-text-centered">
<p>
This website template was borrowed from <a href="https://github.com/nerfies/nerfies.github.io">NeRFies</a> made by <a href="https://keunhong.com/">Keunhong Park</a> and <a href="https://peract.github.io">PerAct</a> made by <a href="https://mohitshridhar.com">Mohit Shridhar</a>.
</p>
</div>
</div>
</div>
</div>
</footer>
</body>
</html>