Hugging Face's logo

robotflowlabs
/
FORGE-Nano-Benchmark

Robotics
English
forge
vla
knowledge-distillation
model-compression
edge-deployment
action-chunking
multi-teacher
Model card Files
xet
 Community
ilessio-aiflowlab commited on
Commit
4184cbe
·
verified ·
1 Parent(s): 4159ff3

Upload FORGE_HERO.html with huggingface_hub

Browse files
Files changed (1) hide show
  1. FORGE_HERO.html +624 -0
FORGE_HERO.html ADDED
@@ -0,0 +1,624 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ <!DOCTYPE html>
2
+ <html lang="en">
3
+ <head>
4
+ <meta charset="UTF-8">
5
+ <meta name="viewport" content="width=device-width, initial-scale=1.0">
6
+ <title>FORGE // Edge Robotics AI</title>
7
+ <link href="https://fonts.googleapis.com/css2?family=Oswald:wght@300;400;700&family=JetBrains+Mono:wght@400;700&display=swap" rel="stylesheet">
8
+ <style>
9
+ :root {
10
+ --orange: #FF3B00;
11
+ --black: #050505;
12
+ --offwhite: #f3f3f3;
13
+ --dark: #0A0A0A;
14
+ --mid: #1A1A1A;
15
+ --green: #00CC66;
16
+ --blue: #0066FF;
17
+ }
18
+ * { margin: 0; padding: 0; box-sizing: border-box; }
19
+ body { font-family: 'JetBrains Mono', monospace; background: var(--black); color: var(--offwhite); overflow-x: hidden; }
20
+ h1,h2,h3,.label { font-family: 'Oswald', sans-serif; text-transform: uppercase; font-weight: 700; }
21
+
22
+ /* HERO */
23
+ .hero {
24
+ min-height: 100vh;
25
+ display: flex;
26
+ flex-direction: column;
27
+ justify-content: center;
28
+ align-items: center;
29
+ text-align: center;
30
+ position: relative;
31
+ padding: 2rem;
32
+ }
33
+ .hero::before {
34
+ content: '';
35
+ position: absolute;
36
+ top: 0; left: 0; right: 0; bottom: 0;
37
+ background:
38
+ repeating-linear-gradient(0deg, transparent, transparent 49px, rgba(255,59,0,0.03) 50px),
39
+ repeating-linear-gradient(90deg, transparent, transparent 49px, rgba(255,59,0,0.03) 50px);
40
+ pointer-events: none;
41
+ }
42
+ .hero::after {
43
+ content: '';
44
+ position: absolute;
45
+ top: 50%; left: 50%;
46
+ width: 800px; height: 800px;
47
+ transform: translate(-50%, -50%);
48
+ background: radial-gradient(circle, rgba(255,59,0,0.06) 0%, transparent 70%);
49
+ pointer-events: none;
50
+ }
51
+ .hero-badge {
52
+ font-family: 'Oswald', sans-serif;
53
+ font-weight: 300;
54
+ font-size: 0.85rem;
55
+ letter-spacing: 8px;
56
+ color: var(--orange);
57
+ text-transform: uppercase;
58
+ border: 1px solid rgba(255,59,0,0.3);
59
+ padding: 0.5rem 2rem;
60
+ margin-bottom: 2rem;
61
+ position: relative;
62
+ z-index: 1;
63
+ }
64
+ .hero h1 {
65
+ font-size: 8rem;
66
+ letter-spacing: 20px;
67
+ line-height: 0.9;
68
+ position: relative;
69
+ z-index: 1;
70
+ }
71
+ .hero h1 span { color: var(--orange); }
72
+ .hero-tagline {
73
+ font-family: 'Oswald', sans-serif;
74
+ font-weight: 300;
75
+ font-size: 1.6rem;
76
+ letter-spacing: 6px;
77
+ color: #666;
78
+ margin-top: 1.5rem;
79
+ position: relative;
80
+ z-index: 1;
81
+ }
82
+ .hero-equation {
83
+ font-size: 1.1rem;
84
+ color: #555;
85
+ margin-top: 2.5rem;
86
+ position: relative;
87
+ z-index: 1;
88
+ }
89
+ .hero-equation span { color: var(--orange); font-weight: 700; }
90
+ .scroll-cue {
91
+ position: absolute;
92
+ bottom: 2rem;
93
+ font-size: 0.7rem;
94
+ letter-spacing: 4px;
95
+ color: #333;
96
+ text-transform: uppercase;
97
+ animation: pulse 2s ease-in-out infinite;
98
+ }
99
+ @keyframes pulse { 0%,100% { opacity: 0.3; } 50% { opacity: 1; } }
100
+
101
+ /* IMPACT BAR */
102
+ .impact {
103
+ display: grid;
104
+ grid-template-columns: repeat(4, 1fr);
105
+ border-top: 6px solid var(--orange);
106
+ border-bottom: 6px solid var(--orange);
107
+ }
108
+ .impact-item {
109
+ text-align: center;
110
+ padding: 3rem 1rem;
111
+ border-right: 1px solid #222;
112
+ position: relative;
113
+ }
114
+ .impact-item:last-child { border-right: none; }
115
+ .impact-num {
116
+ font-family: 'Oswald', sans-serif;
117
+ font-size: 4.5rem;
118
+ font-weight: 700;
119
+ color: var(--orange);
120
+ line-height: 1;
121
+ }
122
+ .impact-unit {
123
+ font-family: 'Oswald', sans-serif;
124
+ font-size: 1.5rem;
125
+ font-weight: 300;
126
+ color: var(--orange);
127
+ }
128
+ .impact-label {
129
+ font-size: 0.65rem;
130
+ letter-spacing: 3px;
131
+ color: #666;
132
+ text-transform: uppercase;
133
+ margin-top: 0.5rem;
134
+ }
135
+
136
+ /* SECTION */
137
+ .section {
138
+ padding: 5rem 3rem;
139
+ max-width: 1200px;
140
+ margin: 0 auto;
141
+ }
142
+ .section-label {
143
+ font-family: 'Oswald', sans-serif;
144
+ font-weight: 300;
145
+ font-size: 0.75rem;
146
+ letter-spacing: 6px;
147
+ color: var(--orange);
148
+ text-transform: uppercase;
149
+ margin-bottom: 0.5rem;
150
+ }
151
+ .section-title {
152
+ font-size: 2.5rem;
153
+ letter-spacing: 6px;
154
+ color: var(--offwhite);
155
+ margin-bottom: 2rem;
156
+ }
157
+
158
+ /* ARCHITECTURE */
159
+ .arch-flow {
160
+ display: flex;
161
+ align-items: stretch;
162
+ gap: 0;
163
+ margin: 3rem 0;
164
+ }
165
+ .arch-block {
166
+ flex: 1;
167
+ background: var(--mid);
168
+ border: 2px solid #2A2A2A;
169
+ padding: 2rem 1.5rem;
170
+ text-align: center;
171
+ position: relative;
172
+ }
173
+ .arch-block.active { border-color: var(--orange); }
174
+ .arch-block .block-icon {
175
+ font-size: 2rem;
176
+ margin-bottom: 0.5rem;
177
+ }
178
+ .arch-block .block-name {
179
+ font-family: 'Oswald', sans-serif;
180
+ font-size: 1rem;
181
+ letter-spacing: 2px;
182
+ color: var(--orange);
183
+ text-transform: uppercase;
184
+ }
185
+ .arch-block .block-detail {
186
+ font-size: 0.7rem;
187
+ color: #888;
188
+ margin-top: 0.5rem;
189
+ }
190
+ .arch-block .block-params {
191
+ font-family: 'Oswald', sans-serif;
192
+ font-size: 1.5rem;
193
+ color: var(--offwhite);
194
+ margin-top: 0.8rem;
195
+ }
196
+ .arch-arrow {
197
+ display: flex;
198
+ align-items: center;
199
+ font-size: 1.5rem;
200
+ color: var(--orange);
201
+ padding: 0 0.3rem;
202
+ }
203
+
204
+ /* VS SECTION */
205
+ .vs-container {
206
+ display: grid;
207
+ grid-template-columns: 1fr 80px 1fr;
208
+ gap: 0;
209
+ align-items: stretch;
210
+ margin: 3rem 0;
211
+ }
212
+ .vs-card {
213
+ background: var(--mid);
214
+ border: 2px solid #2A2A2A;
215
+ padding: 3rem 2rem;
216
+ }
217
+ .vs-card.winner { border-color: var(--green); }
218
+ .vs-card .card-label {
219
+ font-family: 'Oswald', sans-serif;
220
+ font-size: 0.7rem;
221
+ letter-spacing: 4px;
222
+ text-transform: uppercase;
223
+ margin-bottom: 1.5rem;
224
+ }
225
+ .vs-card .card-label.old { color: #555; }
226
+ .vs-card .card-label.new { color: var(--green); }
227
+ .vs-metric {
228
+ display: flex;
229
+ justify-content: space-between;
230
+ align-items: baseline;
231
+ padding: 0.8rem 0;
232
+ border-bottom: 1px solid #222;
233
+ }
234
+ .vs-metric .metric-name { font-size: 0.75rem; color: #888; }
235
+ .vs-metric .metric-val { font-family: 'Oswald', sans-serif; font-size: 1.4rem; }
236
+ .vs-metric .metric-val.dim { color: #555; }
237
+ .vs-metric .metric-val.bright { color: var(--green); }
238
+ .vs-divider {
239
+ display: flex;
240
+ align-items: center;
241
+ justify-content: center;
242
+ font-family: 'Oswald', sans-serif;
243
+ font-size: 2rem;
244
+ color: var(--orange);
245
+ background: var(--dark);
246
+ border-top: 2px solid #2A2A2A;
247
+ border-bottom: 2px solid #2A2A2A;
248
+ }
249
+
250
+ /* RESULTS GRID */
251
+ .results-grid {
252
+ display: grid;
253
+ grid-template-columns: repeat(3, 1fr);
254
+ gap: 2rem;
255
+ margin: 3rem 0;
256
+ }
257
+ .result-card {
258
+ background: var(--mid);
259
+ border: 2px solid #2A2A2A;
260
+ padding: 2rem;
261
+ border-left: 4px solid var(--orange);
262
+ }
263
+ .result-card .rc-title {
264
+ font-family: 'Oswald', sans-serif;
265
+ font-size: 0.8rem;
266
+ letter-spacing: 3px;
267
+ color: var(--orange);
268
+ text-transform: uppercase;
269
+ margin-bottom: 1rem;
270
+ }
271
+ .result-card .rc-num {
272
+ font-family: 'Oswald', sans-serif;
273
+ font-size: 3rem;
274
+ font-weight: 700;
275
+ color: var(--offwhite);
276
+ line-height: 1;
277
+ }
278
+ .result-card .rc-sub {
279
+ font-size: 0.7rem;
280
+ color: #666;
281
+ margin-top: 0.5rem;
282
+ }
283
+
284
+ /* CONFIG TABLE */
285
+ .config-box {
286
+ background: var(--mid);
287
+ border: 2px solid var(--orange);
288
+ padding: 2rem;
289
+ margin: 3rem 0;
290
+ display: grid;
291
+ grid-template-columns: 1fr 1fr;
292
+ gap: 2rem;
293
+ }
294
+ .config-box .config-header {
295
+ grid-column: 1 / -1;
296
+ font-family: 'Oswald', sans-serif;
297
+ font-size: 1.2rem;
298
+ letter-spacing: 4px;
299
+ color: var(--orange);
300
+ }
301
+ .config-row {
302
+ display: flex;
303
+ justify-content: space-between;
304
+ padding: 0.5rem 0;
305
+ border-bottom: 1px solid #222;
306
+ font-size: 0.8rem;
307
+ }
308
+ .config-row .ck { color: #888; }
309
+ .config-row .cv { color: var(--offwhite); font-weight: 700; }
310
+
311
+ /* TEACHERS */
312
+ .teacher-grid {
313
+ display: grid;
314
+ grid-template-columns: repeat(5, 1fr);
315
+ gap: 1rem;
316
+ margin: 2rem 0;
317
+ }
318
+ .teacher-card {
319
+ background: var(--mid);
320
+ border: 1px solid #2A2A2A;
321
+ padding: 1.5rem 1rem;
322
+ text-align: center;
323
+ }
324
+ .teacher-card .tc-name {
325
+ font-family: 'Oswald', sans-serif;
326
+ font-size: 0.85rem;
327
+ color: var(--orange);
328
+ letter-spacing: 1px;
329
+ }
330
+ .teacher-card .tc-params {
331
+ font-family: 'Oswald', sans-serif;
332
+ font-size: 1.5rem;
333
+ color: var(--offwhite);
334
+ margin: 0.5rem 0;
335
+ }
336
+ .teacher-card .tc-type { font-size: 0.65rem; color: #666; }
337
+
338
+ /* CTA */
339
+ .cta {
340
+ text-align: center;
341
+ padding: 5rem 2rem;
342
+ border-top: 6px solid var(--orange);
343
+ background: linear-gradient(180deg, var(--mid) 0%, var(--black) 100%);
344
+ }
345
+ .cta h2 {
346
+ font-size: 2.5rem;
347
+ letter-spacing: 8px;
348
+ margin-bottom: 1rem;
349
+ }
350
+ .cta h2 span { color: var(--orange); }
351
+ .cta-sub {
352
+ font-size: 0.85rem;
353
+ color: #666;
354
+ max-width: 600px;
355
+ margin: 0 auto 2rem;
356
+ }
357
+ .cta-links {
358
+ display: flex;
359
+ justify-content: center;
360
+ gap: 2rem;
361
+ }
362
+ .cta-link {
363
+ font-family: 'Oswald', sans-serif;
364
+ font-size: 0.9rem;
365
+ letter-spacing: 3px;
366
+ text-transform: uppercase;
367
+ color: var(--offwhite);
368
+ text-decoration: none;
369
+ padding: 1rem 2.5rem;
370
+ border: 2px solid var(--orange);
371
+ transition: all 0.2s;
372
+ }
373
+ .cta-link:hover { background: var(--orange); color: var(--black); }
374
+ .cta-link.primary { background: var(--orange); color: var(--black); }
375
+ .cta-link.primary:hover { background: var(--offwhite); border-color: var(--offwhite); }
376
+
377
+ /* FOOTER */
378
+ .footer {
379
+ text-align: center;
380
+ padding: 2rem;
381
+ font-size: 0.65rem;
382
+ color: #333;
383
+ letter-spacing: 2px;
384
+ }
385
+ .footer span { color: var(--orange); }
386
+
387
+ @media (max-width: 900px) {
388
+ .hero h1 { font-size: 4rem; letter-spacing: 10px; }
389
+ .impact { grid-template-columns: repeat(2, 1fr); }
390
+ .arch-flow { flex-direction: column; }
391
+ .arch-arrow { transform: rotate(90deg); justify-content: center; padding: 0.5rem; }
392
+ .vs-container { grid-template-columns: 1fr; }
393
+ .results-grid { grid-template-columns: 1fr; }
394
+ .config-box { grid-template-columns: 1fr; }
395
+ .teacher-grid { grid-template-columns: repeat(2, 1fr); }
396
+ .cta-links { flex-direction: column; align-items: center; }
397
+ }
398
+ </style>
399
+ </head>
400
+ <body>
401
+
402
+ <!-- ========== HERO ========== -->
403
+ <div class="hero">
404
+ <div class="hero-badge">Robot Flow Labs // ANIMA Stack</div>
405
+ <h1><span>FOR</span>GE</h1>
406
+ <div class="hero-tagline">Fast Optimized Robot Generation Engine</div>
407
+ <div class="hero-equation">
408
+ <span>7B</span> teacher &rarr; <span>&lt;1B</span> student &rarr; <span>14.1 fps</span> on edge GPU
409
+ </div>
410
+ <div class="scroll-cue">&darr; scroll &darr;</div>
411
+ </div>
412
+
413
+ <!-- ========== IMPACT ========== -->
414
+ <div class="impact">
415
+ <div class="impact-item">
416
+ <div class="impact-num">9<span class="impact-unit">x</span></div>
417
+ <div class="impact-label">Model Compression</div>
418
+ </div>
419
+ <div class="impact-item">
420
+ <div class="impact-num">14.1</div>
421
+ <div class="impact-label">FPS on NVIDIA L4 (FP16)</div>
422
+ </div>
423
+ <div class="impact-item">
424
+ <div class="impact-num">28<span class="impact-unit">x</span></div>
425
+ <div class="impact-label">Faster Than Teacher</div>
426
+ </div>
427
+ <div class="impact-item">
428
+ <div class="impact-num">&lt;600<span class="impact-unit">MB</span></div>
429
+ <div class="impact-label">INT4 Deploy Size</div>
430
+ </div>
431
+ </div>
432
+
433
+ <!-- ========== THE PROBLEM ========== -->
434
+ <div class="section">
435
+ <div class="section-label">The Problem</div>
436
+ <h2 class="section-title">VLAs Are Too Big for Robots</h2>
437
+ <p style="color:#888;font-size:0.85rem;max-width:800px;margin-bottom:2rem;">
438
+ Vision-Language-Action models (OpenVLA, RT-2, Pi0) achieve state-of-the-art robot manipulation
439
+ — but at 7B+ parameters, they run at 0.5 fps. Robots need 10+ fps for real-time control.
440
+ FORGE solves this with automated knowledge distillation.
441
+ </p>
442
+
443
+ <div class="vs-container">
444
+ <div class="vs-card">
445
+ <div class="card-label old">Before // OpenVLA-7B</div>
446
+ <div class="vs-metric"><span class="metric-name">Parameters</span><span class="metric-val dim">7,000M</span></div>
447
+ <div class="vs-metric"><span class="metric-name">Throughput</span><span class="metric-val dim">~0.5 fps</span></div>
448
+ <div class="vs-metric"><span class="metric-name">Latency</span><span class="metric-val dim">~2,000 ms</span></div>
449
+ <div class="vs-metric"><span class="metric-name">Model Size</span><span class="metric-val dim">~13 GB</span></div>
450
+ <div class="vs-metric"><span class="metric-name">Edge Deploy</span><span class="metric-val dim">No</span></div>
451
+ </div>
452
+ <div class="vs-divider">vs</div>
453
+ <div class="vs-card winner">
454
+ <div class="card-label new">After // FORGE-Nano</div>
455
+ <div class="vs-metric"><span class="metric-name">Parameters</span><span class="metric-val bright">774M</span></div>
456
+ <div class="vs-metric"><span class="metric-name">Throughput</span><span class="metric-val bright">14.1 fps</span></div>
457
+ <div class="vs-metric"><span class="metric-name">Latency</span><span class="metric-val bright">71 ms</span></div>
458
+ <div class="vs-metric"><span class="metric-name">Model Size</span><span class="metric-val bright">&lt;600 MB</span></div>
459
+ <div class="vs-metric"><span class="metric-name">Edge Deploy</span><span class="metric-val bright">Jetson + Apple Silicon</span></div>
460
+ </div>
461
+ </div>
462
+ </div>
463
+
464
+ <!-- ========== ARCHITECTURE ========== -->
465
+ <div class="section">
466
+ <div class="section-label">Architecture</div>
467
+ <h2 class="section-title">4-Stage Distillation Pipeline</h2>
468
+ <div class="arch-flow">
469
+ <div class="arch-block active">
470
+ <div class="block-icon">&#x1F441;</div>
471
+ <div class="block-name">SigLIP-SO400M</div>
472
+ <div class="block-detail">Vision Encoder (frozen)</div>
473
+ <div class="block-params">472M</div>
474
+ </div>
475
+ <div class="arch-arrow">&rarr;</div>
476
+ <div class="arch-block active">
477
+ <div class="block-icon">&#x1F310;</div>
478
+ <div class="block-name">Bridge Attention</div>
479
+ <div class="block-detail">64 queries, 4 layers</div>
480
+ <div class="block-params">40M</div>
481
+ </div>
482
+ <div class="arch-arrow">&rarr;</div>
483
+ <div class="arch-block active">
484
+ <div class="block-icon">&#x1F9E0;</div>
485
+ <div class="block-name">Qwen2.5-0.5B</div>
486
+ <div class="block-detail">LoRA rank=64</div>
487
+ <div class="block-params">494M</div>
488
+ </div>
489
+ <div class="arch-arrow">&rarr;</div>
490
+ <div class="arch-block active">
491
+ <div class="block-icon">&#x1F3AF;</div>
492
+ <div class="block-name">Flow Head</div>
493
+ <div class="block-detail">1-step inference</div>
494
+ <div class="block-params">1.7M</div>
495
+ </div>
496
+ </div>
497
+ <div style="text-align:center;font-size:0.8rem;color:#666;">
498
+ Total: <span style="color:var(--orange);font-weight:700;">967.9M</span> params &rarr;
499
+ Pruned: <span style="color:var(--green);font-weight:700;">774.1M</span> params &rarr;
500
+ INT4: <span style="color:var(--green);font-weight:700;">&lt;600 MB</span>
501
+ </div>
502
+ </div>
503
+
504
+ <!-- ========== BENCHMARK RESULTS ========== -->
505
+ <div class="section">
506
+ <div class="section-label">GPU Benchmarks // 4x NVIDIA L4</div>
507
+ <h2 class="section-title">Measured Performance</h2>
508
+
509
+ <div class="results-grid">
510
+ <div class="result-card">
511
+ <div class="rc-title">Best Speed</div>
512
+ <div class="rc-num">14.1 fps</div>
513
+ <div class="rc-sub">Flow + LoRA-64 + 60% prune, FP16</div>
514
+ </div>
515
+ <div class="result-card">
516
+ <div class="rc-title">Best Training</div>
517
+ <div class="rc-num">92.3%</div>
518
+ <div class="rc-sub">Loss reduction in 30 steps</div>
519
+ </div>
520
+ <div class="result-card">
521
+ <div class="rc-title">Most Compressed</div>
522
+ <div class="rc-num">739M</div>
523
+ <div class="rc-sub">50% pruned, functional model</div>
524
+ </div>
525
+ <div class="result-card">
526
+ <div class="rc-title">Multi-GPU</div>
527
+ <div class="rc-num">33.6 fps</div>
528
+ <div class="rc-sub">Single GPU, batch=32, FP16</div>
529
+ </div>
530
+ <div class="result-card">
531
+ <div class="rc-title">Multi-Teacher</div>
532
+ <div class="rc-num">5</div>
533
+ <div class="rc-sub">Teachers with learned routing</div>
534
+ </div>
535
+ <div class="result-card">
536
+ <div class="rc-title">Test Suite</div>
537
+ <div class="rc-num">524</div>
538
+ <div class="rc-sub">Tests passing, 0 failures</div>
539
+ </div>
540
+ </div>
541
+
542
+ <!-- Recommended Config -->
543
+ <div class="config-box">
544
+ <div class="config-header">Recommended Production Config</div>
545
+ <div>
546
+ <div class="config-row"><span class="ck">variant</span><span class="cv">nano</span></div>
547
+ <div class="config-row"><span class="ck">action_head</span><span class="cv">flow</span></div>
548
+ <div class="config-row"><span class="ck">lora_rank</span><span class="cv">64</span></div>
549
+ <div class="config-row"><span class="ck">prune_ratio</span><span class="cv">0.60</span></div>
550
+ </div>
551
+ <div>
552
+ <div class="config-row"><span class="ck">params</span><span class="cv">774.1M</span></div>
553
+ <div class="config-row"><span class="ck">fps (FP16)</span><span class="cv">14.1</span></div>
554
+ <div class="config-row"><span class="ck">deploy_size</span><span class="cv">&lt;600 MB</span></div>
555
+ <div class="config-row"><span class="ck">target</span><span class="cv">Jetson + Apple Silicon</span></div>
556
+ </div>
557
+ </div>
558
+ </div>
559
+
560
+ <!-- ========== TEACHERS ========== -->
561
+ <div class="section">
562
+ <div class="section-label">Compatibility</div>
563
+ <h2 class="section-title">Any Teacher, Any Robot</h2>
564
+ <div class="teacher-grid">
565
+ <div class="teacher-card">
566
+ <div class="tc-name">OpenVLA-7B</div>
567
+ <div class="tc-params">7.6B</div>
568
+ <div class="tc-type">Token-AR // H=1</div>
569
+ </div>
570
+ <div class="teacher-card">
571
+ <div class="tc-name">RDT2-FM</div>
572
+ <div class="tc-params">1.2B</div>
573
+ <div class="tc-type">Diffusion // H=8</div>
574
+ </div>
575
+ <div class="teacher-card">
576
+ <div class="tc-name">SmolVLA</div>
577
+ <div class="tc-params">0.5B</div>
578
+ <div class="tc-type">Parallel // H=1</div>
579
+ </div>
580
+ <div class="teacher-card">
581
+ <div class="tc-name">BitVLA</div>
582
+ <div class="tc-params">5.9B</div>
583
+ <div class="tc-type">Quantized // H=1</div>
584
+ </div>
585
+ <div class="teacher-card">
586
+ <div class="tc-name">Pi0</div>
587
+ <div class="tc-params">3.0B</div>
588
+ <div class="tc-type">Flow // H=4</div>
589
+ </div>
590
+ </div>
591
+ <div style="text-align:center;margin-top:1rem;">
592
+ <span style="font-size:0.75rem;color:#666;">Robots: </span>
593
+ <span style="font-size:0.8rem;color:var(--orange);">Franka</span>
594
+ <span style="font-size:0.75rem;color:#444;"> // </span>
595
+ <span style="font-size:0.8rem;color:var(--orange);">ALOHA</span>
596
+ <span style="font-size:0.75rem;color:#444;"> // </span>
597
+ <span style="font-size:0.8rem;color:var(--orange);">xArm</span>
598
+ <span style="font-size:0.75rem;color:#444;"> // </span>
599
+ <span style="font-size:0.8rem;color:var(--orange);">UR5e</span>
600
+ <span style="font-size:0.75rem;color:#444;"> // </span>
601
+ <span style="font-size:0.8rem;color:#888;">+ any 6-14 DoF arm</span>
602
+ </div>
603
+ </div>
604
+
605
+ <!-- ========== CTA ========== -->
606
+ <div class="cta">
607
+ <h2><span>FORGE</span> is Open Source</h2>
608
+ <div class="cta-sub">
609
+ Part of the ANIMA agentic robotics stack. Apache 2.0 licensed.
610
+ Built for production deployment on edge hardware.
611
+ </div>
612
+ <div class="cta-links">
613
+ <a href="https://github.com/RobotFlow-Labs/anima-forge-distillation-pipeline" class="cta-link primary">GitHub</a>
614
+ <a href="https://huggingface.co/robotflowlabs/FORGE-Nano-Benchmark" class="cta-link">HuggingFace</a>
615
+ <a href="https://robotflowlabs.com" class="cta-link">Robot Flow Labs</a>
616
+ </div>
617
+ </div>
618
+
619
+ <div class="footer">
620
+ FORGE v2 // <span>All metrics measured on real NVIDIA L4 hardware</span> // March 2026 // Robot Flow Labs
621
+ </div>
622
+
623
+ </body>
624
+ </html>