-
Notifications
You must be signed in to change notification settings - Fork 1
Expand file tree
/
Copy pathindex.html
More file actions
854 lines (784 loc) · 51 KB
/
Copy pathindex.html
File metadata and controls
854 lines (784 loc) · 51 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>See like a Robot: Robot-Centric Pointmaps for Vision-Language-Action Models</title>
<meta name="description" content="Robot-centric pointmaps express per-pixel 3D geometry in the robot frame while preserving the dense image grid used by pretrained VLAs.">
<!-- Open Graph (for link previews on Twitter/Slack etc.) -->
<meta property="og:title" content="See like a Robot: Robot-Centric Pointmaps for VLA Models">
<meta property="og:description" content="Robot-centric pointmaps express per-pixel 3D geometry in the robot frame while preserving the dense image grid used by pretrained VLAs.">
<meta property="og:image" content="static/images/fig1_teaser.png">
<link rel="preconnect" href="https://fonts.googleapis.com">
<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>
<link href="https://fonts.googleapis.com/css2?family=Google+Sans:wght@400;500;700&family=Noto+Sans:ital,wght@0,400;0,600;0,700;1,400&display=swap" rel="stylesheet">
<!-- Google tag (gtag.js) -->
<script async src="https://www.googletagmanager.com/gtag/js?id=G-94PPY030WH"></script>
<script>
window.dataLayer = window.dataLayer || [];
function gtag(){dataLayer.push(arguments);}
gtag('js', new Date());
gtag('config', 'G-94PPY030WH');
</script>
<style>
:root {
--text: #1f2430;
--muted: #5b6270;
--accent: #2563eb; /* link blue */
--accent-dark: #1e40af;
--ours-bg: #eaf3ff; /* shaded "ours" rows */
--border: #e3e6ec;
--card-bg: #f7f8fa;
}
* { margin: 0; padding: 0; box-sizing: border-box; }
html { scroll-behavior: smooth; }
body {
font-family: "Noto Sans", -apple-system, BlinkMacSystemFont, "Segoe UI", Helvetica, Arial, sans-serif;
color: var(--text);
line-height: 1.78;
font-size: 18px;
background: #ffffff;
}
.container { max-width: 1040px; margin: 0 auto; padding: 0 20px; }
h1, h2, h3, .sec-label, .sec-claim, .sub-head, .authors, .affil {
font-family: "Google Sans", "Noto Sans", sans-serif;
}
section.container { margin: 56px auto; }
/* ---------- Hero ---------- */
.hero { text-align: center; padding-top: 56px; }
h1.title {
font-size: clamp(26px, 4.2vw, 40px);
font-weight: 700; line-height: 1.25; margin-bottom: 22px;
}
h1.title .subtitle-break { display: block; }
.authors { font-size: 18px; margin-bottom: 6px; }
.authors a { color: var(--accent); text-decoration: none; }
.authors a:hover { text-decoration: underline; }
.authors sup { font-size: 12px; color: var(--muted); }
.affil { font-size: 16px; color: var(--muted); margin-bottom: 4px; }
.eqnote { font-size: 13.5px; color: var(--muted); margin-bottom: 26px; }
.btn-row { display: flex; flex-wrap: wrap; gap: 10px; justify-content: center; margin-bottom: 8px; }
.btn {
display: inline-flex; align-items: center; gap: 8px;
background: #222831; color: #fff; text-decoration: none;
font-size: 15.5px; font-weight: 600;
border-radius: 999px; padding: 9px 20px;
transition: background .15s ease, transform .15s ease;
}
.btn:hover { background: #000; transform: translateY(-1px); }
.btn.disabled { background: #9aa0ab; cursor: default; pointer-events: none; }
.btn svg { width: 16px; height: 16px; fill: currentColor; }
/* ---------- Figures ---------- */
figure { margin: 28px 0; text-align: center; }
figure img {
max-width: 100%; height: auto; border-radius: 6px;
}
figure.framed img { border: 1px solid var(--border); padding: 10px; background: #fff; }
figcaption {
margin: 12px auto 0; font-size: 14.5px; color: var(--muted); line-height: 1.55;
text-align: left;
}
figcaption b { color: var(--text); }
.tagline {
text-align: center; font-size: 19px; max-width: 780px; margin: 26px auto 30px; color: var(--text);
}
.tagline b { color: var(--accent-dark); }
/* problem-evidence row (intro, CamVLA-style claim + numbers) */
.evidence-row { display: grid; grid-template-columns: 1.2fr 1fr; gap: 30px; align-items: center; margin: 30px auto 0; }
@media (max-width: 760px) { .evidence-row { grid-template-columns: 1fr; } }
.stat-line { padding: 14px 18px; border: 1px solid var(--border); border-radius: 10px; background: var(--card-bg); }
.stat-line + .stat-line { margin-top: 14px; }
.stat-num { font-size: 25px; font-weight: 700; line-height: 1.2; }
.stat-num.bad { color: #111; }
.stat-num.good { color: #229e39; } /* matches "Robot frame" green in the concept figure */
.stat-num.ours { color: #2563eb; }
.stat-desc { font-size: 14px; color: var(--muted); line-height: 1.5; margin-top: 3px; }
.rootcause { text-align: center; font-size: 15.5px; max-width: 780px; margin: 24px auto 0; }
ul.req-list { margin: 0 auto 16px; padding-left: 26px; }
ul.req-list li { margin: 7px 0; }
/* ---------- Sections ---------- */
.sec-label {
margin: 0 auto 16px; font-size: 30px; font-weight: 600;
text-align: left; color: var(--accent);
border-bottom: 1px solid #e5e5e5; padding-bottom: 10px;
}
.sub-head {
margin: 40px auto 14px; font-size: 24px; font-weight: 600;
text-align: left; line-height: 1.3;
}
.sec-claim {
margin: 0 auto 14px; font-size: 24px; font-weight: 600;
text-align: left; line-height: 1.3;
}
h2 {
font-size: 30px; font-weight: 600; text-align: center; margin-bottom: 22px;
}
h3 { font-size: 20px; font-weight: 700; margin: 30px 0 10px; }
p.body-text { margin: 0 auto 20px; }
/* Key takeaways (collapsible) */
.takeaways { margin: 0 auto; }
details.takeaway {
background: var(--card-bg); border: 1px solid var(--border); border-radius: 10px;
margin: 12px 0; overflow: hidden;
}
details.takeaway summary {
list-style: none; cursor: pointer; display: flex; align-items: baseline; gap: 12px;
padding: 16px 20px; font-size: 17px; font-weight: 700; line-height: 1.5;
}
details.takeaway summary::-webkit-details-marker { display: none; }
details.takeaway summary .tk-num {
flex: none; width: 26px; height: 26px; border-radius: 50%;
background: var(--accent-dark); color: #fff; font-size: 14px; font-weight: 700;
display: inline-flex; align-items: center; justify-content: center;
align-self: flex-start; margin-top: 2px;
}
details.takeaway summary .chev {
margin-left: auto; align-self: center; flex: none;
color: var(--muted); font-size: 13px; transition: transform .2s ease;
}
details.takeaway[open] summary .chev { transform: rotate(180deg); }
details.takeaway summary:hover { background: #eef0f4; }
.tk-body { padding: 2px 20px 18px 58px; font-size: 15.5px; color: var(--muted); }
.tk-body b { color: var(--text); }
.tk-body figure { margin: 16px 0 4px; }
.tk-body .rq-tag { font-size: 13px; font-weight: 600; color: var(--accent-dark); letter-spacing: .03em; }
table.mini {
border-collapse: collapse; font-size: 14.5px; margin: 12px auto 4px;
}
table.mini th, table.mini td { padding: 5px 14px; text-align: center; border-bottom: 1px solid var(--border); }
table.mini td:first-child, table.mini th:first-child { text-align: left; }
table.mini thead th { border-bottom: 2px solid var(--text); }
table.mini tr.ours { background: var(--ours-bg); font-weight: 600; }
/* two-column media rows */
.row2 { display: grid; grid-template-columns: 1fr 1fr; gap: 26px; align-items: center; margin: 28px 0; }
.tldr-row { display: grid; grid-template-columns: 0.84fr 1fr; gap: 26px; align-items: center; margin: 26px auto 0; max-width: 960px; }
.tldr-row .cell-r { border-left: 1px solid var(--border); padding-left: 30px; }
@media (max-width: 760px) { .tldr-row { grid-template-columns: 1fr; } .tldr-row .cell-r { border-left: none; padding-left: 0; } }
.row2 .stat-line + .stat-line { margin-top: 0; }
@media (max-width: 760px) { .row2 { grid-template-columns: 1fr; } }
/* equation block */
.eqs {
max-width: 640px; margin: 10px auto 34px; background: var(--card-bg);
border: 1px solid var(--border); border-radius: 10px; padding: 12px 22px;
}
.eq { display: flex; align-items: baseline; justify-content: space-between; gap: 18px; padding: 7px 0; flex-wrap: wrap; }
.eq + .eq { border-top: 1px dashed var(--border); }
.eq-math { font-family: "STIX Two Math", "Cambria Math", Georgia, serif; font-size: 16.5px; white-space: nowrap; }
.eq-note { font-size: 13.5px; color: var(--muted); }
/* ---------- Tables ---------- */
.table-wrap { overflow-x: auto; margin: 24px auto; }
table.results {
border-collapse: collapse; margin: 0 auto; font-size: 15px; min-width: 640px;
}
table.results caption {
caption-side: top; text-align: left; font-size: 14.5px; color: var(--muted); padding-bottom: 10px;
}
table.results th, table.results td { padding: 7px 14px; text-align: center; }
table.results thead th { border-bottom: 2px solid var(--text); font-weight: 700; }
table.results td:first-child, table.results th:first-child { text-align: left; }
table.results tbody tr { border-bottom: 1px solid var(--border); }
table.results tr.group td { font-style: italic; color: var(--muted); border-bottom: none; padding-top: 12px; }
table.results tr.ours { background: var(--ours-bg); }
table.results tr.ours td { font-weight: 600; }
table.results .best { font-weight: 700; }
/* ---------- Video placeholder ---------- */
.video-grid { display: grid; grid-template-columns: repeat(2, 1fr); gap: 16px; }
@media (max-width: 640px) { .video-grid { grid-template-columns: 1fr; } }
.video-slot {
border: 2px dashed #c9ced8; border-radius: 10px; aspect-ratio: 16/9;
display: flex; align-items: center; justify-content: center;
color: var(--muted); font-size: 15px; text-align: center; padding: 12px;
background: var(--card-bg);
}
.video-grid video { width: 100%; border-radius: 10px; display: block; }
.teaser-video { max-width: 860px; margin: 28px auto 0; }
.teaser-video video, .wide-video video { width: 100%; border-radius: 10px; display: block; }
.wide-video { max-width: 900px; margin: 24px auto 0; }
/* per-task rollout gallery */
.task-block { width: min(1120px, 96vw); margin: 34px auto 0; position: relative; left: 50%; transform: translateX(-50%); }
.task-name { font-weight: 700; font-size: 17px; margin-bottom: 8px; text-align: center; }
.vid-pair { display: grid; grid-template-columns: 1fr 1fr; gap: 14px; }
@media (max-width: 700px) { .vid-pair { grid-template-columns: 1fr; } }
.vid-pair video { width: 100%; border-radius: 8px; display: block; }
.vid-tag {
font-size: 12.5px; font-weight: 600; color: var(--muted); margin-bottom: 5px;
text-align: center; letter-spacing: .04em; text-transform: uppercase;
}
.vid-tag.unseen { color: var(--accent-dark); }
.vid-tag.train { color: #b91c1c; }
/* unseen rollout: π0.5 clip stacked above pointmap clip, overlay labels */
.roll-cell { position: relative; }
.roll-cell video { width: 100%; border-radius: 8px; display: block; }
.roll-cell + .roll-cell { margin-top: 8px; }
.train-grid .roll-cell { margin-top: 0; }
.roll-compact .roll-label { font-size: 11px; padding: 2px 8px; top: 6px; left: 6px; }
.roll-label {
position: absolute; top: 9px; left: 9px; z-index: 2;
font-size: 12.5px; font-weight: 600; letter-spacing: .02em;
padding: 3px 11px; border-radius: 999px;
background: rgba(255,255,255,.9); border: 1px solid var(--border);
}
.roll-label.fail { color: #b91c1c; border-color: #fca5a5; }
.roll-label.ok { color: #15803d; border-color: #86efac; }
/* training-view grid */
.train-grid { display: grid; grid-template-columns: 96px repeat(4, 1fr); gap: 8px; align-items: center; }
.tg-task { text-align: center; font-size: 12.5px; font-weight: 600; letter-spacing: .04em; text-transform: uppercase; color: var(--muted); }
.train-grid img, .train-grid video { width: 100%; border-radius: 6px; display: block; }
.tg-head { text-align: center; font-size: 12.5px; font-weight: 600; letter-spacing: .04em; text-transform: uppercase; color: #b91c1c; }
.tg-label { font-size: 13px; font-weight: 700; color: #b91c1c; text-align: right; padding-right: 4px; line-height: 1.3; }
.tg-empty { display: flex; align-items: center; justify-content: center; text-align: center; font-size: 12.5px; color: var(--muted); background: var(--card-bg); border: 1px dashed var(--border); border-radius: 6px; align-self: stretch; }
/* task tab gallery */
.task-tabs { display: flex; flex-wrap: wrap; gap: 10px; justify-content: center; margin: 24px 0 4px; }
.task-pill {
font-family: "Google Sans", "Noto Sans", sans-serif;
font-size: 15px; font-weight: 600; color: var(--text);
background: #fff; border: 1.5px solid var(--border); border-radius: 999px;
padding: 8px 22px; cursor: pointer;
transition: border-color .15s ease, background .15s ease, color .15s ease;
}
.task-pill:hover { border-color: var(--accent); color: var(--accent); }
.task-pill.active { background: var(--accent); border-color: var(--accent); color: #fff; }
.task-pane.hidden-pane { display: none; }
.pane-label { text-align: center; font-size: 13px; font-weight: 700; letter-spacing: .05em; text-transform: uppercase; color: var(--muted); margin-bottom: 8px; }
/* ---------- BibTeX ---------- */
.bibtex-box { position: relative; margin: 20px auto 0; }
pre.bibtex {
background: #f4f5f7; border: 1px solid var(--border); border-radius: 10px;
padding: 22px 24px; overflow-x: auto; font-size: 14px; line-height: 1.5;
font-family: ui-monospace, SFMono-Regular, Menlo, Consolas, monospace;
}
.copy-btn {
position: absolute; top: 12px; right: 12px;
background: #fff; border: 1px solid var(--border); border-radius: 8px;
font-size: 13px; padding: 5px 12px; cursor: pointer; color: var(--muted);
}
.copy-btn:hover { color: var(--text); border-color: var(--muted); }
/* ---------- Footer ---------- */
footer {
border-top: 1px solid var(--border); margin-top: 70px; padding: 28px 0 40px;
font-size: 14px; color: var(--muted); text-align: center;
}
footer a { color: var(--accent); text-decoration: none; }
</style>
</head>
<body>
<!-- =============================== HERO =============================== -->
<div class="container hero">
<h1 class="title">See like a Robot: <span class="subtitle-break">Robot-Centric Pointmaps for Vision-Language-Action Models</span></h1>
<p class="authors">
<a href="https://lee15253.github.io/">Byungkun Lee</a><sup>1*</sup>,
<a href="https://godnpeter.github.io/">Dongyoon Hwang</a><sup>1*</sup>,
<a href="https://k00dj-19.github.io/">Dongjin Kim</a><sup>1</sup>,
<a href="https://joonleesky.github.io/">Hojoon Lee</a><sup>2</sup>,
<a href="https://pmh9960.github.io/">Minho Park</a><sup>1</sup>,
<a href="https://sites.google.com/site/jaegulchoo/">Jaegul Choo</a><sup>1</sup>
</p>
<p class="affil"><sup>1</sup>KAIST AI <sup>2</sup>Holiday Robotics</p>
<p class="eqnote"><sup>*</sup>Equal contribution</p>
<div class="btn-row">
<a class="btn" href="https://arxiv.org/abs/2607.11498">
<svg viewBox="0 0 384 512"><path d="M181.9 256.1c-5-16-4.9-46.9-2-46.9 8.4 0 7.6 36.9 2 46.9zm-1.7 47.2c-7.7 20.2-17.3 43.3-28.4 62.7 18.3-7 39-17.2 62.9-21.9-12.7-9.6-24.9-23.4-34.5-40.8zM86.1 428.1c0 .8 13.2-5.4 34.9-40.2-6.7 6.3-29.1 24.5-34.9 40.2zM248 160h136v328c0 13.3-10.7 24-24 24H24c-13.3 0-24-10.7-24-24V24C0 10.7 10.7 0 24 0h200v136c0 13.2 10.8 24 24 24zm-8 171.8c-20-12.2-33.3-29-42.7-53.8 4.5-18.5 11.6-46.6 6.2-64.2-4.7-29.4-42.4-26.5-47.8-6.8-5 18.3-.4 44.1 8.1 77-11.6 27.6-28.7 64.6-40.8 85.8-.1 0-.1.1-.2.1-27.1 13.9-73.6 44.5-54.5 68 5.6 6.9 16 10 21.5 10 17.9 0 35.7-18 61.1-61.8 25.8-8.5 54.1-19.1 79-23.2 21.7 11.8 47.1 19.5 64 19.5 29.2 0 31.2-32 19.7-43.4-13.9-13.6-54.3-9.7-73.6-7.2zM377 105L279 7c-4.5-4.5-10.6-7-17-7h-6v128h128v-6.1c0-6.3-2.5-12.4-7-16.9zm-74.1 255.3c4.1-2.7-2.5-11.9-42.8-9 37.1 15.8 42.8 9 42.8 9z"/></svg>
arXiv
</a>
<a class="btn" href="static/pdfs/see_like_a_robot.pdf">
<svg viewBox="0 0 512 512"><path d="M0 64C0 28.7 28.7 0 64 0H224V128c0 17.7 14.3 32 32 32H384V304H176c-35.3 0-64 28.7-64 64V512H64c-35.3 0-64-28.7-64-64V64zm384 64H256V0L384 128zM176 352h32c30.9 0 56 25.1 56 56s-25.1 56-56 56H192v32c0 8.8-7.2 16-16 16s-16-7.2-16-16V368c0-8.8 7.2-16 16-16zm32 80c13.3 0 24-10.7 24-24s-10.7-24-24-24H192v48h16zm96-80h32c26.5 0 48 21.5 48 48v64c0 26.5-21.5 48-48 48H304c-8.8 0-16-7.2-16-16V368c0-8.8 7.2-16 16-16zm32 128c8.8 0 16-7.2 16-16V400c0-8.8-7.2-16-16-16H320v96h16zm80-112c0-8.8 7.2-16 16-16h48c8.8 0 16 7.2 16 16s-7.2 16-16 16H448v32h32c8.8 0 16 7.2 16 16s-7.2 16-16 16H448v48c0 8.8-7.2 16-16 16s-16-7.2-16-16V368z"/></svg>
Paper
</a>
<!-- TODO: 코드 공개 시 링크 교체, 비공개면 class="btn disabled" 로 두고 "Code (soon)" -->
<a class="btn disabled" href="#">
<svg viewBox="0 0 496 512"><path d="M165.9 397.4c0 2-2.3 3.6-5.2 3.6-3.3.3-5.6-1.3-5.6-3.6 0-2 2.3-3.6 5.2-3.6 3-.3 5.6 1.3 5.6 3.6zm-31.1-4.5c-.7 2 1.3 4.3 4.3 4.9 2.6 1 5.6 0 6.2-2s-1.3-4.3-4.3-5.2c-2.6-.7-5.5.3-6.2 2.3zm44.2-1.7c-2.9.7-4.9 2.6-4.6 4.9.3 2 2.9 3.3 5.9 2.6 2.9-.7 4.9-2.6 4.6-4.6-.3-1.9-3-3.2-5.9-2.9zM244.8 8C106.1 8 0 113.3 0 252c0 110.9 69.8 205.8 169.5 239.2 12.8 2.3 17.3-5.6 17.3-12.1 0-6.2-.3-40.4-.3-61.4 0 0-70 15-84.7-29.8 0 0-11.4-29.1-27.8-36.6 0 0-22.9-15.7 1.6-15.4 0 0 24.9 2 38.6 25.8 21.9 38.6 58.6 27.5 72.9 20.9 2.3-16 8.8-27.1 16-33.7-55.9-6.2-112.3-14.3-112.3-110.5 0-27.5 7.6-41.3 23.6-58.9-2.6-6.5-11.1-33.3 2.6-67.9 20.9-6.5 69 27 69 27 20-5.6 41.5-8.5 62.8-8.5s42.8 2.9 62.8 8.5c0 0 48.1-33.6 69-27 13.7 34.7 5.2 61.4 2.6 67.9 16 17.7 25.8 31.5 25.8 58.9 0 96.5-58.9 104.2-114.8 110.5 9.2 7.9 17 22.9 17 46.4 0 33.7-.3 75.4-.3 83.6 0 6.5 4.6 14.4 17.3 12.1C428.2 457.8 496 362.9 496 252 496 113.3 383.5 8 244.8 8z"/></svg>
Code
</a>
<a class="btn" href="#bibtex">
<svg viewBox="0 0 448 512"><path d="M448 360V24c0-13.3-10.7-24-24-24H96C43 0 0 43 0 96v320c0 53 43 96 96 96h328c13.3 0 24-10.7 24-24v-16c0-7.5-3.5-14.3-8.9-18.7-4.2-15.4-4.2-59.3 0-74.7 5.4-4.3 8.9-11.1 8.9-18.6zM128 134c0-3.3 2.7-6 6-6h212c3.3 0 6 2.7 6 6v20c0 3.3-2.7 6-6 6H134c-3.3 0-6-2.7-6-6v-20zm0 64c0-3.3 2.7-6 6-6h212c3.3 0 6 2.7 6 6v20c0 3.3-2.7 6-6 6H134c-3.3 0-6-2.7-6-6v-20zm253.4 250H96c-17.7 0-32-14.3-32-32 0-17.6 14.4-32 32-32h285.4c-1.9 17.1-1.9 46.9 0 64z"/></svg>
BibTeX
</a>
</div>
</div>
<!-- =============================== TEASER (1-min overview video) =============================== -->
<div class="container">
<div style="max-width: 960px; margin: 28px auto 0;">
<video autoplay muted loop playsinline controls style="width: 100%; border-radius: 10px; display: block;">
<source src="static/videos/teaser.mp4?v=4" type="video/mp4">
</video>
</div>
</div>
<!-- =============================== TL;DR (highlight first, details below) =============================== -->
<section class="container" id="tldr" style="margin: 44px auto 8px;">
<p style="text-align: center; font-family: 'Google Sans', 'Noto Sans', sans-serif; font-size: 22px; font-weight: 700; line-height: 1.45; margin: 0 auto 6px; max-width: 940px;">
TL;DR — Give a VLA per-pixel 3D in the robot frame,<br>and it stays robust as training-time camera viewpoint variation grows.
</p>
<div class="tldr-row">
<img src="static/images/arch_short.png?v=1" alt="Architecture: a pointmap encoder is added next to the RGB encoder and its tokens are element-wise added before the VLA backbone." style="width: 100%;">
<div class="cell-r">
<img src="static/images/fig6_robustness.png?v=7" alt="As camera variation during training grows from none to high, RGB-only drops from 34.5 to 24.9 while RGB + Pointmap stays near 36, widening the gap from +3.1 to +10.9 points." style="width: 100%;">
</div>
</div>
<p class="body-text" style="text-align: center; font-size: 15.5px; color: var(--muted); max-width: 960px; margin: 22px auto 0;">
Robot datasets mix demonstrations from many camera viewpoints; as that training-time variation grows,<br>
an RGB-only policy degrades while pointmaps keep it robust, widening the gap to <b style="color: var(--text);">+10.9 points</b>.<br>
<b style="color: var(--text);">The whole method is one extra encoder and one element-wise addition.</b>
</p>
</section>
<!-- =============================== SETTING =============================== -->
<section class="container" id="setting">
<h2 class="sec-label">Motivation</h2>
<div class="sec-claim">Large-Scale Robot Data Spans Many Camera Viewpoints,<br>but Actions Are Defined in the Robot Frame</div>
<figure style="margin: 24px 0 0;">
<img src="static/images/droid_viewpoints.png?v=3" alt="Third-person camera viewpoint distribution around a Franka robot in DROID." style="max-width: 680px; width: 100%;">
<figcaption style="text-align: center;">
Third-person camera viewpoints in DROID (subsampled); brighter colors indicate regions of
higher viewpoint density. Figure from <a href="https://droid-dataset.github.io/" target="_blank" rel="noopener">DROID</a>.
</figcaption>
</figure>
<p class="body-text" style="margin-top: 20px;">
Large-scale robot datasets aggregate demonstrations collected across labs and episodes,
often from different camera viewpoints. <a href="https://droid-dataset.github.io/" target="_blank" rel="noopener">DROID</a>, for example, contains 1,417 unique
third-person viewpoints.
</p>
<p class="body-text" style="margin-top: 28px;">
Why does this viewpoint variation matter?
<b>A VLA observes the scene in the camera frame but predicts actions in a robot-centric
frame, creating an observation-to-action frame mismatch.</b>
With a fixed camera, this mapping remains consistent across demonstrations. As training
data spans more viewpoints, however, the policy must generalize the mapping from multiple
camera frames to the shared robot-frame action space, making the task harder to learn.
</p>
<p class="body-text">
To test this, we conduct a controlled study on 24 RoboCasa tasks. Using a π-style model
initialized from a base PaliGemma checkpoint, we vary only the amount of camera viewpoint
randomization across training demonstrations while keeping the remaining training and
evaluation settings fixed.
</p>
<figure style="margin: 24px auto 0; max-width: 1000px;">
<img src="static/images/camera_variation_training.png?v=1" alt="Camera poses in the training data under No variation, Low, and High camera viewpoint randomization, shown as camera frustums around a RoboCasa kitchen scene." style="width: 100%;">
<figcaption style="text-align: center;">
The two third-person cameras' poses in the training demonstrations, shown as frustums over a
RoboCasa scene. From left to right, we widen the camera viewpoint randomization from none to high.
</figcaption>
</figure>
<div class="evidence-row" style="margin-top: 24px;">
<figure style="margin: 0;">
<img src="static/images/fig6_rgb_only.png?v=4" alt="RGB-only success decreases from 34.5% to 24.9% as training-time camera viewpoint variation increases.">
</figure>
<div>
<div class="stat-line">
<div class="stat-num bad">−9.6 points</div>
<div class="stat-desc">RGB-only degrades as training-time viewpoint variation increases.</div>
</div>
</div>
</div>
</section>
<!-- =============================== METHOD =============================== -->
<section class="container" id="method">
<h2 class="sec-label">Solution</h2>
<div class="sec-claim">Bridge Camera-Frame Observations and Robot-Frame Actions<br>with Robot-Centric Pointmaps</div>
<p class="body-text">
Rather than asking the policy to learn how each camera frame maps to robot-frame actions,
we express the observed 3D geometry from different viewpoints in a shared robot frame.
This gives the policy a consistent geometric coordinate system across diverse camera viewpoints.
Robot-centric pointmaps bridge this frame mismatch while preserving the image-form
representation used by pretrained VLAs.
</p>
<ul class="req-list">
<li><b>Robot-frame geometry.</b> Each pixel stores the 3D coordinate of its corresponding scene point in the frame where robot actions are defined.</li>
<li><b>Image-form structure.</b> The dense H × W grid is preserved, maintaining spatial correspondence with RGB.</li>
</ul>
<figure>
<img src="static/images/fig2_concept.png" alt="Observations from different camera frames are expressed in the shared robot frame where actions are defined." style="max-width: 660px;">
</figure>
<p class="body-text">
A robot-centric pointmap stores robot-frame XYZ coordinates at the corresponding image pixels,
preserving the same dense H × W grid as RGB.
</p>
<p class="body-text" style="font-family: 'Google Sans', 'Noto Sans', sans-serif; font-size: 21px; font-weight: 700; margin-top: 30px;">
The whole method: one extra encoder and one element-wise addition.
</p>
<figure class="framed">
<img src="static/images/fig10_architecture.png?v=4" alt="Full model architecture: each RGB-D observation becomes a robot-centric pointmap, encoded by a second image encoder and added element-wise to the RGB tokens.">
<figcaption><b>Full architecture.</b> Each camera's pointmap is encoded by a second image encoder, initialized from the RGB encoder, and added element-wise to the RGB tokens right next to it. Nothing else in the VLA changes.</figcaption>
</figure>
</section>
<!-- =============================== PAYOFF (mirrors the intro study) =============================== -->
<section class="container" id="payoff">
<h2 class="sec-label">Results</h2>
<div class="sec-claim">Pointmaps Keep the Policy Robust as Training-Time Viewpoint Variation Increases</div>
<p class="body-text">
Adding pointmaps to the same controlled study changes the trend: RGB + Pointmap drops by
only 1.8 points, compared with 9.6 points for RGB-only.
</p>
<div class="evidence-row">
<figure style="margin: 0;">
<img src="static/images/fig6_robustness.png?v=7" alt="On the same axes as the intro chart, RGB + Pointmap decreases from 37.6% to 35.8% while RGB-only decreases from 34.5% to 24.9%.">
</figure>
<div style="position: relative; top: -38px;">
<div class="stat-line">
<div class="stat-num bad">−9.6 points</div>
<div class="stat-desc">RGB-only degrades as training-time viewpoint variation increases.</div>
</div>
<div class="stat-line">
<div class="stat-num good">+3.1 → +10.9 points</div>
<div class="stat-desc">The pointmap advantage over RGB-only widens as training-time viewpoint variation increases.</div>
</div>
</div>
</div>
</section>
<!-- =============================== SIMULATION RESULTS =============================== -->
<section class="container" id="results">
<h2 class="sub-head">Pointmaps Improve Both Pretrained VLA Backbones</h2>
<p class="body-text">
On RoboCasa, third-person camera viewpoints are randomized across training demonstrations.
In this setting, pointmaps improve both pretrained backbones:
<b>+7.6</b> on π<sub>0.5</sub> and <b>+4.2</b> on SmolVLA.
</p>
<div class="row2" style="max-width: 720px; margin: 22px auto; align-items: stretch;">
<div class="stat-line">
<div class="stat-num good">55.3 → 62.9</div>
<div class="stat-desc">π<sub>0.5</sub> average success with pointmaps (+7.6).</div>
</div>
<div class="stat-line">
<div class="stat-num good">37.2 → 41.4</div>
<div class="stat-desc">SmolVLA average success with pointmaps (+4.2).</div>
</div>
</div>
<p class="body-text">
With π<sub>0.5</sub>, pointmaps reach <b>62.9%</b> average success, outperforming all
camera-aware, 3D-augmented, and point-cloud baselines.
</p>
<table class="mini" style="margin: 14px auto 6px;">
<thead><tr><th>Method</th><th>Category</th><th>Backbone</th><th>Avg. SR</th></tr></thead>
<tbody>
<tr><td>FP3</td><td>Point-cloud policy</td><td>–</td><td>42.8</td></tr>
<tr><td>π<sub>0.5</sub></td><td>–</td><td>π<sub>0.5</sub></td><td>55.3</td></tr>
<tr><td>OC-VLA</td><td>Camera-aware VLA</td><td>π<sub>0.5</sub></td><td>56.3</td></tr>
<tr><td>KYC</td><td>Camera-aware VLA</td><td>π<sub>0.5</sub></td><td>59.1</td></tr>
<tr><td>GeoVLA</td><td>3D-augmented VLA</td><td>π<sub>0.5</sub></td><td>57.1</td></tr>
<tr><td>PointVLA</td><td>3D-augmented VLA</td><td>π<sub>0.5</sub></td><td>57.3</td></tr>
<tr class="ours"><td>π<sub>0.5</sub> + Pointmap (ours)</td><td>Ours</td><td>π<sub>0.5</sub></td><td>62.9</td></tr>
</tbody>
</table>
<p class="body-text" style="text-align:center; font-size: 13.5px; color: var(--muted); margin-top: 4px;">
FP3 is a DROID-pretrained point-cloud policy without a VLA backbone.
Per-task-category results are in the paper.
</p>
</section>
<!-- =============================== REAL ROBOT =============================== -->
<section class="container" id="real-robot">
<h2 class="sub-head">On a Real Robot, the Gap Widens at an Unseen Camera Viewpoint</h2>
<p class="body-text">
We collect 180 demonstrations over four tasks on a FR3 while repositioning the
external camera across three training viewpoints. We then evaluate the policies at one seen
viewpoint and at a held-out viewpoint not observed during training.
</p>
<p class="body-text" style="margin-top: 34px;">
At evaluation, the external camera is placed either at one of the training viewpoints
(<b>seen</b>) or at a held-out viewpoint not used for data collection (<b>unseen</b>).
Pointmaps improve performance at both camera viewpoints, but the advantage grows from
<b>+5.0</b> points at the seen viewpoint to <b>+11.7</b> points at the unseen viewpoint.
By expressing observed geometry in the robot frame, pointmaps provide a more consistent
spatial representation when the external camera moves beyond the training viewpoints.
</p>
<div class="row2" style="max-width: 720px; margin: 22px auto; align-items: stretch;">
<div class="stat-line">
<div class="stat-num good">73.3 → 78.3</div>
<div class="stat-desc">Seen camera viewpoint, π<sub>0.5</sub> → π<sub>0.5</sub> + pointmap (+5.0).</div>
</div>
<div class="stat-line">
<div class="stat-num good">55.0 → 66.7</div>
<div class="stat-desc">Unseen camera viewpoint, π<sub>0.5</sub> → π<sub>0.5</sub> + pointmap (+11.7).</div>
</div>
</div>
<div class="table-wrap">
<table class="results">
<caption>Real-world success rate (%), 15 rollouts per task per camera viewpoint. DP3 is a from-scratch point-cloud diffusion policy, unlike FP3 in the simulation table (a DROID-pretrained point-cloud foundation policy).</caption>
<thead>
<tr><th>Eval. viewpoint</th><th>Model</th><th>Avg.</th><th>Pick-and-place</th><th>Stack blocks</th><th>Open drawer</th><th>Close drawer</th></tr>
</thead>
<tbody>
<tr><td rowspan="3">Seen</td><td>DP3</td><td>63.3</td><td>60.0</td><td>40.0</td><td>60.0</td><td>93.3</td></tr>
<tr><td>π<sub>0.5</sub></td><td>73.3</td><td>80.0</td><td>53.3</td><td>73.3</td><td>86.7</td></tr>
<tr class="ours"><td>π<sub>0.5</sub> + pointmap</td><td class="best">78.3</td><td class="best">86.7</td><td class="best">60.0</td><td>73.3</td><td class="best">93.3</td></tr>
<tr><td rowspan="3">Unseen</td><td>DP3</td><td>48.3</td><td>33.3</td><td>33.3</td><td>40.0</td><td>86.7</td></tr>
<tr><td>π<sub>0.5</sub></td><td>55.0</td><td>40.0</td><td>26.7</td><td>66.7</td><td>86.7</td></tr>
<tr class="ours"><td>π<sub>0.5</sub> + pointmap</td><td class="best">66.7</td><td class="best">53.3</td><td class="best">46.7</td><td class="best">73.3</td><td class="best">93.3</td></tr>
</tbody>
</table>
</div>
<h3 style="text-align:center; font-size: 18px; margin-top: 46px;">Per-Task Videos, From Training Views to Unseen-Viewpoint Rollouts</h3>
<p class="body-text" style="font-size: 15.5px; color: var(--muted);">
The top row shows one training
demonstration from each training viewpoint, seen from the policy's external input camera.
The rows below show evaluation rollouts at the unseen viewpoint, recorded from a separate
fixed camera, comparing π<sub>0.5</sub> and π<sub>0.5</sub> + Pointmap. All clips
play at 2× speed.
</p>
<div class="task-tabs">
<button class="task-pill active" data-task="stack">Stack blocks</button>
<button class="task-pill" data-task="pnp">Pick & place</button>
<button class="task-pill" data-task="open">Open drawer</button>
<button class="task-pill" data-task="close">Close drawer</button>
</div>
<div class="task-pane" data-task="stack">
<div class="task-block" style="margin-top: 20px;">
<div class="pane-label">Training demonstrations, three camera viewpoints</div>
<div class="train-grid" style="grid-template-columns: repeat(3, 1fr);">
<div class="tg-head">Train view 1</div>
<div class="tg-head">Train view 2</div>
<div class="tg-head">Train view 3</div>
<video data-sync="stacktrain" autoplay muted loop playsinline><source src="static/videos/train_views/train_stack_cfg1.mp4" type="video/mp4"></video>
<video data-sync="stacktrain" autoplay muted loop playsinline><source src="static/videos/train_views/train_stack_cfg2.mp4" type="video/mp4"></video>
<video data-sync="stacktrain" autoplay muted loop playsinline><source src="static/videos/train_views/train_stack_cfg3.mp4" type="video/mp4"></video>
</div>
<div class="pane-label" style="margin-top: 26px;">Evaluation rollouts, unseen viewpoint</div>
<div class="train-grid roll-compact" style="grid-template-columns: repeat(3, 1fr);">
<div class="roll-cell"><span class="roll-label fail">π<sub>0.5</sub> ✗</span>
<video data-sync="stack" autoplay muted loop playsinline><source src="static/videos/rollouts/unseen_stack_rgb.mp4" type="video/mp4"></video></div>
<div class="roll-cell"><span class="roll-label fail">π<sub>0.5</sub> ✗</span>
<video data-sync="stack" autoplay muted loop playsinline><source src="static/videos/rollouts/unseen_stack_rgb2.mp4" type="video/mp4"></video></div>
<div class="roll-cell"><span class="roll-label fail">π<sub>0.5</sub> ✗</span>
<video data-sync="stack" autoplay muted loop playsinline><source src="static/videos/rollouts/unseen_stack_rgb3.mp4" type="video/mp4"></video></div>
<div class="roll-cell"><span class="roll-label ok">π<sub>0.5</sub> + Pointmap ✓</span>
<video data-sync="stack" autoplay muted loop playsinline><source src="static/videos/rollouts/unseen_stack_pmp.mp4" type="video/mp4"></video></div>
<div class="roll-cell"><span class="roll-label ok">π<sub>0.5</sub> + Pointmap ✓</span>
<video data-sync="stack" autoplay muted loop playsinline><source src="static/videos/rollouts/unseen_stack_pmp2.mp4" type="video/mp4"></video></div>
<div class="roll-cell"><span class="roll-label ok">π<sub>0.5</sub> + Pointmap ✓</span>
<video data-sync="stack" autoplay muted loop playsinline><source src="static/videos/rollouts/unseen_stack_pmp3.mp4" type="video/mp4"></video></div>
</div>
</div>
</div>
<div class="task-pane hidden-pane" data-task="pnp">
<div class="task-block" style="margin-top: 20px;">
<div class="pane-label">Training demonstrations, three camera viewpoints</div>
<div class="train-grid" style="grid-template-columns: repeat(3, 1fr);">
<div class="tg-head">Train view 1</div>
<div class="tg-head">Train view 2</div>
<div class="tg-head">Train view 3</div>
<video data-sync="pnptrain" autoplay muted loop playsinline><source src="static/videos/train_views/train_pnp_cfg1.mp4" type="video/mp4"></video>
<video data-sync="pnptrain" autoplay muted loop playsinline><source src="static/videos/train_views/train_pnp_cfg2.mp4" type="video/mp4"></video>
<video data-sync="pnptrain" autoplay muted loop playsinline><source src="static/videos/train_views/train_pnp_cfg3.mp4" type="video/mp4"></video>
</div>
<div class="pane-label" style="margin-top: 26px;">Evaluation rollouts, unseen viewpoint</div>
<div class="train-grid roll-compact" style="grid-template-columns: repeat(3, 1fr);">
<div class="roll-cell"><span class="roll-label fail">π<sub>0.5</sub> ✗</span>
<video data-sync="pnp" autoplay muted loop playsinline><source src="static/videos/rollouts/unseen_pnp_rgb.mp4" type="video/mp4"></video></div>
<div class="roll-cell"><span class="roll-label ok">π<sub>0.5</sub> ✓</span>
<video data-sync="pnp" autoplay muted loop playsinline><source src="static/videos/rollouts/unseen_pnp_rgb3.mp4" type="video/mp4"></video></div>
<div class="roll-cell"><span class="roll-label fail">π<sub>0.5</sub> ✗</span>
<video data-sync="pnp" autoplay muted loop playsinline><source src="static/videos/rollouts/unseen_pnp_rgb2.mp4" type="video/mp4"></video></div>
<div class="roll-cell"><span class="roll-label ok">π<sub>0.5</sub> + Pointmap ✓</span>
<video data-sync="pnp" autoplay muted loop playsinline><source src="static/videos/rollouts/unseen_pnp_pmp.mp4" type="video/mp4"></video></div>
<div class="roll-cell"><span class="roll-label ok">π<sub>0.5</sub> + Pointmap ✓</span>
<video data-sync="pnp" autoplay muted loop playsinline><source src="static/videos/rollouts/unseen_pnp_pmp3.mp4" type="video/mp4"></video></div>
<div class="roll-cell"><span class="roll-label ok">π<sub>0.5</sub> + Pointmap ✓</span>
<video data-sync="pnp" autoplay muted loop playsinline><source src="static/videos/rollouts/unseen_pnp_pmp2.mp4?v=2" type="video/mp4"></video></div>
</div>
</div>
</div>
<div class="task-pane hidden-pane" data-task="open">
<div class="task-block" style="margin-top: 20px;">
<div class="pane-label">Training demonstrations, three camera viewpoints</div>
<div class="train-grid" style="grid-template-columns: repeat(3, 1fr);">
<div class="tg-head">Train view 1</div>
<div class="tg-head">Train view 2</div>
<div class="tg-head">Train view 3</div>
<video data-sync="opentrain" autoplay muted loop playsinline><source src="static/videos/train_views/train_open_cfg1.mp4" type="video/mp4"></video>
<video data-sync="opentrain" autoplay muted loop playsinline><source src="static/videos/train_views/train_open_cfg2.mp4" type="video/mp4"></video>
<video data-sync="opentrain" autoplay muted loop playsinline><source src="static/videos/train_views/train_open_cfg3.mp4" type="video/mp4"></video>
</div>
<div class="pane-label" style="margin-top: 26px;">Evaluation rollouts, unseen viewpoint</div>
<div class="train-grid roll-compact" style="grid-template-columns: repeat(3, 1fr);">
<div class="roll-cell"><span class="roll-label fail">π<sub>0.5</sub> ✗</span>
<video data-sync="open" autoplay muted loop playsinline><source src="static/videos/rollouts/unseen_open_rgb.mp4" type="video/mp4"></video></div>
<div class="roll-cell"><span class="roll-label fail">π<sub>0.5</sub> ✗</span>
<video data-sync="open" autoplay muted loop playsinline><source src="static/videos/rollouts/unseen_open_rgb2.mp4" type="video/mp4"></video></div>
<div class="roll-cell"><span class="roll-label ok">π<sub>0.5</sub> ✓</span>
<video data-sync="open" autoplay muted loop playsinline><source src="static/videos/rollouts/unseen_open_rgb3.mp4" type="video/mp4"></video></div>
<div class="roll-cell"><span class="roll-label ok">π<sub>0.5</sub> + Pointmap ✓</span>
<video data-sync="open" autoplay muted loop playsinline><source src="static/videos/rollouts/unseen_open_pmp3.mp4" type="video/mp4"></video></div>
<div class="roll-cell"><span class="roll-label ok">π<sub>0.5</sub> + Pointmap ✓</span>
<video data-sync="open" autoplay muted loop playsinline><source src="static/videos/rollouts/unseen_open_pmp2.mp4" type="video/mp4"></video></div>
<div class="roll-cell"><span class="roll-label ok">π<sub>0.5</sub> + Pointmap ✓</span>
<video data-sync="open" autoplay muted loop playsinline><source src="static/videos/rollouts/unseen_open_pmp.mp4" type="video/mp4"></video></div>
</div>
</div>
</div>
<div class="task-pane hidden-pane" data-task="close">
<div class="task-block" style="margin-top: 20px;">
<div class="pane-label">Training demonstrations, three camera viewpoints</div>
<div class="train-grid" style="grid-template-columns: repeat(3, 1fr);">
<div class="tg-head">Train view 1</div>
<div class="tg-head">Train view 2</div>
<div class="tg-head">Train view 3</div>
<video data-sync="closetrain" autoplay muted loop playsinline><source src="static/videos/train_views/train_close_cfg1.mp4" type="video/mp4"></video>
<video data-sync="closetrain" autoplay muted loop playsinline><source src="static/videos/train_views/train_close_cfg2.mp4" type="video/mp4"></video>
<video data-sync="closetrain" autoplay muted loop playsinline><source src="static/videos/train_views/train_close_cfg3.mp4" type="video/mp4"></video>
</div>
<div class="pane-label" style="margin-top: 26px;">Evaluation rollouts, unseen viewpoint</div>
<div class="train-grid roll-compact" style="grid-template-columns: repeat(3, 1fr);">
<div class="roll-cell"><span class="roll-label fail">π<sub>0.5</sub> ✗</span>
<video data-sync="close" autoplay muted loop playsinline><source src="static/videos/rollouts/unseen_close_rgb3.mp4?v=3" type="video/mp4"></video></div>
<div class="roll-cell"><span class="roll-label ok">π<sub>0.5</sub> ✓</span>
<video data-sync="close" autoplay muted loop playsinline><source src="static/videos/rollouts/unseen_close_rgb.mp4?v=4" type="video/mp4"></video></div>
<div class="roll-cell"><span class="roll-label ok">π<sub>0.5</sub> ✓</span>
<video data-sync="close" autoplay muted loop playsinline><source src="static/videos/rollouts/unseen_close_rgb2.mp4?v=5" type="video/mp4"></video></div>
<div class="roll-cell"><span class="roll-label ok">π<sub>0.5</sub> + Pointmap ✓</span>
<video data-sync="close" autoplay muted loop playsinline><source src="static/videos/rollouts/unseen_close_pmp3.mp4?v=3" type="video/mp4"></video></div>
<div class="roll-cell"><span class="roll-label ok">π<sub>0.5</sub> + Pointmap ✓</span>
<video data-sync="close" autoplay muted loop playsinline><source src="static/videos/rollouts/unseen_close_pmp.mp4?v=4" type="video/mp4"></video></div>
<div class="roll-cell"><span class="roll-label ok">π<sub>0.5</sub> + Pointmap ✓</span>
<video data-sync="close" autoplay muted loop playsinline><source src="static/videos/rollouts/unseen_close_pmp2.mp4?v=5" type="video/mp4"></video></div>
</div>
</div>
</div>
</section>
<!-- =============================== CONTROLLED ANALYSIS =============================== -->
<section class="container" id="analysis">
<h2 class="sec-label">Ablation Studies</h2>
<p class="body-text">
Why does the design look the way it does? Each choice is backed by its own controlled ablation
on RoboCasa. The backbone and training recipe stay fixed, and only the 3D input changes.
</p>
<p class="body-text" style="font-size: 14px; color: var(--muted); margin-top: -10px;">
For these controlled studies, we use a model without robot pretraining to isolate the effect of
each input design. Its success rates are therefore lower than those of the pretrained VLAs
evaluated later.
</p>
<div class="takeaways">
<details class="takeaway">
<summary><span class="tk-num">1</span>Pre-compute Robot-Frame Geometry<span class="chev">▼</span></summary>
<div class="tk-body">
<p style="margin-top:8px;">Plücker rays describe the camera geometry, and depth provides per-pixel distance. RGB + Plücker + Depth therefore contains the information needed to obtain robot-frame geometry, but leaves it for the policy to infer. Providing robot-frame geometry directly as a pointmap improves success from <b>31.6% to 34.7%</b>.</p>
<table class="mini">
<thead><tr><th>Input</th><th>Robot-frame geometry</th><th>SR</th></tr></thead>
<tbody>
<tr><td>RGB</td><td>not provided</td><td>27.9</td></tr>
<tr><td>RGB + Plücker</td><td>not provided</td><td>28.7</td></tr>
<tr><td>RGB + Plücker + Depth</td><td>left to the policy</td><td>31.6</td></tr>
<tr class="ours"><td>RGB + Pointmap</td><td>provided directly</td><td>34.7</td></tr>
</tbody>
</table>
</div>
</details>
<details class="takeaway">
<summary><span class="tk-num">2</span>Preserve the Image Grid<span class="chev">▼</span></summary>
<div class="tk-body">
<p style="margin-top:8px;">Pointmaps outperform point-cloud inputs while preserving the dense image grid used by the VLA. Because RGB and pointmap tokens share the same spatial layout, element-wise addition further improves success from <b>30.7% to 34.7%</b>.</p>
<table class="mini">
<thead><tr><th>3D input</th><th>Encoder</th><th>Fusion</th><th>SR</th></tr></thead>
<tbody>
<tr><td>None</td><td>–</td><td>–</td><td>27.9</td></tr>
<tr><td>Point cloud</td><td>MLP</td><td>concat</td><td>24.2</td></tr>
<tr><td>Point cloud</td><td>PTv3</td><td>concat</td><td>32.8</td></tr>
<tr><td>Pointmap</td><td>image encoder</td><td>concat</td><td>30.7</td></tr>
<tr class="ours"><td>Pointmap</td><td>image encoder</td><td>add</td><td>34.7</td></tr>
</tbody>
</table>
<figure>
<img src="static/images/fig5_fusion.png" alt="Element-wise addition preserves spatial correspondence; concatenation treats pointmap tokens as a separate sequence." style="max-width: 540px;">
</figure>
</div>
</details>
<details class="takeaway">
<summary><span class="tk-num">3</span>Center Geometry at the End Effector<span class="chev">▼</span></summary>
<div class="tk-body">
<p style="margin-top:8px;">Centering the pointmap at the current end-effector position reduces variation caused by absolute workspace location, making similar interactions appear in a more consistent local coordinate system. It improves success under both fixed and randomized evaluation viewpoints.</p>
<table class="mini">
<thead><tr><th>Pointmap origin</th><th>Fixed eval.</th><th>Randomized eval.</th><th>Drop</th></tr></thead>
<tbody>
<tr><td>Robot base</td><td>34.7</td><td>32.7</td><td>−2.0</td></tr>
<tr class="ours"><td>End effector</td><td>36.9</td><td>36.6</td><td>−0.3</td></tr>
</tbody>
</table>
<figure>
<img src="static/images/fig4_ee_centering.png" alt="End-effector centering concentrates interaction targets near a common origin.">
</figure>
</div>
</details>
</div>
</section>
<!-- =============================== BIBTEX =============================== -->
<section class="container" id="bibtex">
<h2 class="sec-label">BibTeX</h2>
<div class="bibtex-box">
<button class="copy-btn" onclick="copyBibtex(this)">Copy</button>
<!-- TODO: arXiv 공개 / 논문 accept 후 실제 엔트리로 교체 -->
<pre class="bibtex" id="bibtex-text">@article{lee2026seelikearobot,
title = {See like a Robot: Robot-Centric Pointmaps for Vision-Language-Action Models},
author = {Lee, Byungkun and Hwang, Dongyoon and Kim, Dongjin and Lee, Hojoon and Park, Minho and Choo, Jaegul},
journal = {arXiv preprint arXiv:2607.11498},
year = {2026},
eprint = {2607.11498},
archivePrefix = {arXiv},
primaryClass = {cs.RO}
}</pre>
</div>
</section>
<footer>
<div class="container">
Website template adapted from <a href="https://nerfies.github.io" target="_blank" rel="noopener">Nerfies</a>, licensed under <a href="https://creativecommons.org/licenses/by-sa/4.0/" target="_blank" rel="noopener">CC BY-SA 4.0</a>.
</div>
</footer>
<script>
function copyBibtex(btn) {
const text = document.getElementById('bibtex-text').innerText;
navigator.clipboard.writeText(text).then(() => {
btn.textContent = 'Copied!';
setTimeout(() => { btn.textContent = 'Copy'; }, 1600);
});
}
// Rollout pairs: each clip holds its last frame until every clip in the
// same task group has finished, then all restart together.
const syncGroups = new Map();
document.querySelectorAll('.sync-group').forEach((group, i) => {
syncGroups.set('g' + i, Array.from(group.querySelectorAll('video')));
});
document.querySelectorAll('video[data-sync]').forEach((v) => {
const key = 'k' + v.dataset.sync;
if (!syncGroups.has(key)) syncGroups.set(key, []);
syncGroups.get(key).push(v);
});
syncGroups.forEach((vids) => {
if (vids.length < 2) return;
const restartAll = () => vids.forEach((x) => { x.currentTime = 0; x.play(); });
const done = new Set();
vids.forEach((v) => {
v.removeAttribute('loop');
v.addEventListener('ended', () => {
done.add(v);
if (done.size === vids.length) { done.clear(); restartAll(); }
});
});
// start the first cycle together as well (or after 3 s if a clip is slow to load)
vids.forEach((v) => v.pause());
let started = false;
const startAll = () => { if (!started) { started = true; restartAll(); } };
const check = () => { if (vids.every((v) => v.readyState >= 3)) startAll(); };
vids.forEach((v) => v.addEventListener('canplaythrough', check));
setTimeout(startAll, 3000);
check();
});
// Per-task gallery tabs: only the active pane's videos play.
const taskPanes = document.querySelectorAll('.task-pane');
taskPanes.forEach((p) => {
p.querySelectorAll('video').forEach((v) => {
// if anything (autoplay or the sync logic above) starts a hidden pane's video, pause it
v.addEventListener('play', () => { if (p.classList.contains('hidden-pane')) v.pause(); });
});
});
document.querySelectorAll('.task-pill').forEach((btn) => {
btn.addEventListener('click', () => {
document.querySelectorAll('.task-pill').forEach((b) => b.classList.toggle('active', b === btn));
taskPanes.forEach((p) => {
const on = p.dataset.task === btn.dataset.task;
p.classList.toggle('hidden-pane', !on);
if (on) {
p.querySelectorAll('video').forEach((v) => { v.currentTime = 0; v.play().catch(() => {}); });
} else {
p.querySelectorAll('video').forEach((v) => v.pause());
}
});
});
});
</script>
</body>
</html>