Repository navigation
Expand file tree
/
Copy pathindex.html
More file actions
462 lines (441 loc) · 54.3 KB
/
Copy pathindex.html
File metadata and controls
462 lines (441 loc) · 54.3 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
<!doctype html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width,initial-scale=1">
<meta name="description" content="HACo learns force-regulating actions from fingertip tactile and joint-torque feedback for contact-rich dexterous manipulation.">
<meta property="og:title" content="HACo: Learning Haptic Active Compliance for Force-Aware Dexterous Manipulation">
<meta property="og:description" content="HACo learns force-regulating actions from fingertip tactile and joint-torque feedback.">
<meta property="og:image" content="https://opendrivelab.github.io/Haco/public/og.png">
<title>HACo: Learning Haptic Active Compliance for Force-Aware Dexterous Manipulation</title>
<link rel="icon" type="image/svg+xml" href="./favicon.svg">
<link rel="stylesheet" href="app/globals.css?v=20260913-benchmark-ood">
</head>
<body>
<main>
<section class="hero-cover" id="top">
<video autoplay muted loop playsinline controls preload="metadata" poster="/videos/hero/HACo_ICRA2027_demo_v26_green_georgia_ood.jpg" aria-label="HACo demonstration video">
<source src="/videos/hero/HACo_ICRA2027_demo_v26_green_georgia_ood.mp4" type="video/mp4">
</video>
<a class="scroll-cue" href="#article">Scroll to explore ↓</a>
</section>
<section class="article-body" id="article">
<span class="legacy-anchor" id="paper" aria-hidden="true"></span>
<div class="article-layout">
<aside class="article-outline" aria-label="Article outline"><nav>
<div class="outline-group"><span class="outline-group__label">Overview</span><div class="outline-group__links">
<a class="outline-link is-active" href="#abstract" aria-current="location"><span class="outline-link__index">01</span><span>Abstract</span></a>
<a class="outline-link" href="#contributions"><span class="outline-link__index">02</span><span>Contributions</span></a>
</div></div>
<div class="outline-group"><span class="outline-group__label">System</span><div class="outline-group__links">
<a class="outline-link" href="#method"><span class="outline-link__index">03</span><span>Method</span></a>
<a class="outline-link outline-link--sub" href="#architecture"><span class="outline-link__index">·</span><span>Architecture</span></a>
<a class="outline-link outline-link--sub" href="#force-regulated-demonstrations"><span class="outline-link__index">·</span><span>Force-regulated demos</span></a>
<a class="outline-link outline-link--sub" href="#compliant-action-learning"><span class="outline-link__index">·</span><span>Compliant action learning</span></a>
</div></div>
<div class="outline-group"><span class="outline-group__label">Evaluation</span><div class="outline-group__links">
<a class="outline-link" href="#benchmark"><span class="outline-link__index">04</span><span>Benchmark</span></a>
<a class="outline-link" href="#results"><span class="outline-link__index">05</span><span>Experiments</span></a>
<a class="outline-link outline-link--sub" href="#main-comparison"><span class="outline-link__index">·</span><span>Main comparison</span></a>
<a class="outline-link outline-link--sub" href="#perception"><span class="outline-link__index">·</span><span>Haptic sensing</span></a>
<a class="outline-link outline-link--sub" href="#compliance"><span class="outline-link__index">·</span><span>Compliance learning</span></a>
<a class="outline-link outline-link--sub" href="#integration"><span class="outline-link__index">·</span><span>Grounding</span></a>
<a class="outline-link outline-link--sub" href="#wrist-camera"><span class="outline-link__index">·</span><span>Wrist cameras</span></a>
</div></div>
<div class="outline-group"><span class="outline-group__label">Limitation</span><div class="outline-group__links">
<a class="outline-link" href="#failures"><span class="outline-link__index">06</span><span>Failure cases</span></a>
</div></div>
<div class="outline-group"><span class="outline-group__label">Citation</span><div class="outline-group__links">
<a class="outline-link" href="#citation"><span class="outline-link__index">07</span><span>BibTeX</span></a>
</div></div>
</nav></aside>
<article class="article-shell">
<header class="title-block" id="title">
<h1><span class="title-h">H</span><span class="title-a">A</span><span class="title-c">C</span><span class="title-o">o</span>: Learning <span class="title-h">H</span>aptic <span class="title-a">A</span>ctive <span class="title-c">C</span><span class="title-o">o</span>mpliance for Force-Aware Dexterous Manipulation</h1>
<p class="authors">Naisheng Ye<sup>1,2,†</sup>, Yinzhe Zhou<sup>2,3</sup>, Junkai Zhao<sup>2</sup>, Yuhang Lu<sup>1,2</sup><br>Checheng Yu<sup>1</sup>, Zhenjie Yang<sup>1</sup>, Pengwei Wang<sup>2</sup>, and Hongyang Li<sup>1</sup></p>
<p class="affiliation"><sup>1</sup>The University of Hong Kong<br><sup>2</sup>Beijing Academy of Artificial Intelligence (BAAI)<br><sup>3</sup>Johns Hopkins University</p>
<p class="affiliation affiliation-note"><sup>†</sup>Work done during an internship at BAAI.</p>
<div class="article-links">
<a class="article-link" href="https://opendrivelab.github.io/Haco-Page/">☁ Page</a>
<a class="article-link" href="https://arxiv.org/pdf/2609.36596" target="_blank" rel="noreferrer">▤ Paper</a>
<a class="article-link" href="https://github.com/OpenDriveLab/HACo" target="_blank" rel="noreferrer">Code</a>
</div>
</header>
<section id="abstract">
<h2>Abstract</h2>
<div class="abstract-language" lang="en">
<p>Contact-rich dexterous manipulation requires policies to translate physical feedback into motion commands that regulate interaction loads across evolving multi-contact interactions. This requires both haptic observations that capture the contact state and action supervision that demonstrates how motion commands should adapt to it. Existing policies often overlook the complementary roles of fingertip tactile sensing and joint torque. Meanwhile, commonly used action targets include nominal teleoperation commands, which can encode excessive loading, and observed configurations, which omit motion constrained by the object.</p>
<p>We introduce <strong>HACo</strong>, a <strong>Haptic Active Compliance</strong> policy that learns force-regulating actions directly from haptic feedback. Compliance-regulated teleoperation converts operator commands into controller-executable compliant actions that preserve motion intent while regulating interaction loads. HACo learns these compliant actions directly, with their command-state discrepancy providing auxiliary compliant-intent supervision. HACo further combines local contact responses from fingertip tactile sensing with load transmission through the articulated hand captured by joint-torque feedback, including contacts beyond tactile coverage. A Compliance Grounding Module grounds action generation in the evolving haptic state through gated haptic cross-attention, enabling closed-loop force regulation without explicit online contact modeling.</p>
<p>We evaluate HACo on a real-world benchmark spanning multi-contact friction, tangential interaction, fragile curved-surface contact, rotational torque, and deformable-object manipulation. Across 20 trials per task, HACo achieves an 83% mean success rate versus 35% for the strongest evaluated baseline. These results demonstrate HACo's ability to translate complementary haptic feedback into active compliance across diverse force-sensitive dexterous manipulation tasks.</p>
</div>
</section>
<section id="contributions">
<h2>Contributions</h2>
<ul>
<li><strong>We develop complementary haptic perception</strong> that couples fingertip tactile sensing with joint-torque feedback to represent both local contact and loads transmitted through the articulated hand.</li>
<li><strong>We formulate active compliance learning</strong> from regulated demonstrations, combining controller-executable compliant actions, compliant-intent supervision, and the Compliance Grounding Module for haptic-conditioned action generation.</li>
<li><strong>We introduce a real-world dexterous force benchmark</strong> spanning multi-contact friction, tangential interaction, fragile curved-surface contact, rotational torque, and deformable-object manipulation.</li>
</ul>
</section>
<section id="method">
<h2>Method</h2>
<p class="method-lead">HACo learns controller-executable compliant actions from force-regulated demonstrations and conditions action generation on haptic feedback.</p>
<div class="method-detail" id="architecture">
<h3>Architecture</h3>
<p>HACo conditions a flow-based action expert on multi-view images, language, robot state, and haptic histories. Its haptic expert first aligns fingertip wrench and deformation features with joint-torque features from the same finger, then fuses information across fingers into a structured haptic representation. The Compliance Grounding Module lets action features query this representation through gated haptic cross-attention alongside vision-language conditioning. Trained with conditional flow matching, the action expert predicts controller-executable compliant actions with auxiliary compliant-intent supervision.</p>
</div>
<figure class="paper-figure">
<div class="architecture-frame"><img src="/method/haco-architecture.png?v=20260928-website-fig01" width="3091" height="1174" loading="lazy" alt="Overview of HACo with kinematics-aligned haptic perception and gated compliance grounding"></div>
<figcaption><strong>Figure 1.</strong> Overview of HACo with kinematics-aligned haptic perception and gated compliance grounding.</figcaption>
</figure>
<div class="method-detail" id="force-regulated-demonstrations">
<h3>Force-regulated demonstrations</h3>
<p>Motion retargeting provides nominal arm and hand references. Cartesian admittance and fingertip force regulation adjust them before execution, preserving task motion while limiting excessive loads. The resulting compliant references are recorded with synchronized visual, state, tactile, and torque observations.</p>
</div>
<figure class="paper-figure paper-figure--compact">
<div class="architecture-frame"><img src="/method/compliant-teleoperation.png" width="2400" height="2084" loading="lazy" alt="Compliance-regulated teleoperation: retargeting, arm admittance, hand force regulation, and synchronized recording"></div>
<figcaption><strong>Figure 2.</strong> Compliance-regulated teleoperation for force-aware demonstrations.</figcaption>
</figure>
<div class="method-detail" id="compliant-action-learning">
<h3>Compliant action learning</h3>
<p>Observed motion omits commands blocked by contact, while nominal commands can apply excessive force. HACo instead learns the regulated compliant command. Its discrepancy from the observed hand state, <strong>Δq<sup>ci</sup><sub>t</sub> = q<sup>cmp</sup><sub>t</sub> − q<sup>obs</sup><sub>t</sub></strong>, supplies compliant-intent supervision during conditional flow matching.</p>
</div>
<figure class="paper-figure paper-figure--compact">
<div class="architecture-frame"><img src="/method/compliant-intent.png" width="2400" height="1644" loading="lazy" alt="Nominal, observed, and compliant hand configurations under contact, illustrating compliant intent and the correction to the nominal command"></div>
<figcaption><strong>Figure 3.</strong> Hand action semantics under contact: nominal, observed, and compliant configurations.</figcaption>
</figure>
</section>
<section id="benchmark">
<h2>Real-World Force Benchmark</h2>
<p>Most prior work evaluates dexterous manipulation primarily through motion outcomes rather than force regulation. Our benchmark therefore comprises five multi-stage bimanual tasks spanning multi-contact friction, precise tangential-force control, fragile curved-surface contact, rotational torque, and deformable-object compression. While each task emphasizes a distinct interaction regime, all require the policy to complete the intended motion while regulating the physical interaction that makes completion possible.</p>
<div class="benchmark-taxonomy" aria-label="Benchmark force taxonomy">
<div class="active" data-task="poker" aria-current="true" style="--c:var(--task-poker)">FRICTION</div><div data-task="book" style="--c:var(--task-book)">SHEAR</div><div data-task="balloon" style="--c:var(--task-balloon)">CURVATURE</div><div data-task="cap" style="--c:var(--task-cap)">TORQUE</div><div data-task="paste" style="--c:var(--task-paste)">DEFORMATION</div>
</div>
<div class="demo-gallery">
<div class="demo-pills" role="tablist" aria-label="Benchmark tasks">
<button class="demo-pill active" type="button" data-task="poker">Insert poker</button>
<button class="demo-pill" type="button" data-task="book">Open book</button>
<button class="demo-pill" type="button" data-task="balloon">Draw on balloon</button>
<button class="demo-pill" type="button" data-task="cap">Unscrew cap</button>
<button class="demo-pill" type="button" data-task="paste">Squeeze toothpaste</button>
</div>
<div class="demo-view"><video id="benchmark-video" controls muted loop playsinline preload="none" poster="public/videos/benchmark/poker/2.jpg" aria-label="Insert poker rollout 2"><source data-src="public/videos/benchmark/poker/2.mp4" type="video/mp4"></video></div>
<div class="demo-caption"><div class="demo-caption__head"><b id="demo-title">Insert poker · multi-contact friction</b><div class="demo-video-pills" id="demo-video-pills" role="tablist" aria-label="Insert poker rollout selection"></div></div><div class="demo-ood-detail" id="demo-ood-detail" hidden><span>OOD</span><strong id="demo-ood-text"></strong></div><p id="demo-description">Begin with two playing cards lying on the black platform. Use the right hand to pick up one card and transfer it to the left hand. Hold the first card upright and steady with the left hand. Then use the right hand to pick up the remaining card, align it with the card held in the left hand, and slide it into the same grip so that both cards are held together. Finish with both cards securely held in the left hand and the right hand released. Do not bend or drop either card.</p></div>
</div>
</section>
<section id="results">
<h2>Experiments</h2>
<p class="results-overview">We evaluate whether HACo improves force-sensitive dexterous manipulation over visuomotor and tactile-augmented policies, then isolate the contributions of haptic perception, compliant-action learning, compliance grounding, and wrist-mounted visual observations. Every entry reports successful physical rollouts out of 20; Mean is the macro-average across the five tasks.</p>
<div class="experiment-group" id="main-comparison">
<h3>Main comparison</h3><p>Compare HACo with visuomotor and tactile-augmented policies under the same five-task, 20-rollout evaluation protocol.</p>
<figure class="result-figure">
<div class="result-figure__frame">
<div class="result-figure__note"><span>Task success rate by method</span><span>20 evaluation trials per task</span></div>
<div class="grouped-legend"><span><i style="--c:var(--task-poker)"></i>Insert poker</span><span><i style="--c:var(--task-book)"></i>Open book</span><span><i style="--c:var(--task-balloon)"></i>Draw on balloon</span><span><i style="--c:var(--task-cap)"></i>Unscrew cap</span><span><i style="--c:var(--task-paste)"></i>Squeeze toothpaste</span></div>
<div class="grouped-chart">
<div class="model-group"><div class="model-bars"><i class="task-bar" tabindex="0" role="img" data-value="15%" data-tooltip="Insert poker · 15%" aria-label="Insert poker: 3/20, 15%" style="--h:15%;--c:var(--task-poker)"></i><i class="task-bar" tabindex="0" role="img" data-value="0%" data-tooltip="Open book · 0%" aria-label="Open book: 0/20, 0%" style="--h:0%;--c:var(--task-book)"></i><i class="task-bar" tabindex="0" role="img" data-value="5%" data-tooltip="Draw on balloon · 5%" aria-label="Draw on balloon: 1/20, 5%" style="--h:5%;--c:var(--task-balloon)"></i><i class="task-bar" tabindex="0" role="img" data-value="35%" data-tooltip="Unscrew cap · 35%" aria-label="Unscrew cap: 7/20, 35%" style="--h:35%;--c:var(--task-cap)"></i><i class="task-bar" tabindex="0" role="img" data-value="20%" data-tooltip="Squeeze toothpaste · 20%" aria-label="Squeeze toothpaste: 4/20, 20%" style="--h:20%;--c:var(--task-paste)"></i></div><span class="model-label">GR00T</span></div>
<div class="model-group"><div class="model-bars"><i class="task-bar" tabindex="0" role="img" data-value="25%" data-tooltip="Insert poker · 25%" aria-label="Insert poker: 5/20, 25%" style="--h:25%;--c:var(--task-poker)"></i><i class="task-bar" tabindex="0" role="img" data-value="10%" data-tooltip="Open book · 10%" aria-label="Open book: 2/20, 10%" style="--h:10%;--c:var(--task-book)"></i><i class="task-bar" tabindex="0" role="img" data-value="5%" data-tooltip="Draw on balloon · 5%" aria-label="Draw on balloon: 1/20, 5%" style="--h:5%;--c:var(--task-balloon)"></i><i class="task-bar" tabindex="0" role="img" data-value="45%" data-tooltip="Unscrew cap · 45%" aria-label="Unscrew cap: 9/20, 45%" style="--h:45%;--c:var(--task-cap)"></i><i class="task-bar" tabindex="0" role="img" data-value="25%" data-tooltip="Squeeze toothpaste · 25%" aria-label="Squeeze toothpaste: 5/20, 25%" style="--h:25%;--c:var(--task-paste)"></i></div><span class="model-label">GR00T + Tactile</span></div>
<div class="model-group"><div class="model-bars"><i class="task-bar" tabindex="0" role="img" data-value="0%" data-tooltip="Insert poker · 0%" aria-label="Insert poker: 0/20, 0%" style="--h:0%;--c:var(--task-poker)"></i><i class="task-bar" tabindex="0" role="img" data-value="5%" data-tooltip="Open book · 5%" aria-label="Open book: 1/20, 5%" style="--h:5%;--c:var(--task-book)"></i><i class="task-bar" tabindex="0" role="img" data-value="0%" data-tooltip="Draw on balloon · 0%" aria-label="Draw on balloon: 0/20, 0%" style="--h:0%;--c:var(--task-balloon)"></i><i class="task-bar" tabindex="0" role="img" data-value="10%" data-tooltip="Unscrew cap · 10%" aria-label="Unscrew cap: 2/20, 10%" style="--h:10%;--c:var(--task-cap)"></i><i class="task-bar" tabindex="0" role="img" data-value="5%" data-tooltip="Squeeze toothpaste · 5%" aria-label="Squeeze toothpaste: 1/20, 5%" style="--h:5%;--c:var(--task-paste)"></i></div><span class="model-label">ViTacFormer</span></div>
<div class="model-group"><div class="model-bars"><i class="task-bar" tabindex="0" role="img" data-value="20%" data-tooltip="Insert poker · 20%" aria-label="Insert poker: 4/20, 20%" style="--h:20%;--c:var(--task-poker)"></i><i class="task-bar" tabindex="0" role="img" data-value="30%" data-tooltip="Open book · 30%" aria-label="Open book: 6/20, 30%" style="--h:30%;--c:var(--task-book)"></i><i class="task-bar" tabindex="0" role="img" data-value="10%" data-tooltip="Draw on balloon · 10%" aria-label="Draw on balloon: 2/20, 10%" style="--h:10%;--c:var(--task-balloon)"></i><i class="task-bar" tabindex="0" role="img" data-value="60%" data-tooltip="Unscrew cap · 60%" aria-label="Unscrew cap: 12/20, 60%" style="--h:60%;--c:var(--task-cap)"></i><i class="task-bar" tabindex="0" role="img" data-value="55%" data-tooltip="Squeeze toothpaste · 55%" aria-label="Squeeze toothpaste: 11/20, 55%" style="--h:55%;--c:var(--task-paste)"></i></div><span class="model-label">T-Rex</span></div>
<div class="model-group is-ours"><div class="model-bars"><i class="task-bar" tabindex="0" role="img" data-value="90%" data-tooltip="Insert poker · 90%" aria-label="Insert poker: 18/20, 90%" style="--h:90%;--c:var(--task-poker)"></i><i class="task-bar" tabindex="0" role="img" data-value="85%" data-tooltip="Open book · 85%" aria-label="Open book: 17/20, 85%" style="--h:85%;--c:var(--task-book)"></i><i class="task-bar" tabindex="0" role="img" data-value="70%" data-tooltip="Draw on balloon · 70%" aria-label="Draw on balloon: 14/20, 70%" style="--h:70%;--c:var(--task-balloon)"></i><i class="task-bar" tabindex="0" role="img" data-value="95%" data-tooltip="Unscrew cap · 95%" aria-label="Unscrew cap: 19/20, 95%" style="--h:95%;--c:var(--task-cap)"></i><i class="task-bar" tabindex="0" role="img" data-value="75%" data-tooltip="Squeeze toothpaste · 75%" aria-label="Squeeze toothpaste: 15/20, 75%" style="--h:75%;--c:var(--task-paste)"></i></div><span class="model-label">HACo</span></div>
</div>
</div>
<figcaption>Table 1a. Task-wise success-rate summary across the five force-sensitive tasks.</figcaption>
</figure>
<!-- Radar chart temporarily hidden
<figure class="result-figure radar-figure">
<div class="result-figure__frame radar-figure__frame">
<div class="result-figure__note"><span>Force-regulation profile</span><span>Normalized progress score</span></div>
<svg class="radar-chart" viewBox="0 0 560 430" role="img" aria-labelledby="radar-title-static radar-description-static" style="display:block;width:min(100%,620px);height:auto;margin:0 auto;overflow:visible">
<title id="radar-title-static">Five-task progress profile for HACo and evaluated baselines</title>
<desc id="radar-description-static">A five-axis radar chart comparing normalized progress scores for friction, tangential interaction, fragile contact, rotational torque, and deformable-object manipulation. HACo leads on all five axes.</desc>
<g class="radar-grid" fill="rgba(255,255,255,.42)" stroke="rgba(54,73,67,.16)" stroke-width="1">
<polygon points="280,180 318,207.6 303.5,252.4 256.5,252.4 242,207.6"></polygon>
<polygon points="280,140 356.1,195.3 327,284.7 233,284.7 203.9,195.3"></polygon>
<polygon points="280,100 394.1,182.9 350.5,317.1 209.5,317.1 165.9,182.9"></polygon>
<polygon points="280,60 432.2,170.6 374.1,349.4 185.9,349.4 127.8,170.6"></polygon>
<line x1="280" y1="220" x2="280" y2="60"></line>
<line x1="280" y1="220" x2="432.2" y2="170.6"></line>
<line x1="280" y1="220" x2="374.1" y2="349.4"></line>
<line x1="280" y1="220" x2="185.9" y2="349.4"></line>
<line x1="280" y1="220" x2="127.8" y2="170.6"></line>
<text x="287" y="184" fill="rgba(54,73,67,.46)" stroke="none" font-size="9">25%</text><text x="287" y="144" fill="rgba(54,73,67,.46)" stroke="none" font-size="9">50%</text><text x="287" y="104" fill="rgba(54,73,67,.46)" stroke="none" font-size="9">75%</text><text x="287" y="64" fill="rgba(54,73,67,.46)" stroke="none" font-size="9">100%</text>
</g>
<g class="radar-series-group">
<polygon class="radar-series" tabindex="0" aria-label="GR00T: 15, 0, 5, 35, 20 percent across the five benchmark tasks" points="280,196 280,220 284.7,226.5 247.1,265.3 249.6,210.1" fill="#8097a1" fill-opacity=".07" stroke="#8097a1" stroke-width="2.2" style="--radar-color:#8097a1;--radar-delay:0ms"></polygon>
<polygon class="radar-series" tabindex="0" aria-label="GR00T plus Tactile: 25, 10, 5, 45, 25 percent across the five benchmark tasks" points="280,180 295.2,215.1 284.7,226.5 237.7,278.2 242,207.6" fill="#d48149" fill-opacity=".07" stroke="#d48149" stroke-width="2.2" style="--radar-color:#d48149;--radar-delay:70ms"></polygon>
<polygon class="radar-series" tabindex="0" aria-label="ViTacFormer: 0, 5, 0, 10, 5 percent across the five benchmark tasks" points="280,220 287.6,217.5 280,220 270.6,232.9 272.4,217.5" fill="#5f9466" fill-opacity=".07" stroke="#5f9466" stroke-width="2.2" style="--radar-color:#5f9466;--radar-delay:140ms"></polygon>
<polygon class="radar-series" tabindex="0" aria-label="T-Rex: 20, 30, 10, 60, 55 percent across the five benchmark tasks" points="280,188 325.7,205.2 289.4,232.9 223.6,297.7 196.3,192.8" fill="#447bb6" fill-opacity=".07" stroke="#447bb6" stroke-width="2.2" style="--radar-color:#447bb6;--radar-delay:210ms"></polygon>
<polygon class="radar-series is-haco" tabindex="0" aria-label="HACo: 90, 85, 70, 95, 75 percent across the five benchmark tasks" points="280,76 409.3,178 345.8,310.6 190.7,342.9 165.9,182.9" fill="#367e68" fill-opacity=".15" stroke="#367e68" stroke-width="4" style="--radar-color:#367e68;--radar-delay:280ms"></polygon>
</g>
<g class="radar-axis-labels" aria-hidden="true" fill="#25302c" font-size="13" font-weight="700">
<text x="280" y="27" text-anchor="middle">Friction</text>
<text x="475" y="170" text-anchor="start">Tangential</text>
<text x="414" y="402" text-anchor="middle">Fragile contact</text>
<text x="146" y="402" text-anchor="middle">Rotational torque</text>
<text x="85" y="170" text-anchor="end">Deformable</text>
</g>
</svg>
<div class="radar-legend" aria-label="Compared methods" style="display:flex;flex-wrap:wrap;justify-content:center;gap:9px 17px">
<span style="display:inline-flex;align-items:center;gap:6px"><i style="display:inline-block;width:18px;height:3px;border-radius:999px;background:#8097a1"></i>GR00T</span>
<span style="display:inline-flex;align-items:center;gap:6px"><i style="display:inline-block;width:18px;height:3px;border-radius:999px;background:#d48149"></i>GR00T + Tactile</span>
<span style="display:inline-flex;align-items:center;gap:6px"><i style="display:inline-block;width:18px;height:3px;border-radius:999px;background:#5f9466"></i>ViTacFormer</span>
<span style="display:inline-flex;align-items:center;gap:6px"><i style="display:inline-block;width:18px;height:3px;border-radius:999px;background:#447bb6"></i>T-Rex</span>
<span class="is-haco" style="display:inline-flex;align-items:center;gap:6px;color:#367e68;font-weight:800"><i style="display:inline-block;width:18px;height:3px;border-radius:999px;background:#367e68"></i>HACo</span>
</div>
<div class="radar-highlight"><span>HACo mean</span><strong>83%</strong></div>
</div>
<figcaption>Task-wise progress profile across five force-sensitive interaction regimes.</figcaption>
</figure>
-->
<div class="table-wrap"><table class="result-table">
<thead><tr><th>Method</th><th>Insert poker</th><th>Open book</th><th>Draw on balloon</th><th>Unscrew cap</th><th>Squeeze toothpaste</th><th>Mean</th></tr></thead>
<tbody>
<tr><th>GR00T</th><td>3/20</td><td>0/20</td><td>1/20</td><td>7/20</td><td>4/20</td><td>15%</td></tr>
<tr><th>GR00T + Tactile</th><td>5/20</td><td>2/20</td><td>1/20</td><td>9/20</td><td>5/20</td><td>22%</td></tr>
<tr><th>ViTacFormer</th><td>0/20</td><td>1/20</td><td>0/20</td><td>2/20</td><td>1/20</td><td>4%</td></tr>
<tr><th>T-Rex</th><td>4/20</td><td>6/20</td><td>2/20</td><td>12/20</td><td>11/20</td><td>35%</td></tr>
<tr class="ours"><th>HACo</th><td>18/20</td><td>17/20</td><td>14/20</td><td>19/20</td><td>15/20</td><td>83%</td></tr>
</tbody>
</table></div><p class="table-caption">Table 1b. Numerical policy comparison on the real-world force-sensitive dexterous manipulation benchmark.</p>
<div class="experiment-analysis"><span>Insight</span><p>HACo achieves an 83% mean success rate and leads on all five tasks, exceeding T-Rex by 48 percentage points and GR00T by 68. It also reduces mean fingertip force by 19% relative to GR00T, showing that the higher task success is accompanied by more compliant interaction. The modest gain from GR00T to GR00T + Tactile (15% to 22%) further indicates that simple tactile concatenation is not enough; this favors connecting haptic feedback to the evolving action representation.</p></div>
</div>
<div class="experiment-group" id="perception">
<h3>Haptic perception ablation</h3><p>Isolate the contributions of fingertip tactile and joint-torque feedback and compare factorized and coupled haptic representations.</p>
<figure class="result-figure">
<div class="result-figure__frame"><div class="result-figure__note"><span>Macro mean by perception configuration</span><span>Complete five-task results</span></div><div class="horizontal-bars">
<div class="horizontal-bar"><span class="horizontal-bar__label">w/o Haptic Feedback</span><div class="horizontal-bar__track"><i class="horizontal-bar__fill" style="--value:27%"></i></div><span class="horizontal-bar__value">27%</span></div>
<div class="horizontal-bar"><span class="horizontal-bar__label">w/o Tactile Feedback</span><div class="horizontal-bar__track"><i class="horizontal-bar__fill" style="--value:45%"></i></div><span class="horizontal-bar__value">45%</span></div>
<div class="horizontal-bar"><span class="horizontal-bar__label">w/o Torque Feedback</span><div class="horizontal-bar__track"><i class="horizontal-bar__fill" style="--value:68%"></i></div><span class="horizontal-bar__value">68%</span></div>
<div class="horizontal-bar"><span class="horizontal-bar__label">w/o Coupled Encoding</span><div class="horizontal-bar__track"><i class="horizontal-bar__fill" style="--value:70%"></i></div><span class="horizontal-bar__value">70%</span></div>
<div class="horizontal-bar is-ours"><span class="horizontal-bar__label">HACo</span><div class="horizontal-bar__track"><i class="horizontal-bar__fill" style="--value:83%"></i></div><span class="horizontal-bar__value">83%</span></div>
</div></div><figcaption>Table 2a. Mean success-rate summary by haptic-perception configuration.</figcaption>
</figure>
<div class="table-wrap"><table class="result-table">
<thead><tr><th>Haptic perception</th><th>Insert poker</th><th>Open book</th><th>Draw on balloon</th><th>Unscrew cap</th><th>Squeeze toothpaste</th><th>Mean</th></tr></thead>
<tbody>
<tr><th>w/o Haptic Feedback</th><td>7/20</td><td>2/20</td><td>3/20</td><td>8/20</td><td>7/20</td><td>27%</td></tr>
<tr><th>w/o Tactile Feedback</th><td>8/20</td><td>5/20</td><td>7/20</td><td>13/20</td><td>12/20</td><td>45%</td></tr>
<tr><th>w/o Torque Feedback</th><td>15/20</td><td>16/20</td><td>11/20</td><td>14/20</td><td>12/20</td><td>68%</td></tr>
<tr><th>w/o Coupled Encoding</th><td>17/20</td><td>14/20</td><td>12/20</td><td>13/20</td><td>14/20</td><td>70%</td></tr>
<tr class="ours"><th>HACo</th><td>18/20</td><td>17/20</td><td>14/20</td><td>19/20</td><td>15/20</td><td>83%</td></tr>
</tbody>
</table></div><p class="table-caption">Table 2b. Numerical results for the haptic-perception ablation.</p>
<div class="experiment-analysis"><span>Insight</span><p>Removing all haptic feedback drops the mean success rate from 83% to 27%, confirming that physical observations are essential. Tactile feedback alone reaches 68%, compared with 45% for joint torque alone, so local contact is the dominant cue. Joint torque remains complementary: adding it improves every task and raises cap removal from 70% to 95%. Coupled encoding provides a further gain from 70% to 83%, supporting kinematics-aligned fusion across the hand.</p></div>
</div>
<div class="experiment-group" id="compliance">
<h3>Compliance learning ablation</h3><p>Progressively introduce controller-executable compliant actions and auxiliary compliant-intent supervision from regulated demonstrations.</p>
<figure class="result-figure"><div class="result-figure__frame"><div class="result-figure__note"><span>Macro mean by compliant-action learning stage</span><span>Complete five-task results</span></div><div class="horizontal-bars">
<div class="horizontal-bar"><span class="horizontal-bar__label">Nominal Action</span><div class="horizontal-bar__track"><i class="horizontal-bar__fill" style="--value:59%"></i></div><span class="horizontal-bar__value">59%</span></div>
<div class="horizontal-bar"><span class="horizontal-bar__label">Compliant Action</span><div class="horizontal-bar__track"><i class="horizontal-bar__fill" style="--value:73%"></i></div><span class="horizontal-bar__value">73%</span></div>
<div class="horizontal-bar is-ours"><span class="horizontal-bar__label">+ Compliance-Intent Supervision</span><div class="horizontal-bar__track"><i class="horizontal-bar__fill" style="--value:83%"></i></div><span class="horizontal-bar__value">83%</span></div>
</div></div><figcaption>Table 3a. Mean success-rate summary by compliant-action learning stage.</figcaption></figure>
<div class="table-wrap"><table class="result-table result-table--components">
<thead><tr><th>Configuration</th><th class="component-head">Compliant Action</th><th class="component-head component-head--cis">CIS</th><th>Insert poker</th><th>Open book</th><th>Draw on balloon</th><th>Unscrew cap</th><th>Squeeze toothpaste</th><th>Mean</th></tr></thead>
<tbody>
<tr><th>Nominal Action</th><td class="component-cell">—</td><td class="component-cell component-cell--cis">—</td><td>12/20</td><td>11/20</td><td>9/20</td><td>14/20</td><td>13/20</td><td>59%</td></tr>
<tr><th>Compliant Action</th><td class="component-cell is-enabled">✓</td><td class="component-cell component-cell--cis">—</td><td>15/20</td><td>15/20</td><td>13/20</td><td>16/20</td><td>14/20</td><td>73%</td></tr>
<tr class="ours"><th>+ Compliance-Intent Supervision</th><td class="component-cell is-enabled">✓</td><td class="component-cell component-cell--cis is-enabled">✓</td><td>18/20</td><td>17/20</td><td>14/20</td><td>19/20</td><td>15/20</td><td>83%</td></tr>
</tbody>
</table></div><p class="table-caption">Table 3b. Numerical results for the compliance-learning ablation.</p>
<div class="experiment-analysis"><span>Insight</span><p>Keeping the compliant action target but removing compliant-intent supervision lowers the mean from 83% to 73%. Replacing that target with the nominal teleoperation command lowers it further to 59%. The larger loss from changing the action target shows that the controller-executable, force-regulated reference is the primary contribution, while compliant-intent supervision adds a complementary benefit by preserving the force-regulating command-state discrepancy.</p></div>
</div>
<div class="experiment-group" id="integration">
<h3>Compliance grounding ablation</h3><p>Hold haptic observations and compliant action targets fixed while varying how haptic state conditions action generation.</p>
<figure class="result-figure"><div class="result-figure__frame"><div class="result-figure__note"><span>Macro mean by compliance grounding mechanism</span><span>Complete five-task results</span></div><div class="horizontal-bars">
<div class="horizontal-bar"><span class="horizontal-bar__label">Visuo–Haptic Fusion</span><div class="horizontal-bar__track"><i class="horizontal-bar__fill" style="--value:41%"></i></div><span class="horizontal-bar__value">41%</span></div>
<div class="horizontal-bar"><span class="horizontal-bar__label">Action-Suffix Fusion</span><div class="horizontal-bar__track"><i class="horizontal-bar__fill" style="--value:47%"></i></div><span class="horizontal-bar__value">47%</span></div>
<div class="horizontal-bar"><span class="horizontal-bar__label">Compliance Attention</span><div class="horizontal-bar__track"><i class="horizontal-bar__fill" style="--value:66%"></i></div><span class="horizontal-bar__value">66%</span></div>
<div class="horizontal-bar is-ours"><span class="horizontal-bar__label">Gated Compliance Attention</span><div class="horizontal-bar__track"><i class="horizontal-bar__fill" style="--value:83%"></i></div><span class="horizontal-bar__value">83%</span></div>
</div></div><figcaption>Table 4a. Mean success-rate summary by compliance-grounding mechanism.</figcaption></figure>
<div class="table-wrap"><table class="result-table">
<thead><tr><th>Compliance grounding</th><th>Insert poker</th><th>Open book</th><th>Draw on balloon</th><th>Unscrew cap</th><th>Squeeze toothpaste</th><th>Mean</th></tr></thead>
<tbody>
<tr><th>Visuo–Haptic Fusion</th><td>10/20</td><td>6/20</td><td>7/20</td><td>10/20</td><td>8/20</td><td>41%</td></tr>
<tr><th>Action-Suffix Fusion</th><td>14/20</td><td>9/20</td><td>6/20</td><td>11/20</td><td>7/20</td><td>47%</td></tr>
<tr><th>Compliance Attention</th><td>14/20</td><td>15/20</td><td>11/20</td><td>16/20</td><td>10/20</td><td>66%</td></tr>
<tr class="ours"><th>Gated Compliance Attention</th><td>18/20</td><td>17/20</td><td>14/20</td><td>19/20</td><td>15/20</td><td>83%</td></tr>
</tbody>
</table></div><p class="table-caption">Table 4b. Numerical results for the compliance-grounding ablation.</p>
<div class="experiment-analysis"><span>Insight</span><p>Direct Visuo-Haptic Fusion and Action-Suffix Fusion reach only 41% and 47%, while dedicated haptic cross-attention reaches 66% and the complete gated CGM reaches 83%. Cross-attention gives the evolving action representation explicit access to contact evidence without disrupting the pretrained visual or action streams. The zero-initialized gate adds another 17 points by preserving pretrained computation at the start of adaptation and regulating reliance on noisy or intermittent haptic signals.</p></div>
</div>
<div class="experiment-group" id="wrist-camera">
<h3>Wrist-camera ablation</h3><p>Remove both wrist cameras while keeping the haptic observations and HACo policy otherwise unchanged.</p>
<figure class="result-figure"><div class="result-figure__frame"><div class="result-figure__note"><span>Macro mean with and without wrist cameras</span><span>Complete five-task results</span></div><div class="horizontal-bars">
<div class="horizontal-bar"><span class="horizontal-bar__label">w/o Wrist Cameras</span><div class="horizontal-bar__track"><i class="horizontal-bar__fill" style="--value:76%"></i></div><span class="horizontal-bar__value">76%</span></div>
<div class="horizontal-bar is-ours"><span class="horizontal-bar__label">w/ Wrist Cameras</span><div class="horizontal-bar__track"><i class="horizontal-bar__fill" style="--value:83%"></i></div><span class="horizontal-bar__value">83%</span></div>
</div></div><figcaption>Table 5a. Mean success-rate summary with and without wrist cameras.</figcaption></figure>
<div class="table-wrap"><table class="result-table">
<thead><tr><th>Camera observation</th><th>Insert poker</th><th>Open book</th><th>Draw on balloon</th><th>Unscrew cap</th><th>Squeeze toothpaste</th><th>Mean</th></tr></thead>
<tbody>
<tr><th>w/o Wrist Cameras</th><td>13/20</td><td>15/20</td><td>14/20</td><td>18/20</td><td>16/20</td><td>76%</td></tr>
<tr class="ours"><th>w/ Wrist Cameras</th><td>18/20</td><td>17/20</td><td>14/20</td><td>19/20</td><td>15/20</td><td>83%</td></tr>
</tbody>
</table></div><p class="table-caption">Table 5b. Numerical results for the wrist-camera ablation.</p>
<div class="experiment-analysis"><span>Insight</span><p>Removing both wrist cameras reduces the mean success rate by only 7 points, from 83% to 76%. This modest drop suggests that haptic feedback recovers contact state that is difficult to observe under visual occlusion. Wrist views remain useful, but the result indicates that sensing anchored at the interaction interface provides a more direct and robust contact representation.</p></div>
</div>
</section>
<section id="failures">
<h2>Failure Cases & Limitations</h2>
<p>Representative unsuccessful rollouts reveal distinct limitations in contact-rich manipulation.</p>
<div class="failure-grid">
<article class="failure-card"><video class="lazy-video" controls muted loop playsinline preload="none" poster="/videos/failures/open-book.jpg" aria-label="Open Book failure rollout"><source data-src="/videos/failures/open-book.mp4" type="video/mp4"></video><div class="failure-card__copy"><span class="failure-card__type">Failure F1 · Open Book</span><h3>Insufficient Tangential Separation</h3><p>Because the thumb fails to enter the middle of the page stack, it cannot establish enough tangential traction to separate the target section. The pages therefore move together instead of opening at the intended location.</p></div></article>
<article class="failure-card"><video class="lazy-video" controls muted loop playsinline preload="none" poster="/videos/failures/draw-on-balloon.jpg" aria-label="Draw on Balloon failure rollout"><source data-src="/videos/failures/draw-on-balloon.mp4" type="video/mp4"></video><div class="failure-card__copy"><span class="failure-card__type">Failure F2 · Draw on Balloon</span><h3>Missed Contact Onset</h3><p>The balloon slips out of the left hand during drawing. This occurs because the right hand does not recognize that the marker has already made contact and continues its motion, creating a disturbance that exceeds the left hand’s stabilizing grip.</p></div></article>
<article class="failure-card"><video class="lazy-video" controls muted loop playsinline preload="none" poster="/videos/failures/insert-poker.jpg" aria-label="Insert Poker Cards failure rollout"><source data-src="/videos/failures/insert-poker.mp4" type="video/mp4"></video><div class="failure-card__copy"><span class="failure-card__type">Failure F3 · Insert Poker Cards</span><h3>Excessive Pickup Force</h3><p>Due to excessive normal pressure from two right-hand fingers, inter-card friction rises at pickup, causing two cards to be lifted as one stack rather than separating a single layer.</p></div></article>
<article class="failure-card"><video class="lazy-video" controls muted loop playsinline preload="none" poster="/videos/failures/squeeze-toothpaste.jpg" aria-label="Squeeze Toothpaste failure rollout"><source data-src="/videos/failures/squeeze-toothpaste.mp4" type="video/mp4"></video><div class="failure-card__copy"><span class="failure-card__type">Failure F4 · Squeeze Toothpaste</span><h3>Nozzle–Brush Misalignment</h3><p>The toothpaste is dispensed beside the bristles instead of onto them. The reason is inaccurate nozzle–brush alignment: the policy produces the squeezing action without tightly coupling visual placement to haptic evidence of extrusion.</p></div></article>
</div>
</section>
<section id="citation">
<h2>Citation</h2>
<p>If you find our work helpful, please cite it below.</p>
<pre class="bibtex">@misc{ye2026hacolearninghapticactive,
title = {HACo: Learning Haptic Active Compliance for Force-Aware Dexterous Manipulation},
author = {Naisheng Ye and Yinzhe Zhou and Junkai Zhao and Yuhang Lu and Checheng Yu and Zhenjie Yang and Pengwei Wang and Hongyang Li},
year = {2026},
eprint = {2609.36596},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2609.36596}
}</pre>
<p class="citation-credit">This website draws inspiration from the open-source <a href="https://github.com/enpire-research/enpire-research.github.io" target="_blank" rel="noreferrer">ENPIRE project-page template</a> and the <a href="https://tactile-reactive-dexterous.github.io/" target="_blank" rel="noreferrer">T-Rex project page</a>. We thank both teams for sharing their work.</p>
</section>
</article>
</div>
</section>
</main>
<footer class="footer"><span>HACo · PROJECT PAGE</span><a href="#top">Back to top ↑</a></footer>
<script>
const demos = {
poker: ['Insert poker · multi-contact friction', 'Begin with two playing cards lying on the black platform. Use the right hand to pick up one card and transfer it to the left hand. Hold the first card upright and steady with the left hand. Then use the right hand to pick up the remaining card, align it with the card held in the left hand, and slide it into the same grip so that both cards are held together. Finish with both cards securely held in the left hand and the right hand released. Do not bend or drop either card.', ['public/videos/benchmark/poker/2.mp4', 'public/videos/benchmark/poker/3.mp4'], [null, 'Extrapolation·Human Perturbation']],
book: ['Open book · tangential force', 'Begin with the closed book lying flat on the black platform. Use the left hand to separate and lift a thick section of pages near the middle of the book, then turn the section across the spine to open the book to an interior spread. Bring in the right hand to help guide the pages and press the book open and flat. Finish with both hands holding the selected pages flat for at least five seconds. Do not open only the front or back cover, move the book off the platform, or crease or tear any page.', ['public/videos/benchmark/book/1.mp4', 'public/videos/benchmark/book/2.mp4', 'public/videos/benchmark/book/3.mp4', 'public/videos/benchmark/book/4.mp4'], [null, 'Unseen book A', 'Unseen book B', 'Unseen toolkit']],
balloon: ['Draw on balloon · curved fragile contact', 'Begin with the inflated yellow balloon and the black marker on the table. Gently pick up and steady the balloon with the left hand. Use the right hand to pick up the marker and draw a smiley face with two eyes and a curved smile, maintaining light and continuous contact with the balloon. After completing the drawing, set the marker down, lift and orient the balloon so that the finished face is clearly visible to the audience, and present it with the left hand while giving a thumbs-up beside it with the right hand. Do not let the balloon slip, deform excessively, or pop.', ['public/videos/benchmark/balloon/1.mp4', 'public/videos/benchmark/balloon/2.mp4', 'public/videos/benchmark/balloon/3.mp4'], [null, 'Unseen Balloon Color A', 'Unseen Balloon Color B']],
cap: ['Unscrew cap · rotational torque', 'Begin with the bottle standing upright on the white tray. Stabilize the bottle with the left hand while using the right hand to grasp the blue cap and turn it counterclockwise until it is completely detached. At the end, place the bottle upright on the white tray and set the cap down on the table beside the tray, then release both objects and move both hands away. Keep the bottle stable throughout, and do not crush, drop, or spill it.', ['public/videos/benchmark/cap/1.mp4', 'public/videos/benchmark/cap/2.mp4', 'public/videos/benchmark/cap/3.mp4'], [null, 'Unseen bottle color', 'Two consecutive unscrewing tasks·Unseen bottle colors']],
paste: ['Squeeze toothpaste · deformable interaction', 'Begin with the purple toothbrush lying horizontally farther from the robot and the uncapped toothpaste tube positioned below it, closer to the robot. Use the right hand to pick up and hold the toothbrush with its bristles facing the toothpaste tube. Use the left hand to pick up the toothpaste tube, position the nozzle above the toothbrush bristles, and gently squeeze a small amount of toothpaste onto the bristles. Then place the toothpaste tube and toothbrush back on the table, release both objects, and move both hands away. Do not drop either object, damage the toothbrush, or dispense toothpaste anywhere except the bristles.', ['public/videos/benchmark/paste/1.mp4', 'public/videos/benchmark/paste/2.mp4', 'public/videos/benchmark/paste/3.mp4'], [null, null, 'Unseen toothpaste']]
};
const benchmarkVideo = document.getElementById('benchmark-video');
const benchmarkVideoSource = benchmarkVideo.querySelector('source');
const demoVideoPills = document.getElementById('demo-video-pills');
const demoOodDetail = document.getElementById('demo-ood-detail');
const demoOodText = document.getElementById('demo-ood-text');
const rolloutSplit = (task, index) => task === 'paste' ? index < 2 ? 'Standard' : 'OOD' : index === 0 ? 'Standard' : 'OOD';
function selectDemoVideo(task, index) {
const [title, , videos, ood] = demos[task];
benchmarkVideoSource.dataset.src = videos[index];
benchmarkVideo.poster = videos[index].replace(/\.mp4$/, '.jpg');
benchmarkVideo.setAttribute('aria-label', `${title.split(' · ')[0]} rollout ${index + 1}, ${rolloutSplit(task, index)}`);
if (benchmarkVideo.dataset.inView === 'true') {
benchmarkVideoSource.src = videos[index];
benchmarkVideo.load();
benchmarkVideo.play().catch(() => {});
}
demoVideoPills.querySelectorAll('.demo-video-pill').forEach((item, itemIndex) => {
item.classList.toggle('active', itemIndex === index);
item.setAttribute('aria-selected', itemIndex === index ? 'true' : 'false');
});
demoOodDetail.hidden = !ood[index];
demoOodText.textContent = ood[index] || '';
}
function loadBenchmarkVideo() {
benchmarkVideo.dataset.inView = 'true';
benchmarkVideoSource.src = benchmarkVideoSource.dataset.src;
benchmarkVideo.load();
benchmarkVideo.play().catch(() => {});
}
if ('IntersectionObserver' in window) {
const videoObserver = new IntersectionObserver((entries, observer) => {
entries.forEach((entry) => {
if (!entry.isIntersecting) return;
const video = entry.target;
if (video === benchmarkVideo) loadBenchmarkVideo();
else {
const source = video.querySelector('source');
source.src = source.dataset.src;
video.load();
}
observer.unobserve(video);
});
}, { rootMargin: '320px 0px' });
videoObserver.observe(benchmarkVideo);
document.querySelectorAll('video.lazy-video').forEach((video) => videoObserver.observe(video));
} else {
loadBenchmarkVideo();
document.querySelectorAll('video.lazy-video').forEach((video) => {
const source = video.querySelector('source');
source.src = source.dataset.src;
video.load();
});
}
function renderDemoVideos(task) {
const [title, , videos] = demos[task];
demoVideoPills.replaceChildren();
demoVideoPills.setAttribute('aria-label', `${title.split(' · ')[0]} rollout selection`);
videos.forEach((source, index) => {
const split = rolloutSplit(task, index);
const button = document.createElement('button');
button.className = `demo-video-pill${index === 0 ? ' active' : ''}`;
button.type = 'button';
button.dataset.split = split.toLowerCase();
button.setAttribute('aria-label', `Rollout ${index + 1}, ${split}`);
button.setAttribute('aria-selected', index === 0 ? 'true' : 'false');
const number = document.createElement('span');
number.textContent = `#${index + 1}`;
const status = document.createElement('small');
status.textContent = split;
button.append(number, status);
button.addEventListener('click', () => selectDemoVideo(task, index));
demoVideoPills.append(button);
});
selectDemoVideo(task, 0);
}
document.querySelectorAll('.demo-pill').forEach((button) => {
button.addEventListener('click', () => {
document.querySelectorAll('.demo-pill').forEach((item) => item.classList.remove('active'));
button.classList.add('active');
document.querySelectorAll('.benchmark-taxonomy [data-task]').forEach((item) => {
const isActive = item.dataset.task === button.dataset.task;
item.classList.toggle('active', isActive);
if (isActive) item.setAttribute('aria-current', 'true');
else item.removeAttribute('aria-current');
});
const [title, description] = demos[button.dataset.task];
document.getElementById('demo-title').textContent = title;
document.getElementById('demo-description').textContent = description;
renderDemoVideos(button.dataset.task);
});
});
renderDemoVideos('poker');
const resultFigures = document.querySelectorAll('.result-figure');
if (window.matchMedia('(prefers-reduced-motion: reduce)').matches || !('IntersectionObserver' in window)) {
resultFigures.forEach((figure) => figure.classList.add('is-visible'));
} else {
const figureObserver = new IntersectionObserver((entries) => {
entries.forEach((entry) => {
if (entry.isIntersecting) {
entry.target.classList.add('is-visible');
figureObserver.unobserve(entry.target);
}
});
}, { threshold: 0.16, rootMargin: '0px 0px -7% 0px' });
resultFigures.forEach((figure) => figureObserver.observe(figure));
}
const outlineLinks = Array.from(document.querySelectorAll('.outline-link'));
const updateOutline = () => {
let current = outlineLinks[0]?.getAttribute('href')?.slice(1);
outlineLinks.forEach((link) => {
const id = link.getAttribute('href')?.slice(1);
const section = id && document.getElementById(id);
if (section && section.getBoundingClientRect().top <= window.innerHeight * .34) current = id;
});
outlineLinks.forEach((link) => {
const active = link.getAttribute('href') === '#' + current;
link.classList.toggle('is-active', active);
if (active) link.setAttribute('aria-current', 'location');
else link.removeAttribute('aria-current');
});
};
updateOutline();
window.addEventListener('scroll', updateOutline, { passive: true });
window.addEventListener('resize', updateOutline);
</script>
</body>
</html>