Repository navigation
Expand file tree
/
Copy pathindex.html
More file actions
429 lines (407 loc) · 39.7 KB
/
Copy pathindex.html
File metadata and controls
429 lines (407 loc) · 39.7 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
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width,initial-scale=1">
<title>Bench2Dex: A Bimanual Manipulation Benchmark across Dexterous Hands with Simulated Fingertip Visuo-Tactile Support</title>
<meta name="description" content="Bench2Dex is a simulation benchmark for bimanual manipulation across 12 dexterous hands, with 26 tasks, about 1.3K human-teleoperated demonstrations, and a unified visuo-tactile interface for multimodal policy learning.">
<meta property="og:title" content="Bench2Dex: A Bimanual Manipulation Benchmark across Dexterous Hands with Simulated Fingertip Visuo-Tactile Support">
<meta property="og:description" content="Bimanual manipulation across 12 dexterous hands with simulated visuo-tactile support. 26 tasks · ≈1.3K demonstrations.">
<meta property="og:type" content="website">
<meta property="og:url" content="https://bench2dex.github.io/">
<meta property="og:image" content="https://bench2dex.github.io/assets/img/og-cover.png">
<meta property="og:image:type" content="image/png">
<meta property="og:image:width" content="2400">
<meta property="og:image:height" content="1260">
<meta property="og:image:alt" content="Bench2Dex: visuo-tactile bimanual dexterous manipulation benchmark across 12 robot embodiments.">
<meta name="twitter:card" content="summary_large_image">
<link rel="icon" href="assets/img/favicon.png" type="image/png">
<link rel="apple-touch-icon" href="assets/img/apple-touch-icon.png">
<link rel="stylesheet" href="assets/css/style.css?v=score-v1">
<link rel="stylesheet" href="https://bench2dex.github.io/Bench2Dex-Leaderboard/assets/css/leaderboard.css?v=score-v1">
</head>
<body>
<!-- ================= NAV ================= -->
<nav class="nav" id="nav">
<div class="nav-inner">
<a href="#top" class="brand" aria-label="BENCH2DEX home">
<img src="assets/img/figures/overview_logo.png" alt="BENCH2DEX" id="brand-logo" width="1774" height="887">
</a>
<div class="nav-links" id="nav-links">
<a href="#video">Video</a>
<a href="#abstract">Abstract</a>
<a href="#overview">Overview</a>
<a href="#why">Compare</a>
<a href="#embodiments">Dataset</a>
<a href="#tasks">Tasks</a>
<a href="#evaluation">Evaluation</a>
<a href="#results">Leaderboard</a>
<a href="#docs">Docs</a>
</div>
<div class="nav-cta">
<button type="button" class="btn btn-primary" data-copy-bib>Copy BibTeX</button>
<button class="burger" id="burger" aria-label="Open menu" aria-expanded="false"><span></span><span></span><span></span></button>
</div>
</div>
<div class="nav-progress" id="nav-progress" aria-hidden="true"></div>
</nav>
<div class="drawer" id="drawer">
<div class="drawer-head">
<span class="brand-name">BENCH2DEX</span>
<button class="drawer-close" id="drawer-close" aria-label="Close menu">×</button>
</div>
<a href="#video">Video</a>
<a href="#abstract">Abstract</a>
<a href="#overview">Overview</a>
<a href="#why">Compare</a>
<a href="#embodiments">Embodiments</a>
<a href="#data">Teleop & Data</a>
<a href="#tasks">Tasks</a>
<a href="#evaluation">Evaluation</a>
<a href="#generalization">Generalization</a>
<a href="#results">Leaderboard</a>
<a href="#docs">Docs & Get Started</a>
</div>
<main id="top">
<!-- ================= HERO ================= -->
<header class="hero masthead hero-centered">
<div class="hero-atmosphere" aria-hidden="true"><span></span><span></span><span></span></div>
<div class="container">
<div class="hero-content">
<img class="hero-logo" src="assets/img/figures/overview_logo.png" alt="Bench2Dex: two dexterous robot hands linked by visuo-tactile contact" width="1774" height="887" decoding="async" fetchpriority="high">
<h1 class="sr-only">Bench2Dex</h1>
<p class="hero-subtitle">A Bimanual Manipulation Benchmark across Dexterous Hands<br>with Simulated Fingertip Visuo-Tactile Support</p>
<div class="hero-contributors" aria-label="Authors and affiliations">
<p class="hero-author-slot">Zhenjie Yang<sup>3*</sup>, Yideng Zhang<sup>1*</sup>, Dongjie Zhang<sup>2,5*</sup>, Chenyu Jiang<sup>2,5*</sup>, Xianshuai Liu<sup>1</sup>, Yufeng Li<sup>1,5</sup>,<br>
Zuhao Ge<sup>2</sup>, Xingyu Jiao<sup>2,5</sup>, Zheng Zhang<sup>1</sup>, Kaiyu He<sup>1</sup>, Qifeng Li<sup>1</sup>, He Wang<sup>1</sup>, Yuwen Zhong<sup>1</sup>,<br>
Yi Deng<sup>1</sup>, Muyun Jiang<sup>7</sup>, Xianliang Huang<sup>2</sup>, Haisheng Su<sup>1</sup>, Donghang Zhang<sup>4</sup>, Jian Zhang<sup>4</sup>,<br>
Xue Yang<sup>1,6</sup>, Hongyang Li<sup>3</sup>, Xiaosong Jia<sup>2†</sup>, Junchi Yan<sup>1†</sup></p>
<p class="hero-affiliation-slot"><sup>1</sup>Shanghai Jiao Tong University · <sup>2</sup>Fudan University · <sup>3</sup>The University of Hong Kong · <sup>4</sup>Inspire Robots<br>
<sup>5</sup>Zhongguancun Academy · <sup>6</sup>COWARobot Co. Ltd · <sup>7</sup>Nanyang Technological University</p>
<p class="hero-affiliation-slot"><sup>*</sup> Core contribution · <sup>†</sup> Corresponding authors</p>
</div>
<nav class="cta" aria-label="Project resources">
<a class="btn btn-resource" href="https://arxiv.org/abs/2609.15726" target="_blank" rel="noopener"><svg class="btn-mark mark-arxiv" viewBox="0 0 24 24" aria-hidden="true" focusable="false"><path d="M3.8423 0a1.0037 1.0037 0 0 0-.922.6078c-.1536.3687-.0438.6275.2938 1.1113l6.9185 8.3597-1.0223 1.1058a1.0393 1.0393 0 0 0 .003 1.4229l1.2292 1.3135-5.4391 6.4444c-.2803.299-.4538.823-.2971 1.1986a1.0253 1.0253 0 0 0 .9585.635.9133.9133 0 0 0 .6891-.3405l5.783-6.126 7.4902 8.0051a.8527.8527 0 0 0 .6835.2597.9575.9575 0 0 0 .8777-.6138c.1577-.377-.017-.7502-.306-1.1407l-7.0518-8.3418 1.0632-1.13a.9626.9626 0 0 0 .0089-1.3165L4.6336.4639s-.3733-.4535-.768-.463zm0 .272h.0166c.2179.0052.4874.2715.5644.3639l.005.006.0052.0055 10.169 10.9905a.6915.6915 0 0 1-.0072.945l-1.0666 1.133-1.4982-1.7724-8.5994-10.39c-.3286-.472-.352-.6183-.2592-.841a.7307.7307 0 0 1 .6704-.4401Zm14.341 1.5701a.877.877 0 0 0-.6554.2418l-5.6962 6.1584 1.6944 1.8319 5.3089-6.5138c.3251-.4335.479-.6603.3247-1.0292a1.1205 1.1205 0 0 0-.9763-.689zm-7.6557 12.2823 1.3186 1.4135-5.7864 6.1295a.6494.6494 0 0 1-.4959.26.7516.7516 0 0 1-.706-.4669c-.1119-.2682.0359-.6864.2442-.9083l.0051-.0055.0047-.0055z"/></svg>Paper</a>
<a class="btn btn-resource" href="https://bench2dex.github.io/Bench2Dex-Leaderboard/"><svg class="btn-mark" viewBox="0 0 24 24" aria-hidden="true" focusable="false"><path d="M3 12h4v9H3zm7-9h4v18h-4zm7 5h4v13h-4z"/></svg>Leaderboard</a>
<a class="btn btn-resource" href="https://bench2dex.github.io/doc/" target="_blank" rel="noopener"><span class="btn-mark mark-doc" aria-hidden="true"></span>Docs</a>
<a class="btn btn-resource" href="https://github.com/Bench2Dex/Bench2Dex" target="_blank" rel="noopener"><svg class="btn-mark mark-github" viewBox="0 0 24 24" aria-hidden="true" focusable="false"><path d="M12 .297c-6.63 0-12 5.373-12 12 0 5.303 3.438 9.8 8.205 11.385.6.113.82-.258.82-.577 0-.285-.01-1.04-.015-2.04-3.338.724-4.042-1.61-4.042-1.61C4.422 18.07 3.633 17.7 3.633 17.7c-1.087-.744.084-.729.084-.729 1.205.084 1.838 1.236 1.838 1.236 1.07 1.835 2.809 1.305 3.495.998.108-.776.417-1.305.76-1.605-2.665-.3-5.466-1.332-5.466-5.93 0-1.31.465-2.38 1.235-3.22-.135-.303-.54-1.523.105-3.176 0 0 1.005-.322 3.3 1.23.96-.267 1.98-.399 3-.405 1.02.006 2.04.138 3 .405 2.28-1.552 3.285-1.23 3.285-1.23.645 1.653.24 2.873.12 3.176.765.84 1.23 1.91 1.23 3.22 0 4.61-2.805 5.625-5.475 5.92.42.36.81 1.096.81 2.22 0 1.606-.015 2.896-.015 3.286 0 .315.21.69.825.57C20.565 22.092 24 17.592 24 12.297c0-6.627-5.373-12-12-12"/></svg>GitHub</a>
<a class="btn btn-resource btn-huggingface" href="https://huggingface.co/Bench2Dex" target="_blank" rel="noopener"><svg class="btn-mark mark-hf" viewBox="0 0 24 24" aria-hidden="true" focusable="false"><path d="M1.4446 11.5059c0 1.1021.1673 2.1585.4847 3.1563-.0378-.0028-.0691-.0058-.1058-.0058-.4209 0-.8015.16-1.0704.4512-.3454.3737-.4984.8335-.4316 1.293a1.576 1.576 0 0 0 .2148.5978c-.2319.1864-.4018.4456-.4844.7578-.0646.2448-.131.7543.2149 1.2794a1.4552 1.4552 0 0 0-.0625.1055c-.208.3923-.2207.8372-.0371 1.25.2783.6258.9696 1.1175 2.3126 1.6467.8356.3292 1.5988.5411 1.6056.543 1.1046.2847 2.104.4277 2.969.4277 1.4173 0 2.4754-.3849 3.1525-1.1446 1.538.2651 2.791.1403 3.592.006.6773.7555 1.7332 1.1387 3.1467 1.1387.8649 0 1.8643-.143 2.969-.4278.0068-.0019.77-.2138 1.6056-.543 1.343-.5292 2.0343-1.0208 2.3126-1.6466.1836-.4129.171-.8577-.037-1.25a1.4685 1.4685 0 0 0-.0626-.1056c.346-.525.2795-1.0346.2149-1.2793-.0826-.3122-.2525-.5714-.4844-.7579.11-.1816.1831-.3788.2148-.5977.0669-.4595-.0862-.9193-.4316-1.293-.2688-.2913-.6495-.4513-1.0704-.4513-.0209 0-.0376.0008-.0588.0018.3162-.9966.4846-2.0518.4846-3.1523 0-5.807-4.7362-10.5144-10.5789-10.5144-5.8426 0-10.5788 4.7073-10.5788 10.5144Zm10.5788-9.4831c5.2727 0 9.5476 4.246 9.5476 9.483a9.4201 9.4201 0 0 1-.2696 2.2365c-.0039-.0047-.0079-.011-.0117-.0156-.274-.3255-.6679-.5059-1.1075-.5059-.352 0-.714.1155-1.0763.3438-.2403.1517-.5058.422-.7793.7598-.2534-.3492-.608-.5832-1.0137-.6465a1.5174 1.5174 0 0 0-.2344-.0176c-.9263 0-1.4828.7993-1.6935 1.5177-.1046.2426-.6065 1.3482-1.3614 2.0978-1.1681 1.1601-1.4458 2.3534-.8396 3.6382-.843.1029-1.5836.0927-2.365-.006.5906-1.212.3626-2.4388-.8426-3.6322-.755-.7496-1.2568-1.8552-1.3614-2.0978-.2107-.7184-.7673-1.5177-1.6935-1.5177-.078 0-.1568.0054-.2344.0176-.4057.0633-.7604.2973-1.0137.6465-.2735-.3379-.539-.6081-.7794-.7598-.3622-.2283-.7243-.3438-1.0762-.3438-.4266 0-.8094.171-1.0821.4786a9.4208 9.4208 0 0 1-.2598-2.1936c0-5.237 4.2749-9.483 9.5475-9.483zM8.6443 7.0036c-.4838.0043-.9503.2667-1.1934.7227-.3536.6633-.1006 1.4873.5645 1.84.351.1862.4883-.5261.836-.6485.3107-.1095.841.399 1.0078.086.3536-.6634.1025-1.4874-.5625-1.84a1.3659 1.3659 0 0 0-.6524-.1602Zm6.8403 0c-.2199-.002-.4426.05-.6504.1602-.665.3526-.9181 1.1766-.5645 1.84.1669.313.6971-.1955 1.0079-.086.3476.1224.4867.8347.838.6485.6649-.3527.916-1.1767.5624-1.84-.243-.456-.7096-.7184-1.1934-.7227Zm-9.7565 1.418a.8768.8768 0 0 0-.877.877c0 .4846.3925.877.877.877a.8768.8768 0 0 0 .877-.877.8768.8768 0 0 0-.877-.877zm12.6434 0c-.4845 0-.879.3925-.879.877 0 .4846.3945.877.879.877a.8768.8768 0 0 0 .877-.877.8768.8768 0 0 0-.877-.877zM8.7927 11.459c-.179-.003-.2793.1107-.2793.416 0 .8097.3874 2.125 1.4279 2.924.207-.7123 1.3453-1.2832 1.5079-1.2012.2315.1167.2191.4417.6074.7266.3884-.285.374-.6098.6056-.7266.1627-.082 1.3009.4889 1.5079 1.2012 1.0404-.799 1.4278-2.1144 1.4278-2.924 0-1.2212-1.583.6402-3.5413.6485-1.4686-.0061-2.7266-1.0558-3.2639-1.0645zM4.312 14.4768c.5792.365 1.6964 2.2751 2.1056 3.0177.1371.2488.371.3536.582.3536.4188 0 .7465-.4138.0391-.9395-1.0636-.791-.6914-2.0846-.1836-2.1642a.4302.4302 0 0 1 .0664-.004c.4616 0 .666.7892.666.7892s.5959 1.4898 1.6213 2.508c.942.9356 1.062 1.703.4961 2.6661-.0164-.004-.0159.0236-.1484.2149-.1853.2673-.4322.4688-.7188.6152-.5062.2269-1.1397.2696-1.7833.2696-1.037 0-2.1017-.1824-2.6975-.336-.0293-.0075-3.6505-.9567-3.1916-1.8224.0771-.1454.2033-.2031.3633-.2031.6463 0 1.823.9551 2.3283.9551.113 0 .196-.0865.2285-.2031.2249-.8045-3.2787-1.0522-2.9846-2.1642.0519-.1967.193-.2757.3907-.2754.854 0 2.7704 1.4923 3.172 1.4923.0307 0 .0525-.0085.0645-.0274.2012-.3227.1096-.5865-1.3087-1.4395-1.4182-.8533-2.4315-1.329-1.8653-1.9416.0651-.0707.1574-.1015.2695-.1015.8611.0002 2.8948 1.84 2.8948 1.84s.5487.5683.8809.5683c.0762 0 .1416-.0315.1855-.1054.2355-.3946-2.1858-2.2183-2.3224-2.971-.0926-.51.0641-.7676.3555-.7676-.0006.008.1701-.0285.4942.1759zm16.2257.5918c-.1366.7526-2.5579 2.5764-2.3224 2.9709.044.074.1092.1055.1855.1055.3321 0 .881-.5684.881-.5684s2.0336-1.8397 2.8947-1.84c.1121 0 .2044.0308.2695.1016.5662.6125-.447 1.0882-1.8653 1.9415-1.4183.853-1.51 1.1168-1.3087 1.4396.012.0188.0337.0273.0644.0273.4016 0 2.3181-1.4923 3.1721-1.4923.1977-.0002.3388.0787.3907.2754.294 1.112-3.2095 1.3597-2.9846 2.1642.0325.1166.1156.2032.2285.2032.5054 0 1.682-.9552 2.3283-.9552.16 0 .2862.0577.3633.2032.459.8656-3.1623 1.8149-3.1916 1.8224-.5958.1535-1.6605.336-2.6975.336-.6351 0-1.261-.0409-1.7638-.2599-.2949-.1472-.5488-.3516-.7383-.625-.0411-.0682-.1026-.1476-.1426-.205-.5726-.9679-.455-1.7371.4903-2.676 1.0254-1.0182 1.6212-2.508 1.6212-2.508s.2044-.7891.666-.7891a.4318.4318 0 0 1 .0665.0039c.5078.0796.88 1.3732-.1836 2.1642-.7074.5257-.3797.9395.039.9395.211 0 .445-.1047.5821-.3535.4092-.7426 1.5264-2.6527 2.1056-3.0178.5588-.3524.99-.1816.8497.5918z"/></svg>Hugging Face</a>
<a class="btn btn-resource btn-modelscope" href="https://modelscope.cn/organization/Bench2Dex" target="_blank" rel="noopener"><svg class="btn-mark mark-ms" viewBox="0 0 24 24" aria-hidden="true" focusable="false"><path d="M2.667 5.333V8H0v5.333h2.667v-2.666H.5V8.5h2.166v2.166h2.666V8H8V5.333Zm0 8v5.334H8V16H5.333v-2.667Zm13.333-8V8h2.667v2.667h2.666V8.5H23.5v2.166h-2.166v2.666H24V8h-2.667V5.333Zm5.333 8h-2.666V16H16v2.667h5.333zM8 10.667v2.666h2.667v-2.666zm2.667 2.666V16h2.666v-2.667zm2.666 0H16v-2.666h-2.667z"/></svg>ModelScope</a>
<a class="btn btn-primary" href="#tasks">Explore Tasks <span class="btn-arrow" aria-hidden="true">→</span></a>
</nav>
</div>
</div>
<section class="proof hero-proof" aria-label="Bench2Dex at a glance">
<div class="proof-grid" id="proof-grid"></div>
</section>
</header>
<!-- ================= VIDEO ================= -->
<section class="section video-section" id="video">
<div class="container hero-banner">
<div class="section-head">
<p class="eyebrow">Video</p>
<h2>A short project showcase of Bench2Dex.</h2>
</div>
<figure class="hero-evidence">
<div class="hero-evidence-frame">
<video class="hero-video" src="assets/video/demo_v2.mp4" poster="assets/img/frames/demo_v2_poster.png" autoplay muted loop playsinline preload="metadata" controls aria-label="Bench2Dex project demo"></video>
</div>
</figure>
</div>
</section>
<!-- ================= ABSTRACT ================= -->
<section class="section evidence-band evidence-band--paper" id="abstract" data-section-index="01">
<div class="container">
<div class="section-head">
<p class="eyebrow">Abstract</p>
<h2>A bimanual manipulation benchmark with simulated visuo-tactile support.</h2>
</div>
<p class="abstract-text">Bimanual dexterous manipulation requires coordinated control of two hands in multi-stage, contact-rich tasks. We introduce <b>Bench2Dex</b>, a simulation benchmark for bimanual manipulation supporting 12 dexterous hands. The benchmark includes 26 tasks involving tool use, interaction with articulated objects, and multi-stage coordination, together with approximately 1.3K human-teleoperated demonstrations. It integrates demonstration collection, synchronized multimodal data, and automated metrics for task completion and intermediate progress within a common framework. We equip every supported hand model with simulated vision-based tactile sensing at the fingertips, encoding local contact geometry as image-like tactile observations. Through a <b>unified visuo-tactile interface</b>, we make these observations available alongside external vision and proprioception to support research on tactile representations and multimodal policy learning. We establish reference results using ACT, Diffusion Policy, π₀.₅, and GR00T N1.5 with visual and proprioceptive inputs. We assess robustness across four evaluation channels—<b>None</b>, <b>Equivariance-only</b>, <b>Invariance-only</b>, and <b>Full</b>—covering an unperturbed baseline and controlled combinations of seven scene perturbation types. By combining a bimanual manipulation benchmark across dexterous hands with simulated visuo-tactile support and automated policy evaluation, Bench2Dex provides the community with a common starting point for investigating when tactile information is useful and how manipulation policies can effectively incorporate it.</p>
</div>
</section>
<!-- ================= SYSTEM OVERVIEW (paper main figure) ================= -->
<section class="section overview-fig evidence-band evidence-band--figure" id="overview" data-section-index="02">
<div class="container">
<div class="section-head">
<p class="eyebrow">System Overview</p>
<h2>One pipeline: teleoperation → multimodal acquisition → shared visuo-tactile representation → evaluation.</h2>
</div>
<figure class="overview-frame">
<img src="assets/img/figures/overview.png" alt="BENCH2DEX system overview: human teleoperation drives 12 bimanual dexterous embodiments through Isaac Lab, recording 8 synchronized modalities including a shared visuo-tactile representation, with controlled generalization and executable evaluation." fetchpriority="high">
<figcaption class="evidence-caption">BENCH2DEX connects human teleoperation, multimodal data acquisition (RGB, depth, joint/object state, a shared visuo-tactile representation, 2D / 3D boxes, occupancy grids), and executable evaluation under controlled generalization — across 26 long-horizon tasks and 12 bimanual dexterous embodiments.</figcaption>
</figure>
</div>
</section>
<!-- ================= WHY ================= -->
<section class="section evidence-band evidence-band--data" id="why" data-section-index="03">
<div class="container">
<div class="section-head">
<p class="eyebrow">Comparison</p>
<h2>Comparison with representative manipulation benchmarks.</h2>
<p class="lead">Most manipulation benchmarks and datasets use a limited set of robot platforms, one gripper or hand design, or a narrow set of sensing modalities. Even large datasets spanning multiple embodiments lack evaluation protocols designed to compare dexterous manipulation policies across different arm–hand designs and contact surfaces. Bench2Dex connects demonstration collection, policy learning, and evaluation within a common framework.</p>
</div>
<div class="card-frame">
<div class="bm-table-wrap">
<table class="bm-table" aria-label="Comparison of BENCH2DEX with representative manipulation benchmarks" aria-describedby="dexverse-note">
<thead><tr>
<th class="bm-name">Benchmark</th>
<th>Task Num</th>
<th>Embodiment</th>
<th>Tool/Device Use</th>
<th>Dexterous Hand</th>
<th>Teleoperation</th>
<th>Vision-Based Tactile</th>
<th>Articulated</th>
</tr></thead>
<tbody>
<tr><td>LIBERO</td><td>130</td><td>1</td><td><span class="nx">✗</span></td><td><span class="nx">✗</span></td><td><span class="yk">✓</span></td><td><span class="nx">✗</span></td><td><span class="yk">✓</span></td></tr>
<tr><td>RLBench2</td><td>13</td><td>1</td><td><span class="yk">✓</span></td><td><span class="nx">✗</span></td><td><span class="nx">✗</span></td><td><span class="nx">✗</span></td><td><span class="yk">✓</span></td></tr>
<tr><td>RoboCasa</td><td>100</td><td>1</td><td><span class="yk">✓</span></td><td><span class="nx">✗</span></td><td><span class="yk">✓</span></td><td><span class="nx">✗</span></td><td><span class="yk">✓</span></td></tr>
<tr><td>DROID</td><td>86</td><td>1</td><td><span class="nx">✗</span></td><td><span class="nx">✗</span></td><td><span class="yk">✓</span></td><td><span class="nx">✗</span></td><td><span class="nx">✗</span></td></tr>
<tr><td>DexMimicGen</td><td>9</td><td>3</td><td><span class="nx">✗</span></td><td><span class="yk">✓</span></td><td><span class="yk">✓</span></td><td><span class="nx">✗</span></td><td><span class="yk">✓</span></td></tr>
<tr><td>RealMirror</td><td>5</td><td>1</td><td><span class="yk">✓</span></td><td><span class="yk">✓</span></td><td><span class="yk">✓</span></td><td><span class="nx">✗</span></td><td><span class="yk">✓</span></td></tr>
<tr><td>RoboTwin 2.0</td><td>50</td><td>5</td><td><span class="yk">✓</span></td><td><span class="nx">✗</span></td><td><span class="nx">✗</span></td><td><span class="nx">✗</span></td><td><span class="yk">✓</span></td></tr>
<tr><td>RoboMIND 2.0</td><td>739</td><td>6</td><td><span class="yk">✓</span></td><td><span class="yk">✓</span></td><td><span class="yk">✓</span></td><td><span class="nx">✗</span></td><td><span class="yk">✓</span></td></tr>
<tr><td>MuJoCo Manipulus</td><td>16</td><td>1</td><td><span class="yk">✓</span></td><td><span class="nx">✗</span></td><td><span class="nx">✗</span></td><td><span class="nx">✗</span></td><td><span class="nx">✗</span></td></tr>
<tr><td>RoboCasa365</td><td>365</td><td>1</td><td><span class="yk">✓</span></td><td><span class="nx">✗</span></td><td><span class="yk">✓</span></td><td><span class="nx">✗</span></td><td><span class="yk">✓</span></td></tr>
<tr><td>RoboDojo</td><td>60</td><td>3</td><td><span class="yk">✓</span></td><td><span class="nx">✗</span></td><td><span class="yk">✓</span></td><td><span class="nx">✗</span></td><td><span class="yk">✓</span></td></tr>
<tr><td>BiCoord</td><td>18</td><td>1</td><td><span class="nx">✗</span></td><td><span class="nx">✗</span></td><td><span class="nx">✗</span></td><td><span class="nx">✗</span></td><td><span class="nx">✗</span></td></tr>
<tr><td>DexJoCo</td><td>11</td><td>2</td><td><span class="yk">✓</span></td><td><span class="yk">✓</span></td><td><span class="yk">✓</span></td><td><span class="nx">✗</span></td><td><span class="yk">✓</span></td></tr>
<tr><td>DexVerse<sup>∗</sup></td><td>19</td><td>1</td><td><span class="yk">✓</span></td><td><span class="yk">✓</span></td><td><span class="yk">✓</span></td><td><span class="nx">✗</span></td><td><span class="yk">✓</span></td></tr>
<tr class="us">
<td><b>Bench2Dex (Ours)</b></td><td><b>26</b></td><td><b>12</b></td>
<td><span class="yk">✓</span></td><td><span class="yk">✓</span></td><td><span class="yk">✓</span></td><td><span class="yk">✓</span></td><td><span class="yk">✓</span></td>
</tr>
</tbody>
</table>
</div>
<p id="dexverse-note" class="evidence-caption"><em>Note.</em> <sup>∗</sup> Statistics based on publicly available data.</p>
</div>
</div>
</section>
<!-- ================= EMBODIMENTS ================= -->
<section class="section evidence-band evidence-band--atlas" id="embodiments" data-section-index="04">
<div class="container">
<div class="section-head">
<p class="eyebrow">Embodiment Atlas · 12 Arm–Hand Morphologies</p>
<h2>One benchmark, twelve arm–hand morphologies.</h2>
<p class="lead">Bench2Dex supports 12 dexterous hand models with a common teleoperation framework and a unified visuo-tactile interface. The interface provides image-like observations of local contact geometry while preserving hand-specific sensing geometry across different arm–hand designs.</p>
</div>
<div class="atlas-grid" id="atlas-grid"></div>
</div>
</section>
<!-- ================= TELEOP + DATA ================= -->
<section class="section evidence-band evidence-band--split" id="data" data-section-index="05">
<div class="container">
<div class="section-head">
<p class="eyebrow">Teleoperation Pipeline</p>
<h2>From human hands to simulated bimanual dexterity.</h2>
<p class="lead">Manus gloves and ARKit wrist tracking control all 12 embodiments through a shared retargeting interface.</p>
</div>
<div class="teleop-split">
<figure class="ph evidence-figure">
<img src="assets/img/figures/teleop_detail.png" alt="Teleoperation pipeline schematic: Manus glove + ARKit wrist to simulated bimanual robot">
<figcaption class="evidence-caption">A shared capture and retargeting path maps human wrist and hand motion onto every arm–hand embodiment.</figcaption>
</figure>
<div>
<div class="teleop-steps">
<div class="teleop-step"><div><h4>Capture Wrist and Hand Motion</h4><p>A Manus glove provides a 25-node hand skeleton via shared memory; an ARKit stream provides wrist translation and orientation.</p></div></div>
<div class="teleop-step"><div><h4>Retarget Skeleton and Solve Arm IK</h4><p>The skeleton is converted to 21 MediaPipe-style keypoints; DexPilot retargeting solves hand joints, a Pinocchio IK controller solves the arm.</p></div></div>
<div class="teleop-step"><div><h4>Record Actions, States, Observations, Traces</h4><p>Per step, the commanded action is stored, then post-step observations and evaluator state are written to a time-aligned HDF5 episode.</p></div></div>
</div>
</div>
</div>
</div>
</section>
<!-- ================= EIGHT MODALITIES ================= -->
<section class="section evidence-band evidence-band--media" id="modalities" data-section-index="06">
<div class="container">
<div class="section-head">
<p class="eyebrow">Eight Synchronized Data Modalities</p>
<h2>Eight time-aligned views of interaction.</h2>
<p class="lead">Explore eight synchronized sensor and state modalities from one recorded episode. Play the video or drag the timeline to inspect individual frames.</p>
</div>
<div class="modalplayer" data-modalplayer data-manifest="assets/video/modal/manifest.json">
<div class="modal-grid"></div>
<div class="mp-ctrl">
<button class="mp-play" type="button" aria-label="Play synchronized modalities">▶</button>
<input type="range" class="mp-seek" min="0" max="1000" value="0" aria-label="Episode timeline scrubber">
<span class="mp-frame">0:00 / 0:00</span>
<span class="mp-ep">Jigsaw Puzzle Assembly · IIWA7 + Sharpa</span>
</div>
</div>
</div>
</section>
<!-- ================= TACTILE INTERFACE ================= -->
<section class="section evidence-band evidence-band--split" id="tactile" data-section-index="07">
<div class="container">
<div class="section-head">
<p class="eyebrow">Unified Visuo-Tactile Interface</p>
<h2>Different hands, one unified visuo-tactile interface.</h2>
<p class="lead">Bench2Dex equips every supported hand model with simulated vision-based tactile sensing at the fingertips. The unified visuo-tactile interface provides image-like observations of local contact geometry alongside external vision and proprioception, while retaining hand-specific sensing geometry. This supports research on when tactile information is useful and how policies can effectively incorporate it.</p>
</div>
<div class="tactile-wrap">
<div>
<div class="tactile-switch" id="tactile-switch"></div>
</div>
<div>
<div class="tactile-img"><img id="tactile-img" src="assets/img/tactile/ur5_shadow_hand_flange_43_000002.png" alt="tactile image sample"></div>
<p class="dim evidence-caption">An 8-bit tactile image <span class="mono">T<sub>s,t</sub> ∈ {0,…,255}</span> is produced per tactile site per frame: ray-cast contact distance is piecewise-quantized (0.005 mm/level below 0.5 mm, 0.03 mm/level above, saturating near 5.15 mm → 255) for fine near-contact sensitivity, then Gaussian-smoothed.</p>
</div>
</div>
<div class="tactile-steps">
<span class="tactile-step">Hand mesh</span><span class="tactile-step"><span class="arrow">→</span> contact surface</span><span class="tactile-step"><span class="arrow">→</span> ray samples</span><span class="tactile-step"><span class="arrow">→</span> depth response</span><span class="tactile-step"><span class="arrow">→</span> 8-bit visuo-tactile image</span>
</div>
</div>
</section>
<!-- ================= TASKS ================= -->
<section class="section evidence-band evidence-band--atlas" id="tasks" data-section-index="08">
<div class="container">
<div class="section-head">
<p class="eyebrow">Task Suite · 26 Long-Horizon Tasks</p>
<h2>The long-horizon task suite.</h2>
<p class="lead">Each task has multiple stages and a machine-checkable success condition.</p>
</div>
<div class="task-rail" id="task-filters" role="tablist" aria-label="Task filter"></div>
<div class="task-grid" id="all-tasks"></div>
</div>
</section>
<!-- ================= EVALUATION ================= -->
<section class="section evidence-band evidence-band--protocol" id="evaluation" data-section-index="09">
<div class="container">
<div class="section-head">
<p class="eyebrow">Executable Evaluation</p>
<h2>From “success” to an executable protocol.</h2>
<p class="lead">A task is successful only when its executable terminal predicate remains satisfied for the required dwell time. The evaluation suite measures task completion, stage-level progress, and success-conditioned efficiency under a fixed protocol, step budget, dwell time, random seed, and sampled generalization parameters.</p>
</div>
<div class="eval-tiers">
<div class="eval-tier"><div class="top-bar" style="background:var(--green);"></div><div class="num">01</div><div class="n">Stable Success Rate (SR)</div><div class="d">Fraction of episodes in which the terminal predicate holds for the configured dwell time (0.5 s default).</div></div>
<div class="eval-tier"><div class="top-bar" style="background:var(--cyan);"></div><div class="num">02</div><div class="n">Latched Stage Completion Rate (LSCR)</div><div class="d">LSCR counts ever-reached, dependency-valid milestones — progress that survives intentional reversals.</div></div>
<div class="eval-tier"><div class="top-bar" style="background:var(--amber);"></div><div class="num">03</div><div class="n">Success-Conditioned Efficiency</div><div class="d">Mean human expert teleoperation steps to completion divided by pooled mean steps to stable success among successful policy episodes.</div></div>
</div>
<div class="card-frame">
<div class="eval-stages" id="eval-stages">
<div class="eval-stage" data-i="0"><div class="check">✓</div><div class="nm">Stage 1</div><div class="st">Open Microwave Door</div></div>
<div class="eval-stage" data-i="1"><div class="check">✓</div><div class="nm">Stage 2</div><div class="st">Put Baguette in Bowl</div></div>
<div class="eval-stage" data-i="2"><div class="check">✓</div><div class="nm">Stage 3</div><div class="st">Put Bowl in Microwave</div></div>
<div class="eval-stage" data-i="3"><div class="check">✓</div><div class="nm">Stage 4</div><div class="st">Close Microwave Door</div></div>
</div>
<p class="evidence-caption">Example: Microwave Bowl Loading is evaluated through four executable stages; the dwell latch is what makes “closed” a success rather than a transient.</p>
</div>
</div>
</section>
<!-- ================= GENERALIZATION ================= -->
<section class="section evidence-band evidence-band--data" id="generalization" data-section-index="10">
<div class="container">
<div class="section-head">
<p class="eyebrow">Controlled Generalization</p>
<h2>Four channels, seven controlled factors.</h2>
<p class="lead">Each evaluation keeps the semantic goal, object set, and success conditions unchanged while controlling two groups of scene factors. Invariance factors alter task-irrelevant visual conditions without changing the intended task behavior; equivariance factors alter task-relevant geometry and require corresponding changes in reaching, grasping, and contact trajectories.</p>
</div>
<div class="channel-grid" aria-label="Four evaluation channels">
<div class="channel-card ch-none"><span class="ch-bar"></span><span class="ch-name">None</span><span class="ch-desc">Matched anchor; all factors are replayed exactly.</span></div>
<div class="channel-card ch-equi"><span class="ch-bar"></span><span class="ch-name">Equivariance-only</span><span class="ch-desc">Matched anchor; only equivariance factors are resampled while the anchor's invariant context is preserved.</span></div>
<div class="channel-card ch-inv"><span class="ch-bar"></span><span class="ch-name">Invariance-only</span><span class="ch-desc">Matched anchor; only invariance factors are resampled while the anchor's task geometry is preserved.</span></div>
<div class="channel-card ch-full"><span class="ch-bar"></span><span class="ch-name">Full</span><span class="ch-desc">Independent full-scene sample; no anchor is used. Both factor groups are resampled.</span></div>
</div>
<p class="data-note">None / Equi. / Inv. use matched anchors aligned by episode index; Full is sampled independently without an anchor.</p>
<div class="gen-axes">
<div class="gen-group">
<div class="group-head inv">Invariance Factors <span class="muted">— task-irrelevant visual conditions; resampled by the Inv. channel</span></div>
<div class="axis-grid">
<div class="axis-card"><video class="axis-video" src="assets/video/7_generalizations/background_smooth.mp4#t=0.1" muted loop playsinline preload="metadata" aria-label="Scene background generalization"></video><div class="axis-meta"><span class="axis-name">Scene Background</span><span class="axis-note">Background context used for visual rendering only.</span></div></div>
<div class="axis-card"><video class="axis-video" src="assets/video/7_generalizations/table_surface_smooth.mp4#t=0.1" muted loop playsinline preload="metadata" aria-label="Tabletop texture generalization"></video><div class="axis-meta"><span class="axis-name">Tabletop Texture</span><span class="axis-note">Support-surface visual material.</span></div></div>
<div class="axis-card"><video class="axis-video" src="assets/video/7_generalizations/light_smooth.mp4#t=0.1" muted loop playsinline preload="metadata" aria-label="Lighting conditions generalization"></video><div class="axis-meta"><span class="axis-name">Lighting Conditions</span><span class="axis-note">Lighting embedded in the sampled scene.</span></div></div>
<div class="axis-card"><video class="axis-video" src="assets/video/7_generalizations/clutter_smooth.mp4#t=0.1" muted loop playsinline preload="metadata" aria-label="Distractor objects generalization"></video><div class="axis-meta"><span class="axis-name">Distractor Objects</span><span class="axis-note">Distractor objects on the tabletop.</span></div></div>
<div class="axis-card"><video class="axis-video" src="assets/video/7_generalizations/camera_generalization.mp4#t=0.1" muted loop playsinline preload="metadata" aria-label="Camera pose generalization"></video><div class="axis-meta"><span class="axis-name">Camera Pose</span><span class="axis-note">World-mounted and wrist-mounted camera perturbations.</span></div></div>
</div>
</div>
<div class="gen-group">
<div class="group-head equi">Equivariance Factors <span class="muted">— task-relevant geometry; resampled by the Equi. channel</span></div>
<div class="axis-grid two">
<div class="axis-card"><video class="axis-video" src="assets/video/7_generalizations/object_pose_generalization.mp4#t=0.1" muted loop playsinline preload="metadata" aria-label="Object pose generalization"></video><div class="axis-meta"><span class="axis-name">Object Pose</span><span class="axis-note">Task-relevant initial object poses within valid bounds.</span></div></div>
<div class="axis-card"><video class="axis-video" src="assets/video/7_generalizations/table_height_generalization.mp4#t=0.1" muted loop playsinline preload="metadata" aria-label="Table height generalization"></video><div class="axis-meta"><span class="axis-name">Table Height</span><span class="axis-note">Worksurface height resampling.</span></div></div>
</div>
</div>
</div>
</div>
</section>
<!-- ================= RESULTS ================= -->
<section class="section evidence-band evidence-band--results" id="results" data-section-index="11">
<div class="container">
<div class="section-head">
<p class="eyebrow">Results</p>
<h2>Policy Leaderboard</h2>
<p class="lead">Reported policy evaluations across 26 tasks.</p>
</div>
<div class="paper-summary board-section">
<div class="section-toolbar">
<div class="section-intro"><div class="section-heading"><h2>Overall Performance</h2><span class="section-count" id="home-ranking-count"></span></div><p id="home-ranking-label" class="table-sort-status" aria-live="polite"></p></div>
<div class="section-meta"><p class="table-updated" id="home-ranking-updated"></p><div class="section-actions"><a class="btn btn-primary" href="https://bench2dex.github.io/Bench2Dex-Leaderboard/">Explore Leaderboard →</a></div></div>
</div>
<div class="explorer-table-wrap" tabindex="0" role="region" aria-label="Policy Leaderboard"><table class="policy-table ranking-table" id="lb-table"></table></div>
<div class="table-footer"><p class="table-note" id="home-ranking-caption"></p><div class="table-legend-compact"><strong class="best-value" title="Highest reported SR, LSCR or efficiency, including ties">Best</strong><span class="heat-key" aria-label="Color scale: 0 to 100 percent"><span>0</span><span class="heat-swatch" aria-hidden="true"></span><span>100%</span></span></div></div>
</div>
</div>
</section>
<!-- ================= GET STARTED / DOCS ================= -->
<section class="section evidence-band evidence-band--resources" id="docs" data-section-index="12">
<div class="container">
<div class="section-head">
<p class="eyebrow">Get Started</p>
<h2>From a scene to an evaluation.</h2>
<p class="lead">Environment setup, policy training and evaluation, and the full 26-task catalog are in the Bench2Dex Documentation.</p>
</div>
<div class="resource-grid" id="resource-grid"></div>
<div style="margin-top:30px;">
<div class="cite-box">
<pre id="bibtex" tabindex="0"></pre>
<div class="actions">
<button type="button" class="btn" id="copy-bib" data-copy-bib>Copy BibTeX</button>
</div>
</div>
</div>
</div>
</section>
</main>
<!-- ================= FOOTER ================= -->
<footer class="footer" id="contact">
<div class="container">
<div class="footer-grid">
<div>
<div class="brand" style="margin-bottom:14px;">
<img src="assets/img/figures/overview_logo.png" alt="BENCH2DEX" width="1774" height="887">
</div>
<p class="dim" style="max-width:42ch; font-size:14px;">A teleoperated Isaac Lab benchmark for long-horizon bimanual dexterous manipulation. Built on NVIDIA Isaac Sim / Isaac Lab.</p>
</div>
<div>
<h5>Explore</h5>
<a href="#tasks">Tasks</a>
<a href="#embodiments">Embodiments</a>
<a href="#data">Teleop & Data</a>
<a href="#generalization">Generalization</a>
<a href="https://bench2dex.github.io/Bench2Dex-Leaderboard/">Leaderboard</a>
</div>
<div>
<h5>Reproduce</h5>
<a href="#docs">Documentation</a>
<a href="#evaluation">Evaluation</a>
</div>
<div>
<h5>Cite</h5>
<a href="#docs">BibTeX →</a>
</div>
</div>
<div class="footer-bottom">
<span class="ver" id="version-badge">Bench2Dex homepage</span>
<span>Dataset & asset licenses listed in the release; uses NVIDIA Isaac Lab / Isaac Sim. © 2026 BENCH2DEX authors.</span>
</div>
</div>
</footer>
<div class="copy-toast" id="copy-toast" role="status" aria-live="polite">BibTeX copied</div>
<script src="assets/js/data.js"></script>
<script type="application/json" id="modal-manifest">{"fps": 20, "duration_s": 41.65, "frame_count": 833, "video_dir": "assets/video/modal", "asset_revision": "20260822-semantic-depth-v3", "modalities": ["rgb", "depth", "joint", "object", "tactile", "box2d", "box3d", "occ"], "labels": {"rgb": "RGB (Overhead)", "depth": "Depth", "joint": "Joint States", "object": "Object States", "tactile": "Tactile Signals", "box2d": "2D Boxes", "box3d": "3D Boxes", "occ": "Occupancy Grid", "occ_table": "Occupancy Grid + Table"}}</script>
<script type="module" src="assets/js/leaderboard-loader.js?v=score-v1"></script>
<script src="assets/js/citation.js"></script>
</body>
</html>