Repository navigation
Expand file tree
/
Copy pathindex.html
More file actions
494 lines (472 loc) · 39.6 KB
/
Copy pathindex.html
File metadata and controls
494 lines (472 loc) · 39.6 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
<!DOCTYPE html>
<html lang="zh-CN">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<meta name="description" content="周苍田,哈尔滨工业大学控制科学与工程硕士。聚焦具身智能、VLA、世界模型与机器人学习,包含京东、无界动力、小米实习项目、论文与开源实践。">
<meta name="theme-color" content="#faf6ee">
<link rel="canonical" href="https://dev-embodiedai.github.io/">
<link rel="icon" href="favicon.svg" type="image/svg+xml">
<meta property="og:type" content="website">
<meta property="og:locale" content="zh_CN">
<meta property="og:title" content="周苍田 · 具身智能算法工程师">
<meta property="og:description" content="哈尔滨工业大学控制科学与工程硕士,聚焦 VLA、世界模型与机器人学习。教育背景、论文、实习与开源项目。">
<meta property="og:url" content="https://dev-embodiedai.github.io/">
<title>周苍田 · 具身智能算法工程师</title>
<style>
:root{
--bg:#faf6ee; --paper:#fffdf9; --ink:#2b2723; --sub:#6f675a; --faint:#9a917f;
--line:#eae1d0; --brand:#c05f2c; --brand2:#e08b4f; --gold:#c99a3b;
--green:#4e8d5b; --blue:#356089; --violet:#6b4fa0; --chip:#f3ebdc;
--mono:"SFMono-Regular",Consolas,Menlo,monospace;
--sans:-apple-system,BlinkMacSystemFont,"Segoe UI","PingFang SC","Microsoft YaHei",sans-serif;
}
*{margin:0;padding:0;box-sizing:border-box}
html{scroll-behavior:smooth}
body{font-family:var(--sans);background:var(--bg);color:var(--ink);line-height:1.65;font-size:15px}
img{max-width:100%;display:block}
a{color:var(--brand);text-decoration:none}
#prog{position:fixed;top:0;left:0;height:3px;background:linear-gradient(90deg,var(--brand),var(--gold));width:0;z-index:200}
/* nav */
#nav{position:fixed;top:0;left:0;right:0;height:54px;background:rgba(255,253,249,.9);
backdrop-filter:blur(12px);border-bottom:1px solid var(--line);z-index:100;
display:flex;align-items:center;justify-content:space-between;padding:0 28px}
#nav .brand{font-weight:700;font-size:16px;color:var(--brand)}
#nav .links{display:flex;gap:4px;flex-wrap:wrap}
#nav .links a{font-size:12.5px;color:var(--sub);padding:5px 10px;border-radius:7px;transition:.15s}
#nav .links a:hover{background:var(--chip);color:var(--brand)}
@media(max-width:920px){#nav .links{display:none}}
/* hero */
.hero{min-height:92vh;display:flex;align-items:center;justify-content:center;padding:80px 24px 40px;position:relative;overflow:hidden}
.hero::before{content:'';position:absolute;inset:0;
background:radial-gradient(ellipse 55% 45% at 12% 18%,rgba(192,95,44,.10),transparent 60%),
radial-gradient(ellipse 45% 40% at 88% 82%,rgba(78,141,91,.08),transparent 55%)}
.hero-inner{max-width:1000px;width:100%;position:relative;z-index:1}
.hero-top{display:flex;align-items:center;gap:30px;margin-bottom:28px}
.avatar{flex-shrink:0;width:120px;height:120px;border-radius:24px;overflow:hidden;
box-shadow:0 14px 34px -8px rgba(192,95,44,.4);border:3px solid #fff;background:var(--chip)}
.avatar img{width:100%;height:100%;object-fit:cover}
.hero h1{font-size:44px;font-weight:800;letter-spacing:-1px;line-height:1.1;margin-bottom:8px}
.hero h1 .en{font-size:19px;color:var(--faint);font-weight:500;margin-left:10px}
.hero .role{font-size:19px;color:var(--brand);font-weight:600;margin-bottom:12px}
.hero .tags{display:flex;gap:8px;flex-wrap:wrap}
.hero .tags span{font-size:12.5px;color:var(--sub);background:var(--paper);border:1px solid var(--line);padding:4px 11px;border-radius:16px}
/* edu/company logo row */
.logo-row{display:flex;align-items:center;gap:14px;flex-wrap:wrap;margin-bottom:24px;
padding:16px 20px;background:var(--paper);border:1px solid var(--line);border-radius:14px}
.logo-row .lbl{font-size:12px;color:var(--faint);font-weight:600;margin-right:4px}
.logo-chip{display:flex;align-items:center;gap:7px;font-size:13px;font-weight:600;color:var(--ink);
background:var(--bg);border:1px solid var(--line);padding:6px 12px;border-radius:10px}
.logo-chip img{width:20px;height:20px;object-fit:contain;border-radius:4px}
.logo-chip .yr{font-size:11px;color:var(--faint);font-weight:400}
/* stat row - horizontal */
.stat-row{display:grid;grid-template-columns:repeat(4,1fr);gap:14px;margin-bottom:20px}
.stat{background:var(--paper);border:1px solid var(--line);border-radius:14px;padding:18px 14px;text-align:center;transition:.2s}
.stat:hover{transform:translateY(-3px);box-shadow:0 10px 24px -12px rgba(192,95,44,.3);border-color:var(--brand2)}
.stat .n{font-size:30px;font-weight:800;color:var(--brand);line-height:1;letter-spacing:-1px}
.stat .n small{font-size:15px;font-weight:700}
.stat .l{font-size:12px;color:var(--sub);margin-top:6px;line-height:1.35}
/* timeline - horizontal */
.tl{display:flex;gap:0;align-items:stretch;background:var(--paper);border:1px solid var(--line);border-radius:14px;padding:8px;overflow-x:auto}
.tl .node{flex:1;min-width:130px;padding:14px 12px;position:relative;text-align:center}
.tl .node:not(:last-child)::after{content:'›';position:absolute;right:-6px;top:50%;transform:translateY(-50%);color:var(--brand2);font-size:20px;font-weight:800;z-index:2}
.tl .node .t{font-size:14px;font-weight:700;color:var(--ink);margin-bottom:3px}
.tl .node .d{font-size:11.5px;color:var(--sub);line-height:1.35}
.tl .node.hl .t{color:var(--brand)}
/* section - compact */
section{max-width:1000px;margin:0 auto;padding:48px 24px 4px;scroll-margin-top:60px}
.sec-head{margin-bottom:22px}
.sec-head .kicker{display:inline-flex;align-items:center;gap:7px;font-size:12.5px;font-weight:600;color:var(--brand);
background:var(--chip);padding:5px 13px;border-radius:16px;margin-bottom:10px}
.sec-head .kicker img{width:16px;height:16px;object-fit:contain;border-radius:3px}
.sec-head h2{font-size:29px;font-weight:800;letter-spacing:-.5px;line-height:1.15}
.sec-head h2 em{font-style:normal;color:var(--brand)}
.sec-head .sub{font-size:14.5px;color:var(--sub);margin-top:6px}
/* card - compact */
.card{background:var(--paper);border:1px solid var(--line);border-radius:16px;padding:26px;margin-bottom:16px}
.card h3{font-size:18px;font-weight:700;margin-bottom:11px;display:flex;align-items:center;gap:8px}
.card h3 .dot{width:7px;height:7px;border-radius:50%;background:var(--brand);flex-shrink:0}
.card p{color:var(--sub);margin-bottom:11px}
.card p strong,.card li strong{color:var(--ink);font-weight:600}
.card .em{color:var(--brand);font-weight:600}
ul.feat{list-style:none;margin:12px 0}
ul.feat li{padding-left:22px;position:relative;margin-bottom:8px;color:var(--sub)}
ul.feat li::before{content:'';position:absolute;left:3px;top:9px;width:6px;height:6px;border-radius:2px;background:var(--brand2);transform:rotate(45deg)}
/* figure */
figure{margin:16px 0}
figure img{width:100%;border-radius:12px;border:1px solid var(--line);background:#fff}
figcaption{font-size:12.5px;color:var(--faint);margin-top:7px;text-align:center;font-style:italic}
.fig-grid{display:grid;gap:14px;margin:16px 0}
.fig-grid.g2{grid-template-columns:1fr 1fr}
.fig-grid.g3{grid-template-columns:repeat(3,1fr)}
.fig-grid figure{margin:0}
@media(max-width:760px){.fig-grid.g2,.fig-grid.g3{grid-template-columns:1fr}.stat-row{grid-template-columns:1fr 1fr}}
/* callout */
.callout{border-radius:12px;padding:16px 20px;margin:14px 0;border-left:4px solid var(--brand);background:linear-gradient(100deg,rgba(192,95,44,.06),transparent)}
.callout.green{border-left-color:var(--green);background:linear-gradient(100deg,rgba(78,141,91,.07),transparent)}
.callout.blue{border-left-color:var(--blue);background:linear-gradient(100deg,rgba(53,96,137,.07),transparent)}
.callout .ct{font-weight:700;margin-bottom:6px;color:var(--ink);font-size:14.5px}
.callout ul.feat{margin:6px 0 0}
/* horizontal flow */
.flow{display:flex;align-items:stretch;gap:0;flex-wrap:nowrap;overflow-x:auto;margin:16px 0;padding-bottom:4px}
.flow .fstep{flex:1;min-width:110px;background:var(--chip);border:1px solid var(--line);border-radius:10px;padding:11px 10px;text-align:center;position:relative}
.flow .fstep .ft{font-weight:700;font-size:13px;color:var(--ink)}
.flow .fstep .fd{font-size:11px;color:var(--sub);margin-top:2px;line-height:1.3}
.flow .fstep.hot{border-color:var(--brand);background:linear-gradient(135deg,#fdeee2,#faf0e0)}
.flow .fstep.hot .ft{color:var(--brand)}
.flow .conn{display:flex;align-items:center;padding:0 6px;color:var(--brand2);font-weight:800;font-size:16px;flex-shrink:0}
/* table */
table{width:100%;border-collapse:collapse;margin:14px 0;font-size:13.5px;background:#fff;border-radius:10px;overflow:hidden;border:1px solid var(--line)}
th{background:var(--chip);color:var(--ink);text-align:left;padding:10px 12px;font-weight:700;font-size:12.5px}
td{padding:8px 12px;border-top:1px solid var(--line);color:var(--sub)}
td.win{color:var(--green);font-weight:700}
td.down{color:var(--brand);font-weight:600}
tr.total td{background:#fbf5ea;font-weight:700;color:var(--ink)}
.pills{display:flex;gap:7px;flex-wrap:wrap;margin:12px 0}
.pills span{font-size:12px;background:var(--chip);color:var(--sub);padding:4px 11px;border-radius:7px;font-family:var(--mono)}
/* tech stack - 3 col horizontal */
.stack{display:grid;grid-template-columns:repeat(3,1fr);gap:16px}
.stack .col{background:var(--bg);border:1px solid var(--line);border-radius:12px;padding:18px}
.stack .col h4{font-size:14.5px;margin-bottom:10px;padding-bottom:7px;border-bottom:2px solid var(--line)}
.stack .col .tag{display:inline-block;font-size:12px;color:var(--sub);background:var(--paper);border:1px solid var(--line);padding:3px 9px;border-radius:7px;margin:3px 3px 0 0}
@media(max-width:760px){.stack{grid-template-columns:1fr}}
code{font-family:var(--mono);background:var(--chip);padding:1px 5px;border-radius:4px;font-size:.87em;color:var(--brand)}
footer{text-align:center;padding:50px 24px 40px;color:var(--faint);font-size:13.5px;border-top:1px solid var(--line);margin-top:40px}
footer .big{font-size:17px;color:var(--ink);font-weight:600;margin-bottom:6px}
.reveal{opacity:0;transform:translateY(20px);transition:opacity .5s,transform .5s}
.reveal.in{opacity:1;transform:none}
</style>
<link rel="stylesheet" href="styles.css">
</head>
<body>
<a class="skip-link" href="#bg">跳到主要内容</a>
<div id="prog"></div>
<!--BODY-->
<nav id="nav">
<a class="brand" href="#hero">周苍田 <small>ZHOU CANGTIAN</small></a>
<button class="menu-toggle" type="button" aria-expanded="false" aria-controls="nav-links">菜单</button>
<div class="links" id="nav-links">
<a href="#bg">教育背景</a>
<a href="#publications">论文</a>
<a href="#experience">实习经历</a>
<a href="#projects">项目经历</a>
<a href="#skills">专业技能</a>
<a class="nav-download" href="output/Zhou_Cangtian_Resume.pptx" download>下载 PPT ↗</a>
</div>
</nav>
<main>
<!-- HERO -->
<section class="hero resume-header" id="hero" aria-label="个人信息">
<div class="hero-inner">
<div class="resume-heading">
<div>
<div class="resume-name"><h1>周苍田</h1><span class="english-name">Zhou Cangtian</span></div>
<p class="role">具身智能算法工程师</p>
<p class="hero-intro">VLA / WAM / 机器人学习</p>
</div>
<div class="actions"><a class="button primary" href="output/Zhou_Cangtian_Resume.pptx" download>下载 PPT <span aria-hidden="true">↗</span></a><a class="button secondary" href="https://github.com/Dev-EmbodiedAI" target="_blank" rel="noopener noreferrer">GitHub <span aria-hidden="true">↗</span></a></div>
</div>
<div class="hero-bottom"><div class="hero-social"><a href="https://github.com/Dev-EmbodiedAI" target="_blank" rel="noopener noreferrer">GitHub / Dev-EmbodiedAI ↗</a><span class="social-text"><img src="assets/pseudo.png" alt="个人头像插画" width="24" height="24">小红书 @Dev.E · 3500+ 粉丝</span></div><p>3 段产业实习 · 多模态理解 / 机器人控制 / 世界模型</p></div>
</div>
</section>
<section id="bg" aria-labelledby="education-title">
<div class="section-title"><div><span class="eyebrow">01 / 教育背景</span><h2 id="education-title">教育背景</h2></div></div>
<div class="education-layout">
<div><div class="education-heading"><img src="assets/hit.svg" alt="哈尔滨工业大学校徽" width="70" height="70"><div><h3>哈尔滨工业大学</h3><p>Harbin Institute of Technology · 2020 — 2027</p></div></div><dl class="education-facts"><div><dt>硕士 · 控制科学与工程(推免)</dt><dd>2027 届</dd></div><div><dt>本科 / 硕士 GPA</dt><dd>91.51 / 85.08</dd></div><div><dt>本科阶段</dt><dd>2020 — 2024</dd></div><div><dt>个人信息</dt><dd>2003.09 · 中共党员</dd></div></dl></div>
<div class="award-block"><div><p class="eyebrow" style="margin-bottom:12px">奖项</p><div class="award-result">全国一等奖</div><h3>第 17 届全国大学生<br>智能汽车竞赛</h3><p>百度智慧交通组 · 全国总决赛</p></div><figure><img src="assets/ug_cert1.jpg" alt="第17届全国大学生智能汽车竞赛全国一等奖证书" width="1820" height="2573" loading="lazy"><figcaption>获奖证书 · 点击放大</figcaption></figure></div>
</div>
</section>
<section id="publications" aria-labelledby="publications-title">
<div class="section-title"><div><span class="eyebrow">02 / 论文</span><h2 id="publications-title">论文</h2></div><p>视觉感知、工业缺陷检测、具身多模态理解</p></div>
<div class="publication-list">
<article class="publication"><span class="pub-number">01</span><div><h3>From Isolated Views to Holistic Logic: Learning Joint Multi-Scene Understanding for Embodied AI with Vision-Language Models.</h3><p class="pub-meta"><span>AAAI 2027</span><span class="pub-badge">CCF-A</span><span>HUGIN</span></p></div><p class="pub-status"><strong>共同第一作者</strong><span>在投</span></p></article>
<article class="publication"><span class="pub-number">02</span><div><h3>OpenMorphoQuery: From Visual Morphology Prompts to Target-Grounded Queries for Unseen Industrial Defect Localization.</h3><p class="pub-meta"><span>ICLR 2027</span><span class="pub-badge">CCF-A</span></p></div><p class="pub-status"><strong>第三作者</strong><span>在投</span></p></article>
<article class="publication"><span class="pub-number">03</span><div><h3>A robust defect detection method with a generalization enhancer and cross-modality aggregator for cylinder bores.</h3><p class="pub-meta"><span>EAAI 2024</span><span class="pub-badge">CCF-C</span></p></div><p class="pub-status"><strong>第三作者</strong></p></article>
<article class="publication"><span class="pub-number">04</span><div><h3>RoMATer: An end-to-end robust multiaircraft tracker with transformers.</h3><p class="pub-meta"><span>IJCNN 2024</span><span class="pub-badge">CCF-C</span></p></div><p class="pub-status"><strong>第三作者</strong></p></article>
</div>
</section>
<section id="experience" aria-labelledby="experience-title">
<div class="section-title"><div><span class="eyebrow">03 / 实习经历</span><h2 id="experience-title">实习经历</h2></div></div>
<div class="experience-list">
<article class="experience-item"><div class="employer"><span class="employer-logo"><img src="assets/jd.svg" alt="京东 logo" width="38" height="38"></span><div><h3>京东物流</h3><p>X 科技部</p></div></div><time>2025.08 — 2026.02</time><p class="experience-topic">跨视角 VLM 决策 · Sim2Real 强化学习</p><a class="experience-link" href="#jd">HUGIN 与 CoPack 项目 ↓</a></article>
<article class="experience-item"><div class="employer"><span class="employer-logo anyverse"><img src="assets/无界动力LOGO.png" alt="无界动力 logo" width="46" height="36"></span><div><h3>无界动力</h3><p>Anyverse Dynamics</p></div></div><time>2026.05 — 2026.07</time><p class="experience-topic">π0.5 VLA · 跨本体混训与模块消融</p><a class="experience-link" href="#vla">VLA 混训与评测 ↓</a></article>
<article class="experience-item"><div class="employer"><span class="employer-logo"><img src="assets/xiaomi.svg" alt="小米 logo" width="38" height="38"></span><div><h3>小米<span class="current">至今</span></h3><p>WAM 模型方向</p></div></div><time>2026.07 — 至今</time><p class="experience-topic">RGB + Optical Flow 双流世界模型</p><a class="experience-link" href="#wam">双流 WAM 项目 ↓</a></article>
</div>
<!-- ═══ 京东 · 分拣 ═══ -->
<article id="jd" class="project-section" aria-labelledby="jd-title">
<div class="project-topline"><span class="project-label">01 / HUGIN · 京东 X 科技部</span><time>2025.11 — 2026.02</time></div>
<div class="project-summary">
<figure class="project-image"><img src="assets/jd_sysarch.png" alt="HUGIN 多相机观测、VLM 规划与机械臂控制系统架构"><figcaption>多视角局部观测 → 联合场景理解 → 机械臂动作规划</figcaption></figure>
<div class="project-copy"><h2 id="jd-title">HUGIN:跨视角物流分拣</h2><p class="project-description">面向多格口、多笼车的物流分拣,结合 EDA 数据扩增、GCR 全局上下文排序损失与辅助感知监督,提升跨视角 VLM 决策准确率。</p><div class="result-line"><div class="result-item"><strong>78.8<small>%</small></strong><span>SortingBench 准确率</span></div><div class="result-item"><strong>+15.2<small> 个百分点</small></strong><span>相对 SFT 提升 23.9%</span></div></div><div class="method-tags"><span>VLM / Qwen3-VL</span><span>GCR · 零推理开销</span><span>AAAI 2027 在投 · 共同一作</span></div></div>
</div>
<details class="project-details"><summary>深入了解 HUGIN:任务场景、数据扩增与完整实验</summary>
<div class="card reveal">
<h3><span class="dot"></span>任务背景:多格口 × 多笼车的复杂分拣</h3>
<p>京东物流单臂需同时服务 <strong>2+ 分拣格口 × 4 笼车</strong>,每个相机只能看到局部场景,跨视角互补观测是核心挑战。</p>
<div class="fig-grid g3">
<figure><img src="assets/jd_scene.png" alt=""><figcaption>实拍:分拣格口与笼车,包裹杂乱</figcaption></figure>
<figure><img src="assets/jd_layout.png" alt=""><figcaption>俯视布局:2 格口 × 4 笼车 × 暂存区</figcaption></figure>
<figure><img src="assets/jd_multiview.png" alt=""><figcaption>多相机跨视角观测</figcaption></figure>
</div>
<figure><img src="assets/jd_sysarch.png" alt=""><figcaption>系统架构:多相机 → VLM Planner 输出分拣计划 → Arm-Controllers 执行</figcaption></figure>
<figure><img src="assets/jd_actionseq.png" alt=""><figcaption>产线实录:Reset → Move(格口) → Grasp → Move(笼车) → Place 完整动作序列</figcaption></figure>
</div>
<div class="card reveal">
<h3><span class="dot"></span>方案一:EDA 数据扩增(2k → 2.6w)</h3>
<p>原始人工标注仅 2000 条,直接 SFT 严重不足。设计三阶段数据扩增:</p>
<div class="flow">
<div class="fstep"><div class="ft">2k</div><div class="fd">原始样本</div></div>
<div class="conn">→</div>
<div class="fstep"><div class="ft">6k</div><div class="fd">子任务拆分<br>sub-task VQA</div></div>
<div class="conn">+</div>
<div class="fstep"><div class="ft">2k</div><div class="fd">重组 task</div></div>
<div class="conn">+</div>
<div class="fstep"><div class="ft">1.4w</div><div class="fd">辅助感知<br>YOLO/SAM3 生成</div></div>
<div class="conn">+</div>
<div class="fstep"><div class="ft">4k</div><div class="fd">通用 VQA<br>防过拟合</div></div>
<div class="conn">→</div>
<div class="fstep hot"><div class="ft">2.6w</div><div class="fd">训练数据</div></div>
</div>
<div class="fig-grid g2">
<figure><img src="assets/jd_detect.png" alt=""><figcaption>辅助感知:YOLO 生成包裹检测 VQA,抑制小目标定位幻觉</figcaption></figure>
<figure><img src="assets/jd_hugin_arch.png" alt=""><figcaption>HUGIN 核心架构:Encoder → Projector → LLM 生成动作,同时算 GCR 对比损失</figcaption></figure>
</div>
</div>
<div class="card reveal">
<h3><span class="dot"></span>方案二:GCR Loss — 强制利用全局多图上下文</h3>
<p>SFT 易让模型只看单图作答。引入 <strong>Global Context Ranking Loss</strong>:要求"含全部视角"的输入比"含部分视角"产生更高的指令相似度,<span class="em">零推理开销</span>。</p>
<div class="callout">
<div class="ct">GCR 损失</div>
<p style="font-family:var(--mono);font-size:13px;margin:0;color:var(--sub)">𝓛<sub>GCR</sub> = 𝔼<sub>n~𝒰{1,N-1}</sub>[ max(0, α + SG(ẽ<sub>n</sub>)ᵀq̃ − ẽ<sub>N</sub>ᵀq̃) ] ,全局上下文相似度须高于部分上下文,margin=α</p>
</div>
<table>
<tr><th>方法</th><th>SortingBench 准确率</th><th>说明</th></tr>
<tr><td>Base(零样本 Qwen3-VL)</td><td class="down">0%</td><td>完全无法完成</td></tr>
<tr><td>SFT(仅 2k 原始数据)</td><td>63.6%</td><td>基线微调</td></tr>
<tr class="total"><td>HUGIN(EDA + GCR + 辅助感知)</td><td class="win">78.8%</td><td>相对 SFT +23.9%</td></tr>
</table>
<div class="callout green"><div class="ct">关键结论</div><p style="margin:0">GCR 零推理成本;EDA 在数据受限时增益显著;辅助感知监督解决小目标定位幻觉。方法在 Ovis2.5-2B / Gemma3-4B / MiniCPM-V4.5 上均验证有效。</p></div>
</div>
</details>
</article>
<!-- ═══ 京东 · 码垛 ═══ -->
<article id="pack" class="project-section" aria-labelledby="pack-title">
<div class="project-topline"><span class="project-label">02 / CoPack · 京东 X 科技部</span><time>2025.09 — 2025.11</time></div>
<div class="project-summary reverse">
<figure class="project-image photo"><img src="assets/pack_sim.png" alt="CoPack 在线三维装箱与机械臂码垛仿真场景"><figcaption>以物理稳定性约束连接仿真训练与产线码垛</figcaption></figure>
<div class="project-copy"><h2 id="pack-title">CoPack:强化学习码垛</h2><p class="project-description">以 GAT 图编码器和 ACKTR 策略处理在线三维装箱,通过静态预训练、域随机化仿真强化学习与真实轨迹微调,完成 Sim2Real 部署。</p><div class="result-line"><div class="result-item"><strong>35.2<small>% ↓</small></strong><span>物品塌落率相对降低</span></div><div class="result-item"><strong>7.31 → 4.74<small>%</small></strong><span>部署前后 ICR</span></div></div><div class="method-tags"><span>GAT / ACKTR</span><span>MuJoCo</span><span>ABB 工业机械臂</span></div></div>
</div>
<details class="project-details"><summary>深入了解 CoPack:MDP 设计、Sim2Real 流程与部署结果</summary>
<div class="card reveal">
<h3><span class="dot"></span>在线三维装箱 + 物理稳定性约束</h3>
<p>包裹逐个到来,每步决策放入笼车最优位,最大化体积利用率同时避免倒垛。难点:变长动作空间 + 真实物理稳定性 + sim2real gap。</p>
<div class="fig-grid g3">
<figure><img src="assets/pack_sim.png" alt=""><figcaption>MuJoCo 仿真:整齐码放</figcaption></figure>
<figure><img src="assets/pack_train.png" alt=""><figcaption>大规模并行训练环境</figcaption></figure>
<figure><img src="assets/pack_real.png" alt=""><figcaption>真实部署:ABB 工业机械臂产线</figcaption></figure>
</div>
</div>
<div class="card reveal">
<h3><span class="dot"></span>GAT 编码器 + ACKTR Actor-Critic</h3>
<div class="fig-grid g2" style="grid-template-columns:1.3fr 1fr;align-items:center">
<figure style="margin:0"><img src="assets/pack_gat.png" alt=""><figcaption>GAT 图编码:三类异构节点注意力聚合,学习支撑/稳定关系</figcaption></figure>
<div>
<div class="callout blue" style="margin-top:0"><div class="ct">MDP 设计</div>
<ul class="feat" style="margin:0">
<li><strong>State</strong>:已装包裹 + 候选位 + Next Item 尺寸</li>
<li><strong>Action</strong>:从 EMS 候选位选一个(+朝向)</li>
<li><strong>Reward</strong>:成功→体积;倒垛→终止清零</li>
</ul>
</div>
</div>
</div>
<div class="callout"><div class="ct">三阶段 Sim2Real</div>
<div class="flow" style="margin:6px 0 0">
<div class="fstep"><div class="ft">静态预训练</div><div class="fd">离线数据学基础策略</div></div>
<div class="conn">→</div>
<div class="fstep"><div class="ft">仿真 RL</div><div class="fd">MuJoCo + 域随机化<br>重量/重心/摩擦</div></div>
<div class="conn">→</div>
<div class="fstep hot"><div class="ft">真实微调</div><div class="fd">结合产线倒垛轨迹</div></div>
</div>
</div>
<table>
<tr><th>指标</th><th>部署前</th><th>部署后</th><th>变化</th></tr>
<tr class="total"><td>物品塌落率 ICR</td><td>7.31%</td><td class="win">4.74%</td><td class="win">↓ 35.2%</td></tr>
</table>
</div>
</details>
</article>
<!-- ═══ 无界动力 ═══ -->
<article id="vla" class="project-section" aria-labelledby="vla-title">
<div class="project-topline"><span class="project-label">03 / Cross-Embodiment VLA · 无界动力</span><time>2026.05 — 2026.07</time></div>
<div class="project-summary">
<figure class="project-image"><img src="assets/cotrain_arch.png" alt="piper 与 DOS-W1 的跨本体动作空间对齐及混合训练架构"><figcaption>UnifiedEEF:统一动作表征,连接不同机器人本体</figcaption></figure>
<div class="project-copy"><h2 id="vla-title">π0.5:跨本体 VLA 混训</h2><p class="project-description">基于 π0.5 联合训练 piper 与 DOS-W1 数据,完成 UnifiedEEF 坐标对齐、多尺度记忆及 FAST CE 消融。建立 7 任务 / 146h 的统一多任务 benchmark。</p><div class="result-line"><div class="result-item"><strong>5 <small>/ 31</small></strong><span>RoboChallenge Table30-v2 双臂</span></div><div class="result-item"><strong>93<small>%</small></strong><span>放碗泛化成功率 · 15 次真机测试</span></div></div><div class="method-tags"><span>π0.5 / UnifiedEEF</span><span>Memory / Flow Matching</span><span>排名截至 2026.07.14</span></div></div>
</div>
<details class="project-details"><summary>深入了解 VLA:跨本体混训、消融实验与 RoboChallenge 评测</summary>
<div class="card reveal">
<h3><span class="dot"></span>跨本体 Co-Train:piper + DOS-W1 混训</h3>
<p>自采 piper 数据与 DOS-W1 数据融入 π0.5 跨臂型混训,核心是 <strong>UnifiedEEF 动作空间对齐</strong>——EEF 增量从基座坐标系变换到相机坐标系,确保同一物理运动跨数据集数值接近。</p>
<figure><img src="assets/cotrain_arch.png" alt=""><figcaption>跨本体动作空间统一:Joint 14-dim → EEF rot6d,统一表征后混训</figcaption></figure>
<div class="fig-grid g2">
<div>
<table>
<tr><th>数据集</th><th>Cross-Emb</th><th>Bipiper</th><th>Multi-Joint</th></tr>
<tr><td>v0620.1</td><td>4.9</td><td>6.2</td><td>28.7</td></tr>
<tr><td>v0620.2</td><td>4.6</td><td>6.0</td><td>12.3</td></tr>
<tr><td>v0621.10*</td><td>6.3</td><td>6.8</td><td>21.8</td></tr>
<tr><td>v0622.1*</td><td>8.1</td><td>6.6</td><td>26.3</td></tr>
<tr class="total"><td>原汇总值*</td><td class="win">7.2</td><td>7.6</td><td class="down">19.2</td></tr>
</table>
<p class="table-note">开环 MSE;数据集名称后的 * 为 held-out 泛化集。末行保留原记录汇总值,统计范围待核,并非上方 4 行的算术均值。</p>
</div>
<div class="callout green" style="margin-top:0"><div class="ct">真机 · 放碗 15 次泛化</div>
<ul class="feat" style="margin:0">
<li>Multi-Task Joint:<strong>33%</strong></li>
<li>Bipiper EEF:<strong>80%</strong></li>
<li>Cross-Emb EEF:<strong class="em">93%</strong> ← UnifiedEEF 是关键</li>
</ul>
<p style="margin:10px 0 0;font-size:13px">无坐标对齐时混训性能崩溃至 0%。</p>
</div>
</div>
</div>
<div class="card reveal">
<h3><span class="dot"></span>子模块消融</h3>
<div class="fig-grid g2">
<div>
<p style="font-weight:600;color:var(--ink);margin-bottom:6px">Memory 多尺度记忆 <a href="https://arxiv.org/abs/2603.03596" target="_blank" rel="noopener noreferrer"><code>arxiv:2603.03596 ↗</code></a></p>
<p style="font-size:13px">5 帧历史 + SigLIP 每 4 层 ViT 插时间注意力。真机总成功率 <strong class="em">86.5% → 88.5%(+2.0 个百分点)</strong>,复杂任务收益明显。</p>
<img src="assets/vla_mse1.png" alt="" style="border-radius:10px;border:1px solid var(--line)">
</div>
<div>
<p style="font-weight:600;color:var(--ink);margin-bottom:6px">FAST CE vs 纯 Flow Matching <a href="https://arxiv.org/abs/2505.23705" target="_blank" rel="noopener noreferrer"><code>arxiv:2505.23705 ↗</code></a></p>
<p style="font-size:13px">KI 梯度隔离防 FM loss 污染 VLM。<strong>结论:多任务场景 FAST CE 相较纯 FM 未见整体收益</strong>,多数任务 MSE 上升——精细连续控制下 FM 更稳健。</p>
<img src="assets/vla_mse2.png" alt="" style="border-radius:10px;border:1px solid var(--line)">
</div>
</div>
</div>
<div class="card reveal">
<h3><span class="dot"></span>RoboChallenge Table30-v2 打榜</h3>
<div class="fig-grid g2">
<figure style="margin:0"><img src="assets/robo_board.png" alt=""><figcaption>排行榜(截至 2026.07.14)</figcaption></figure>
<figure style="margin:0"><img src="assets/robo_res.png" alt=""><figcaption>AVP 模型详细评测</figcaption></figure>
</div>
<div class="callout green"><div class="ct">双臂臂型排名 5 / 31 · 截至 2026.07.14</div><p style="margin:0">基于自采 piper + DOS-W1 跨本体混训,从数据扩充、训练策略、测试时增广多维度持续迭代。</p></div>
</div>
</details>
</article>
<!-- ═══ 小米 WAM ═══ -->
<article id="wam" class="project-section" aria-labelledby="wam-title">
<div class="project-topline"><span class="project-label">04 / World Action Model · 小米新业务部</span><time>2026.07 — 至今</time></div>
<div class="project-summary reverse">
<figure class="project-image"><img src="assets/wam_arch.png" alt="RGB 与光流双流 Transformer 结构及训练、推理模式示意"><figcaption>RGB + Optical Flow 双流结构与训练 / 推理模式示意</figcaption></figure>
<div class="project-copy"><h2 id="wam-title">双流 WAM 世界模型</h2><p class="project-description">基于 Wan2.2-TI2V-5B 构建 RGB 与 Optical Flow 双流 DiT,引入显式运动先验,结合文字与本体状态条件,进行动作驱动的未来场景生成。</p><div class="result-line"><div class="result-item"><strong>5 <small>/ 30</small></strong><span>WorldArena 2.0 · Track1</span></div><div class="result-item"><strong>120 <small>帧</small></strong><span>未来视觉场景预测</span></div></div><div class="method-tags"><span>Wan2.2 / DiT</span><span>RAFT Optical Flow</span><span>SeedVR2</span></div></div>
</div>
<details class="project-details"><summary>深入了解 WAM:双流模型、光流数据构建与比赛结果</summary>
<div class="card reveal">
<h3><span class="dot"></span>任务:预测未来 120 帧视觉场景</h3>
<p>输入首帧 RGB + 动作轨迹 + 文字指令,世界模型预测未来场景(PSNR / SSIM / LPIPS / FVD)。基础模型 <strong>Wan2.2-TI2V-5B</strong>,引入光流作为显式 Motion Prior。</p>
<p>原实验结构:RGB 分支 + Flow 分支,注入 Text / Proprio 条件,经过 <strong>30 层 Joint Self-Attention</strong> 生成未来帧。</p>
<figure><img src="assets/wam_arch.png" alt="双流 WAM 模型结构与训练、推理模式示意"><figcaption>双流 WAM:RGB 与 Flow 分支、条件注入及训练 / 推理模式示意</figcaption></figure>
</div>
<div class="card reveal">
<h3><span class="dot"></span>RGB–Optical Flow 双流 DiT</h3>
<div class="callout blue"><div class="ct">Flow 数据流(训练)</div>
<div class="flow" style="margin:6px 0 0">
<div class="fstep"><div class="ft">robot-only 视频</div><div class="fd">仿真单独渲染</div></div>
<div class="conn">→</div>
<div class="fstep"><div class="ft">RAFT 光流</div><div class="fd">逐帧对</div></div>
<div class="conn">→</div>
<div class="fstep"><div class="ft">HSV 编码</div><div class="fd">方向=色相<br>速度=饱和</div></div>
<div class="conn">→</div>
<div class="fstep"><div class="ft">robot mask</div><div class="fd">只留臂部运动</div></div>
<div class="conn">→</div>
<div class="fstep hot"><div class="ft">同 VAE</div><div class="fd">编码 latent</div></div>
</div>
</div>
<ul class="feat">
<li><strong>RGB 分支</strong>:处理外观语义,Cross-Attention 注入文字(UMT5)+ 14D 本体状态(→MLP→121 token)</li>
<li><strong>Flow 分支</strong>:独立 patchify + RoPE,与 RGB token 拼接后联合过 Self-Attention</li>
<li><strong>训练用 GT Flow,推理置零</strong>:靠训练学到的"动作→运动"隐式映射生成画面</li>
<li><strong>后处理</strong>:SeedVR2 超分 + 时序下采样(121→40/20 帧)提升画质</li>
</ul>
<div class="fig-grid g2" style="grid-template-columns:1fr 1.2fr;align-items:center">
<figure style="margin:0"><img src="assets/wam_result.png" alt=""><figcaption>WorldArena 2.0 Track1 排名结果</figcaption></figure>
<div class="callout green" style="margin:0"><div class="ct">WorldArena 2.0 · Track1 排名 5 / 30</div><p style="margin:0">超过 25 支参赛队伍,Track1 排名第 5。光流显式编码机器人运动场,比让 RGB 隐式学运动更高效。</p></div>
</div>
</div>
</details>
</article>
</section>
<!-- ═══ 个人项目 ═══ -->
<section id="projects" aria-labelledby="projects-title">
<div class="section-title"><div><span class="eyebrow">04 / 项目经历</span><h2 id="projects-title">项目经历</h2></div><a class="text-link" href="https://github.com/Dev-EmbodiedAI" target="_blank" rel="noopener noreferrer">访问 GitHub 主页 <span aria-hidden="true">↗</span></a></div>
<div class="open-source-grid">
<article class="open-source-card">
<span class="project-number">01 / ROBOT MANIPULATION</span>
<h3>GraspFruit · 自主水果抓取</h3>
<p style="font-size:13.5px">VLM + 分割 + 6DoF 抓取的端到端闭环系统,MuJoCo 仿真验证。</p>
<div class="flow" style="margin:10px 0">
<div class="fstep"><div class="ft">Qwen-VL</div><div class="fd">语义规划</div></div>
<div class="conn">→</div>
<div class="fstep"><div class="ft">SAM3</div><div class="fd">实例分割</div></div>
<div class="conn">→</div>
<div class="fstep"><div class="ft">GraspNet</div><div class="fd">6DoF 抓取</div></div>
<div class="conn">→</div>
<div class="fstep hot"><div class="ft">MuJoCo</div><div class="fd">执行验证</div></div>
</div>
<figure style="margin:10px 0 0"><img src="assets/grasp_demo2.gif" alt=""><figcaption>演示:苹果分拣到蓝色篮子</figcaption></figure>
<div class="pills"><span>Qwen3-VL</span><span>SAM3</span><span>GraspNet</span><span>UR5e+2F-85</span><span>ROS2</span></div>
</article>
<article class="open-source-card">
<span class="project-number">02 / VISUAL MEMORY</span>
<h3>LeMEM · VLA 记忆模块</h3>
<p style="font-size:13.5px">为 VLA 模型设计的轻量级多尺度视觉记忆模块,将短期历史观测融入策略网络,提升长时序任务的时序一致性。开源实现,含完整训练/评测框架。</p>
<figure style="margin:10px 0 0"><img src="assets/lemem_arch.png" alt=""><figcaption>LeMEM 架构:多尺度记忆注入 VLA backbone</figcaption></figure>
<div class="pills"><span>π0.5</span><span>SigLIP</span><span>Temporal-Attn</span><span>PyTorch</span></div>
</article>
</div>
<details class="project-details background-details"><summary>查看本科机器人实践与智能车竞赛现场</summary>
<div class="card reveal">
<h3><span class="dot"></span>本科代表性技术经历</h3>
<div class="fig-grid g3">
<figure style="margin:0"><img src="assets/ug_detect.png" alt=""><figcaption>目标检测:车辆/行人感知</figcaption></figure>
<figure style="margin:0"><img src="assets/ug_arm.png" alt=""><figcaption>真实双臂机械臂操作</figcaption></figure>
<figure style="margin:0"><img src="assets/ug_sim.png" alt=""><figcaption>MuJoCo 仿真操作环境</figcaption></figure>
</div>
<div class="fig-grid g2" style="margin-top:14px">
<figure style="margin:0"><img src="assets/ug_car.jpg" alt=""><figcaption>第17届全国大学生智能汽车竞赛(全国一等奖 · 百度智慧交通组)</figcaption></figure>
<div class="callout"><div class="ct">本科项目与竞赛</div><p>本科阶段参与车辆 / 行人目标检测、真实双臂机械臂操作与 MuJoCo 仿真,在全国大学生智能汽车竞赛百度智慧交通组获得全国一等奖。</p></div>
</div>
</div>
</details>
</section>
<section id="skills" aria-labelledby="skills-title">
<div class="section-title"><div><span class="eyebrow">05 / 专业技能</span><h2 id="skills-title">专业技能</h2></div></div>
<div class="skills-area">
<div class="stack">
<div class="col">
<h4 style="color:var(--brand)">具身多模态大模型</h4>
<span class="tag">Qwen3-VL</span><span class="tag">SAM3</span><span class="tag">YOLO</span><span class="tag">GraspNet</span><span class="tag">GCR Loss</span><span class="tag">EDA</span><span class="tag">SFT / LoRA</span><span class="tag">DeepSpeed ZeRO</span>
</div>
<div class="col">
<h4 style="color:var(--green)">仿真 & 强化学习</h4>
<span class="tag">MuJoCo</span><span class="tag">Isaac Sim</span><span class="tag">PyBullet</span><span class="tag">A2C / PPO / ACKTR</span><span class="tag">GAT</span><span class="tag">MoveIt / OMPL</span><span class="tag">Domain Random.</span><span class="tag">Sim2Real</span>
</div>
<div class="col">
<h4 style="color:var(--blue)">VLA & WAM 模型</h4>
<span class="tag">π0.5</span><span class="tag">Flow Matching</span><span class="tag">FAST Token</span><span class="tag">KI / Memory</span><span class="tag">Wan2.2 / DiT</span><span class="tag">RAFT 光流</span><span class="tag">SeedVR2</span><span class="tag">跨本体 Co-Train</span>
</div>
</div>
</div>
</section>
<section class="contact-section" aria-label="联系与资料"><div><span class="eyebrow" style="margin-bottom:12px">其他</span><h2>周苍田</h2><p>具身智能算法 · VLA / WAM · 多模态模型与机器人学习</p></div><div class="actions"><a class="button primary" href="output/Zhou_Cangtian_Resume.pptx" download>下载 PPT <span aria-hidden="true">↗</span></a><a class="button secondary" href="https://github.com/Dev-EmbodiedAI" target="_blank" rel="noopener noreferrer">GitHub <span aria-hidden="true">↗</span></a></div></section>
</main>
<footer>
<p>周苍田 · 哈尔滨工业大学 · 2027 届</p>
<p>具身智能研究与工程实践 <span aria-hidden="true">/</span> <button type="button" class="print-button" data-print>打印 / 保存 PDF</button></p>
</footer>
<dialog id="image-viewer" class="lightbox" aria-label="项目图片大图查看" aria-describedby="image-caption"><div class="lightbox-inner"><button class="lightbox-close" type="button" aria-label="关闭图片">×</button><img alt=""><p class="lightbox-caption" id="image-caption"></p></div></dialog>
<script src="app.js" defer></script>
</body>
</html>