-
Notifications
You must be signed in to change notification settings - Fork 1
Expand file tree
/
Copy pathindex.html
More file actions
505 lines (458 loc) · 26.4 KB
/
index.html
File metadata and controls
505 lines (458 loc) · 26.4 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
500
501
502
503
504
505
<!DOCTYPE html>
<html>
<head>
<meta charset="utf-8">
<meta name="description" content="SAFE: Multitask Failure Detection for Vision-Language-Action Models">
<meta name="keywords"
content="SAFE, Vision-Language-Action Model, VLA, Failure Detection, Failure Estimation, Robotics, AI">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>SAFE: Multitask Failure Detection for Vision-Language-Action Models</title>
<!-- Google tag (gtag.js) -->
<script async src="https://www.googletagmanager.com/gtag/js?id=G-3W1SBVVN87"></script>
<script>
window.dataLayer = window.dataLayer || [];
function gtag(){dataLayer.push(arguments);}
gtag('js', new Date());
gtag('config', 'G-3W1SBVVN87');
</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">
<!-- <link rel="icon" href="./static/images/favicon.svg"> -->
<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>
<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" href="https://vla-safe.github.io">
<span class="icon">
<i class="fas fa-home"></i>
</span>
</a>
</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-2 publication-title">SAFE: Multitask Failure Detection for <br>Vision-Language-Action Models</h1>
<h2 class="title is-6 publlication-title">Neural Information Processing Systems (NeurIPS) 2025</h2>
<div class="is-size-6 publication-authors">
<span class="author-block">
<a href="https://georgegu1997.github.io/"><i>Qiao Gu</i></a> <sup>1,2,3</sup>,
</span>
<span class="author-block">
<a href="https://scholar.google.com/citations?user=rG90YVAAAAAJ&hl=zh-CN"><i>Yuanliang Ju</i></a> <sup>1,2,3</sup>,
</span>
<span class="author-block">
<a href="https://owensun2004.github.io/"><i>Shengxiang Sun</i></a> <sup>1,2</sup>,
</span>
<span class="author-block">
<a href="https://www.gilitschenski.org/igor/"><i>Igor Gilitschenski</i></a> <sup>1,2,3</sup>,
</span>
<span class="author-block">
<a href="https://harukins.github.io/"><i>Haruki Nishimura</i></a> <sup>4</sup>,
</span>
<span class="author-block">
<a href="https://mashaitkina.weebly.com/"><i>Masha Itkina</i></a> <sup>4</sup>,
</span>
<span class="author-block">
<a href="https://www.cs.toronto.edu/~florian/"><i>Florian Shkurti</i></a> <sup>1,2,3</sup>,
</span>
</div>
<div class="is-size-6 publication-authors">
<span class="author-block"><sup>1</sup>University of Toronto (UofT),</span>
<span class="author-block"><sup>2</sup>UofT Robotics Institute</span>
<span class="author-block"><sup>3</sup>Vector Institute</span>
<span class="author-block"><sup>4</sup>Toyota Research Institute (TRI)</span>
<br>
</div>
<div class="column has-text-centered">
<div class="publication-links">
<!-- PDF Link. -->
<span class="link-block">
<a href="https://arxiv.org/pdf/2506.09937" class="external-link button is-normal">
<span class="icon">
<i class="fas fa-file-pdf"></i>
</span>
<span>Paper</span>
</a>
</span>
<span class="link-block">
<a href="https://arxiv.org/abs/2506.09937" class="external-link button is-normal">
<span class="icon">
<i class="ai ai-arxiv"></i>
</span>
<span>arXiv</span>
</a>
</span>
<!-- Video Link. -->
<!-- <span class="link-block">
<a href="" class="external-link button is-normal" disabled="true">
<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/vla-safe/SAFE" class="external-link button is-normal">
<span class="icon">
<i class="fab fa-github"></i>
</span>
<span>Code</span>
</a>
</span>
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<section class="hero teaser">
<div class="container is-max-desktop is-full-width">
<div class="columns is-centered">
<video id="teaser" autoplay muted playsinline height="100%">
<source src="./static/videos/safe-teaser.mp4" type="video/mp4">
</video>
</div>
<div class="columns is-centered">
<h2 class="subtitle has-text-centered">
We introduce the <b>multitask failure detection</b> problem for VLA models, and propose <b><span class="coolname">SAFE</span></b>, a failure detector that can detect failures for unseen tasks zero-shot and achieve state-of-the-art performance.
</h2>
</div>
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
<!-- Paper video. -->
<!-- <div class="columns is-centered has-text-centered">
<div class="column is-full">
<h2 class="title is-3">Explainer Video</h2>
<div class="publication-video">
<iframe width="100%" height="473" src="https://www.youtube.com/embed/dWuZyeiOXyM?si=14CuGsvIrmxstIFK" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share" referrerpolicy="strict-origin-when-cross-origin" allowfullscreen></iframe>
</div>
</div>
</div> -->
<!--/ Paper video. -->
<!-- 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>
While vision-language-action models (VLAs) have shown promising robotic behaviors across a diverse set of manipulation tasks, they achieve limited success rates when deployed on novel tasks out-of-the-box. To allow these policies to safely interact with their environments, we need a failure detector that gives a timely alert such that the robot can stop, backtrack, or ask for help. However, existing failure detectors are trained and tested only on one or a few specific tasks, while VLAs require the detector to generalize and detect failures also in unseen tasks and novel environments. In this paper, we introduce the multitask failure detection problem and propose <b><span class="coolname">SAFE</span></b>, a failure detector for generalist robot policies such as VLAs. We analyze the VLA feature space and find that VLAs have sufficient high-level knowledge about task success and failure, which is generic across different tasks. Based on this insight, we design <span class="coolname">SAFE</span> to learn from VLA internal features and predict a single scalar indicating the likelihood of task failure. <span class="coolname">SAFE</span> is trained on both successful and failed rollouts, and is evaluated on unseen tasks. <span class="coolname">SAFE</span> is compatible with different policy architectures. We test it on OpenVLA, π<sub>0</sub>, and π<sub>0</sub>-FAST in both simulated and real-world environments extensively. We compare <span class="coolname">SAFE</span> with diverse baselines and show that <span class="coolname">SAFE</span> achieves state-of-the-art failure detection performance and the best trade-off between accuracy and detection time using conformal prediction.
</p>
</div>
</div>
</div>
<!--/ Abstract. -->
</div>
</section>
<section class="hero is-light is-small">
<div class="hero-body has-text-centered">
<h1 class="title"">
VLA Latent Feature Analysis
</h1>
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column is-full-width">
<div class="content has-text-justified">
<p>
We find that the <b>VLA's internal features capture high-level information about task success and failure, and such information is general across different tasks</b>. As shown in the figure below, when a VLA is failing, even though from different tasks, the features fall in the same <i>failure zone</i>. This motivates <span class="coolname">SAFE</span>, an efficient multitask failure detector that is based on VLA internal features and can generalize to unseen tasks.
</p>
<!-- <p>
In the figure below, we visualize the latent features of <a href="https://github.com/Physical-Intelligence/openpi">π<sub>0</sub>-FAST</a> on the <a href="https://github.com/Lifelong-Robot-Learning/LIBERO">LIBERO-10</a> benchmark.
In (a), features from successful rollouts are shown in blue and those from failed ones are shown in blue-red color gradient.
In (b), we visualizes the same set of t-SNE features, colored by task ID.
In (c), we show two example rollouts over time and mark their corresponding projected features in (a) and (b).
</p> -->
<img src="static/images/vla-feature.png" />
</div>
<br />
</div>
</div>
</div>
</section>
<section class="hero is-light is-small">
<div class="hero-body has-text-centered">
<h1 class="title"">
SAFE: Multitask Failure Detector for VLA Models
</h1>
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column is-full-width">
<div class="content has-text-justified">
<p>
Based on the above observation, we propose <span class="coolname">SAFE</span>, a failure detector that learns from VLA internal features and predicts a single scalar indicating the likelihood of task failure. <span class="coolname">SAFE</span> has 3 main components:
</p>
<ul>
<li><b>Feature Extraction</b>: <span class="coolname">SAFE</span> extracts the latent feature from the last layer of a VLA model. In experiments, we ablate different ways of extracting features and aggregate them into a single feature vector.</li>
<li><b>Learning Failure Detector</b>: <span class="coolname">SAFE</span> sequentially processes the latent feature and predicts a failure score, using an MLP or an LSTM backbone. These models are of 1 or 2 layers to reduce overfitting and improve generalization.</li>
<li><b>Calibration and Deployment</b>: <span class="coolname">SAFE</span> determines a time-varying threshold using functional conformal prediction (CP) on a hold-out calibration set. If the predicted score exceeds the threshold during testing, <span class="coolname">SAFE</span> confidently detects a failure. </li>
</ul>
<img src="static/images/safe-pipeline-v2.jpg" />
</div>
</div>
</div>
</div>
</section>
<section class="hero is-light is-small">
<div class="hero-body has-text-centered">
<h1 class="title"">
Experiments
</h1>
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column is-full-width">
<div class="content has-text-justified">
<p>
We evaluate the following diverse baselines. All the baselines use the same conformal prediction framework as <span class="coolname">SAFE</span> to determine the time-varying threshold.
</p>
<ul>
<li><b>Token Uncertainty</b>: Failure scores are computed based on token-wise uncertainty (probability and entropy). </li>
<li><b>Embedding Distribution</b>: Failure scores are computed based on the embedding distances to the calibration distribution. </li>
<li><b>Sample Consistency</b>: Multiple actions are sampled and failure scores are the inconsistency among the samples. </li>
<li><b>Action Consistency</b>: We adopt <a href="https://arxiv.org/abs/2410.04640">STAC scores</a> and also STAC-single that only uses a single sample per timestep. </li>
</ul>
<p>
We conduct experiments on <a href="https://github.com/openvla/openvla">OpenVLA</a>, <a href="https://github.com/Physical-Intelligence/openpi">π<sub>0</sub> and π<sub>0</sub>-FAST</a> VLA models on <a href="https://github.com/Lifelong-Robot-Learning/LIBERO">LIBERO</a>, <a href="https://github.com/simpler-env/SimplerEnv">SimplerEnv</a> benchmarks and a real-world Franka robot.
</p>
</div>
<div class="content has-text-justified">
<h3 class="is-3">How well do failure detectors distinguish failures from successes?</h3>
<p>
Following the LLM uncertainty quantification literature, we report the area under the ROC curve (ROC-AUC) metric in the following figure.
ROC-AUC results are computed based on the max predicted failure score in each rollout.
This metric averages the performance over all possible thresholds, reflecting the overall failure detection performance regardless of threshold selection.
In the following table, the <b style="color: red;">best</b> and <span style="color: orange; text-decoration: underline;">second best</span> results are highlighted in <b style="color: red;">red</b> and <span style="color: orange; text-decoration: underline;">orange</span>, respectively.
</p>
<img src="./static/images/safe-results-roc.png" />
</div>
<div class="content has-text-justified">
<h3 class="is-3">How do detection accuracy and detection time trade off using functional CP?</h3>
<p>
By varying the significance level α used in functional conformal prediction (CP), we can control the conservativeness of failure detection, which gives a trade-off between detection accuracy and detection time.
In the following figure, we plot the balanced accuracy ((TPR + TNR)/2) against the average detection time for different α values.
</p>
<img src="./static/images/safe-acc-tdet.png" />
</div>
</div>
</div>
</div>
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column">
<div class="content">
<h3 class="is-3">Are the detected failures aligned with human intuition?</h3>
<p>
In the following, we show a few example successful and failed rollouts together with the failure scores predicted by <span class="coolname">SAFE</span>. The green shaded region indicates the failure detection threshold determined by functional conformal prediction. Video frames with red border mean the a failure alert has been raised at that time step.
</p>
</div>
</div>
</div>
<div class="columns is-centered">
<div class="column content">
<video poster="" id="snoopy" controls muted loop playsinline autoplay height="100%" width="100%">
<source src="static/videos/openvla-libero-1-succ.mp4" type="video/mp4">
</video>
<p class="has-text-centered"><span class="succ-text">Success</span></p>
</div>
<div class="column content">
<video poster="" id="snoopy" controls muted loop playsinline autoplay height="100%" width="100%">
<source src="static/videos/openvla-libero-1-fail.mp4" type="video/mp4">
</video>
<p class="has-text-centered"><span class="fail-text">Failure</span>: When the robot gets stuck while picking up alphabet soup, it raises a failure signal</p>
</div>
</div>
<div class="columns is-centered">
<div class="column content">
<video poster="" id="snoopy" controls muted loop playsinline autoplay height="100%" width="100%">
<source src="static/videos/openvla-libero-2-succ.mp4" type="video/mp4">
</video>
<p class="has-text-centered"><span class="succ-text">Success</span></p>
</div>
<div class="column content">
<video poster="" id="snoopy" controls muted loop playsinline autoplay height="100%" width="100%">
<source src="static/videos/openvla-libero-2-fail.mp4" type="video/mp4">
</video>
<p class="has-text-centered"><span class="fail-text">Failure</span>: The robot gets stuck in its initial state.</p>
</div>
</div>
<div class="columns is-centered">
<div class="column content">
<video poster="" id="snoopy" controls muted loop playsinline autoplay height="100%" width="100%">
<source src="static/videos/pi0fast-libero-1-succ.mp4" type="video/mp4">
</video>
<p class="has-text-centered"><span class="succ-text">Success</span></p>
</div>
<div class="column content">
<video poster="" id="snoopy" controls muted loop playsinline autoplay height="100%" width="100%">
<source src="static/videos/pi0fast-libero-1-fail.mp4" type="video/mp4">
</video>
<p class="has-text-centered"><span class="fail-text">Failure</span>: The robot knocks down the tomato sauce and fails grasping, subsequently exhibits dangerous behavior. </p>
</div>
</div>
<div class="columns is-centered">
<div class="column content">
<video poster="" id="snoopy" controls muted loop playsinline autoplay height="100%" width="100%">
<source src="static/videos/pi0fast-libero-2-succ.mp4" type="video/mp4">
</video>
<p class="has-text-centered"><span class="succ-text">Success</span></p>
</div>
<div class="column content">
<video poster="" id="snoopy" controls muted loop playsinline autoplay height="100%" width="100%">
<source src="static/videos/pi0fast-libero-2-fail.mp4" type="video/mp4">
</video>
<p class="has-text-centered"><span class="fail-text">Failure</span>: When the robot attempts to place the bowl, it exhibits unexpected dangerous behavior. </p>
</div>
</div>
<div class="columns is-centered">
<div class="column content">
<video poster="" id="snoopy" controls muted loop playsinline autoplay height="100%" width="100%">
<source src="static/videos/pi0fast-libero-3-succ.mp4" type="video/mp4">
</video>
<p class="has-text-centered"><span class="succ-text">Success</span></p>
</div>
<div class="column content">
<video poster="" id="snoopy" controls muted loop playsinline autoplay height="100%" width="100%">
<source src="static/videos/pi0fast-libero-3-fail.mp4" type="video/mp4">
</video>
<p class="has-text-centered"><span class="fail-text">Failure</span>: The robot misses the insertion attempt and subsequently exhibits unstable behavior. </p>
</div>
</div>
<div class="columns is-centered">
<div class="column content">
<video poster="" id="snoopy" controls muted loop playsinline autoplay height="100%" width="100%">
<source src="static/videos/pi0-libero-1-succ.mp4" type="video/mp4">
</video>
<p class="has-text-centered"><span class="succ-text">Success</span></p>
</div>
<div class="column content">
<video poster="" id="snoopy" controls muted loop playsinline autoplay height="100%" width="100%">
<source src="static/videos/pi0-libero-1-fail.mp4" type="video/mp4">
</video>
<p class="has-text-centered"><span class="fail-text">Failure</span>: The robot gets stuck while attempting to grasp, triggering a failure signal. </p>
</div>
</div>
<div class="columns is-centered">
<div class="column content">
<video poster="" id="snoopy" controls muted loop playsinline autoplay height="100%" width="100%">
<source src="static/videos/pi0-simpler-1-succ.mp4" type="video/mp4">
</video>
<p class="has-text-centered"><span class="succ-text">Success</span></p>
</div>
<div class="column content">
<video poster="" id="snoopy" controls muted loop playsinline autoplay height="100%" width="100%">
<source src="static/videos/pi0-simpler-1-fail.mp4" type="video/mp4">
</video>
<p class="has-text-centered"><span class="fail-text">Failure</span>: The robot repeatedly fails to grasp the carrot. </p>
</div>
</div>
<div class="columns is-centered">
<div class="column content">
<video poster="" id="snoopy" controls muted loop playsinline autoplay height="100%" width="100%">
<source src="static/videos/pi0-fast-real-1-succ.mp4" type="video/mp4">
</video>
<p class="has-text-centered"><span class="succ-text">Success</span>: Note that failure scores stop increasing after the robot finishes the task.</p>
</div>
<div class="column content">
<video poster="" id="snoopy" controls muted loop playsinline autoplay height="100%" width="100%">
<source src="static/videos/pi0-fast-real-1-fail.mp4" type="video/mp4">
</video>
<p class="has-text-centered"><span class="fail-text">Failure</span>: The robot gets stuck while picking up the handle of the lid. </p>
</div>
</div>
<div class="columns is-centered">
<div class="column content">
<video poster="" id="snoopy" controls muted loop playsinline autoplay height="100%" width="100%">
<source src="static/videos/pi0-fast-real-2-succ.mp4" type="video/mp4">
</video>
<p class="has-text-centered"><span class="succ-text">Success</span>: Note that failure scores stop increasing after the robot finishes the task.</p>
</div>
<div class="column content">
<video poster="" id="snoopy" controls muted loop playsinline autoplay height="100%" width="100%">
<source src="static/videos/pi0-fast-real-2-fail.mp4" type="video/mp4">
</video>
<p class="has-text-centered"><span class="fail-text">Failure</span>: The robot repeatedly fails to grasp the carrot despite multiple attempts. </p>
</div>
</div>
</div>
</section>
<section class="section" id="BibTeX">
<div class="container is-max-desktop content">
<h2 class="title">BibTeX</h2>
<pre><code>@article{gu2025safe,
author = {Gu, Qiao and Ju, Yuanliang and Sun, Shengxiang and Gilitschenski, Igor and Nishimura, Haruki and Itkina, Masha and Shkurti, Florian},
title = {SAFE: Multitask Failure Detection for Vision-Language-Action Models},
journal = {arXiv preprint arXiv:2506.09937},
year = {2025},
} </code></pre>
</div>
</section>
<footer class="footer">
<div class="container">
<div class="content has-text-centered">
<a class="icon-link" href="./assets/pdf/_ECCV2024_EgoLifter.pdf">
<i class="fas fa-file-pdf"></i>
</a>
<a class="icon-link" href="https://github.com/concept-graphs" class="external-link" disabled>
<i class="fab fa-github"></i>
</a>
<a rel="license" href="http://creativecommons.org/licenses/by-sa/4.0/">
<img alt="Creative Commons License" style="border-width:0"
src="https://i.creativecommons.org/l/by-sa/4.0/88x31.png" />
</a>
</div>
<div class="columns is-centered">
<div class="column is-8">
<div class="content">
<p>
This website adapted from the Nerfies templates, which is licensed under a <a rel="license"
href="http://creativecommons.org/licenses/by-sa/4.0/">Creative Commons Attribution-ShareAlike 4.0
International License</a>.
</p>
<p>
This means you are free to borrow the <a
href="https://github.com/concept-fusion/concept-fusion.github.io">source code</a> of this website,
we just ask that you link back to the <a href="https://github.com/nerfies/nerfies.github.io">Nerfies
source code</a> in the footer.
Please remember to remove the analytics code included in the header of the website which you do not want
on your website.
</p>
</div>
</div>
</div>
</div>
</footer>
</body>
</html>