File size: 10,312 Bytes
fe4bbfa
 
 
 
 
596ac86
16a39bb
05637a9
01f57c3
16a39bb
d9e465e
 
596ac86
16a39bb
01f57c3
16a39bb
05637a9
fe4bbfa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
596ac86
05637a9
16a39bb
596ac86
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
fe4bbfa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
596ac86
05637a9
16a39bb
596ac86
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
fe4bbfa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
596ac86
05637a9
16a39bb
596ac86
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
fe4bbfa
 
 
 
 
 
 
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
{
  "title": "Three Foundation Pipeline Tracks",
  "status": "pipeline_plan",
  "source_document": "THREE_FOUNDATION_PIPELINES.md",
  "claim_boundary": "These are supported pipeline directions, not three completed model-quality claims.",
  "diagram_assets": {
    "status": "published_high_resolution_slide_diagrams",
    "asset_root": "docs/assets/foundation-pipelines",
    "source": "Clean direction-slide PNGs supplied for the three public direction figures, with original presentation photos retained as provenance",
    "source_slide_root": "docs/assets/foundation-pipelines/source-slides",
    "source_photo_root": "docs/assets/foundation-pipelines/source-photos",
    "provenance_file": "docs/assets/foundation-pipelines/prompts.md",
    "renderer_script": "scripts/render_foundation_pipeline_diagrams.py",
    "diagram_type": "direction_slide_diagram",
    "source_update": "2026-06-19: clean Spatial intelligence, Human-video world model, and Vision-language-action PNGs are committed as source-slide assets and published as 2560-pixel public images.",
    "note": "Images are slide-diagram communication assets for pipeline tracks. Technical claims remain governed by the Markdown/JSON contracts and verified metrics."
  },
  "shared_principles": [
    "Use episode-level train/validation/test separation.",
    "Build manifest-first exporters before training.",
    "Keep target-side future labels and captions out of inputs unless the task explicitly queries them.",
    "Report task-specific metrics and saved predictions before updating public cards.",
    "Exclude raw private data and heavyweight base model weights from public packages."
  ],
  "tracks": [
    {
      "id": "spatial_intelligence",
      "title": "Spatial intelligence models",
      "question": "Can the model recover and reason over space from video?",
      "core_inputs": [
        "multiview RGB",
        "egocentric video",
        "depth",
        "camera pose",
        "calibration",
        "object cues",
        "language questions"
      ],
      "intermediate_artifacts": [
        "synchronized camera window manifest",
        "pose and depth availability report",
        "scene and object memory records",
        "object permanence targets",
        "spatial relation targets",
        "spatial QA prompts"
      ],
      "outputs": [
        "object count",
        "object persistence",
        "relative location",
        "3D geometry consistency",
        "multiview retrieval",
        "camera-motion-aware scene memory",
        "language answers grounded in the scene"
      ],
      "first_pipeline": "Build a spatial-memory exporter, start with metric depth and pose consistency tasks, then evaluate spatial QA, object permanence, counting, retrieval, and pose-aware consistency.",
      "current_maturity": "Ready as a pipeline and evaluation contract.",
      "next_gate": "Raw depth and pose artifacts plus held-out multi-episode spatial metrics.",
      "diagram_image": "docs/assets/foundation-pipelines/spatial-intelligence-pipeline.png",
      "website_image": "assets/foundation-pipelines/spatial-intelligence-pipeline.png",
      "image_alt": "High-resolution slide diagram showing the Spatial intelligence models direction for Xperience-10M.",
      "diagram_flow": [
        {
          "stage": "inputs",
          "items": [
            "multiview RGB plus egocentric video",
            "metric depth and confidence",
            "camera pose, calibration, SLAM",
            "object, contact, and language cues"
          ]
        },
        {
          "stage": "tasks_targets",
          "items": [
            "spatial QA and object count",
            "object permanence across windows",
            "relative location and retrieval",
            "pose-aware 3D consistency"
          ]
        },
        {
          "stage": "train_models",
          "items": [
            "export scene/object memory records",
            "train spatial-memory encoder",
            "add geometry-aware QA and retrieval heads",
            "keep episode-level split discipline"
          ]
        },
        {
          "stage": "evaluate_gates",
          "items": [
            "held-out episode spatial metrics",
            "count and relation accuracy",
            "retrieval rank and consistency",
            "saved predictions before public claim"
          ]
        }
      ],
      "avoid_claiming_now": [
        "full neural rendering",
        "full 3D reconstruction",
        "general spatial intelligence without artifact-level evidence"
      ]
    },
    {
      "id": "human_video_world_models",
      "title": "Human-video world models",
      "question": "Can the model predict what happens next?",
      "core_inputs": [
        "observed video windows",
        "audio",
        "sensor windows",
        "hand and body motion",
        "object and contact state",
        "action and subtask labels",
        "future windows"
      ],
      "intermediate_artifacts": [
        "observed and future window pairs",
        "future label targets",
        "action-conditioned target records",
        "visual or latent reconstruction targets",
        "temporal consistency metadata"
      ],
      "outputs": [
        "next action",
        "next subtask",
        "future object set",
        "future state embedding",
        "camera-motion delta",
        "contact transition",
        "future-window quality metrics"
      ],
      "first_pipeline": "Keep Qwen-style structured future probes for task interpretability, keep Cosmos-style dynamics branches separate, and add latent or feature-reconstruction metrics before claiming world-model quality.",
      "current_maturity": "Partially evidenced by current future-task probes and Cosmos-style branch artifacts.",
      "next_gate": "Stronger future-state metrics, qualitative future examples, and held-out episode breakdowns.",
      "diagram_image": "docs/assets/foundation-pipelines/human-video-world-model-pipeline.png",
      "website_image": "assets/foundation-pipelines/human-video-world-model-pipeline.png",
      "image_alt": "High-resolution slide diagram showing the Human-video world models direction for Xperience-10M.",
      "diagram_flow": [
        {
          "stage": "inputs",
          "items": [
            "observed video/audio/sensor window",
            "hand/body motion and camera pose",
            "object/contact state",
            "action and subtask labels"
          ]
        },
        {
          "stage": "tasks_targets",
          "items": [
            "next action and next subtask",
            "future object set",
            "contact transition",
            "camera-motion delta or latent future"
          ]
        },
        {
          "stage": "train_models",
          "items": [
            "Qwen structured future probes",
            "Cosmos/dynamics branch separately",
            "latent rollout or reconstruction loss",
            "no target-side future leakage"
          ]
        },
        {
          "stage": "evaluate_gates",
          "items": [
            "held-out future-task metrics",
            "contact and object-set F1",
            "rollout or latent consistency",
            "per-episode breakdown and examples"
          ]
        }
      ],
      "avoid_claiming_now": [
        "strong world model from structured future-task scores alone",
        "visual future quality without visual or latent future metrics"
      ]
    },
    {
      "id": "vision_language_action",
      "title": "Vision-language-action models",
      "question": "Can the model turn what it sees and reads into action?",
      "core_inputs": [
        "egocentric video",
        "language captions",
        "hand and body motion",
        "contacts",
        "objects",
        "procedure and subtask labels"
      ],
      "intermediate_artifacts": [
        "action-token vocabulary",
        "action-chunk windows",
        "normalization stats",
        "retargeting report",
        "leakage audit",
        "action-space model card"
      ],
      "outputs": [
        "next action",
        "action chunk",
        "object-conditioned action",
        "contact state",
        "subtask transition",
        "policy or VLA held-out metrics"
      ],
      "first_pipeline": "Define the action space, use existing 20-task next-action/contact/object-conditioned tasks first, then add hand-trajectory or policy-compatible action chunks after conversion is traceable.",
      "current_maturity": "Feasible but gated by action-target conversion.",
      "next_gate": "Traceable action tokens, normalization, retargeting metadata, and held-out policy metrics.",
      "diagram_image": "docs/assets/foundation-pipelines/vision-language-action-pipeline.png",
      "website_image": "assets/foundation-pipelines/vision-language-action-pipeline.png",
      "image_alt": "High-resolution slide diagram showing the Vision-language-action models direction for Xperience-10M.",
      "diagram_flow": [
        {
          "stage": "inputs",
          "items": [
            "egocentric video and captions",
            "objects, contacts, and procedures",
            "hand/body motion windows",
            "subtask labels and language context"
          ]
        },
        {
          "stage": "tasks_targets",
          "items": [
            "action-token vocabulary",
            "next action and action chunks",
            "object-conditioned actions",
            "contact state and subtask transition"
          ]
        },
        {
          "stage": "train_models",
          "items": [
            "build action-space converter",
            "normalize and audit action chunks",
            "train VLA/policy-compatible head",
            "track leakage and retargeting reports"
          ]
        },
        {
          "stage": "evaluate_gates",
          "items": [
            "held-out action metrics",
            "chunk and next-action accuracy",
            "object/contact-conditioned scores",
            "policy card before robot-policy claim"
          ]
        }
      ],
      "avoid_claiming_now": [
        "robot policy quality",
        "policy generalization before action-space evidence exists"
      ]
    }
  ]
}