File size: 18,062 Bytes
c325020
 
 
2d80be0
 
c325020
 
 
 
a8124a8
c325020
 
 
a8124a8
c325020
2bd8497
 
 
 
 
 
 
 
 
 
 
 
2d80be0
a07660e
2d80be0
 
 
 
6fdcae7
 
 
2d80be0
a07660e
 
 
 
 
 
 
c325020
 
 
 
 
 
 
 
 
 
45c1706
c325020
 
 
 
 
 
 
 
 
 
 
2bd8497
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2d80be0
2bd8497
 
 
 
 
 
 
 
3a10443
2bd8497
2bd560e
 
 
 
 
 
 
 
 
 
 
2bd8497
 
 
 
 
 
 
 
 
 
 
c325020
 
 
 
2bd8497
 
 
 
 
 
 
 
 
 
 
 
 
 
c325020
 
 
 
 
 
 
 
 
cf07180
c325020
 
 
 
 
 
04c0bde
c325020
 
 
 
 
 
 
 
 
 
 
a07660e
 
 
 
 
 
 
 
 
 
2d80be0
 
 
 
 
 
 
 
6fdcae7
2d80be0
c325020
 
2d80be0
c325020
2bd8497
eeac43c
2d80be0
2bd8497
91b502e
 
c325020
eeac43c
2d80be0
 
 
 
 
 
 
 
 
 
 
c325020
6fdcae7
 
 
 
 
 
 
 
 
 
 
3a10443
 
 
 
 
 
 
 
 
 
 
 
c325020
 
 
 
 
 
 
 
 
 
 
 
2bd8497
 
 
 
2bd560e
2bd8497
 
 
 
 
 
a07660e
6fdcae7
2bd8497
c325020
cf07180
2d80be0
6fdcae7
2d80be0
a07660e
3a10443
cf07180
45c1706
 
3a10443
2bd560e
bfcf156
c325020
 
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
{
  "title": "Ropedia Xperience-10M Task Suite Project Status",
  "version": "2026-06-01",
  "decision": "public_sample_pipeline_verified_128_aligned_baselines_qwen3_cosmos_comparison",
  "research_positioning": "A research-engineering study that makes one public Xperience-10M sample episode inspectable, defines embodied-AI tasks over synchronized modalities, records baseline behavior, aligns simple/NN baselines to the selected 128-episode split, and compares verified Qwen3-Omni and Cosmos3 branch packages as early cross-episode diagnostics.",
  "scope_boundary": {
    "validated_episode_count": 1,
    "aligned_frames": 5821,
    "sliding_windows": 1161,
    "current_feature_dimensions": 8546,
    "core_task_count": 12,
    "neural_head_count": 12,
    "direction_extension_probe_count": 4,
    "audio_featurized": true,
    "raw_xperience10m_data_redistributed": false,
    "qwen3_omni_32_episode_claim": false,
    "qwen3_omni_verified_diagnostic_pilot": true,
    "qwen3_omni_selected_episode_counts": {
      "train": 96,
      "val": 16,
      "test": 16
    },
    "qwen3_omni_exported_window_counts": {
      "train": 2848,
      "val": 512,
      "test": 448
    },
    "qwen3_omni_json_validity_rate": 0.9977678571428571,
    "qwen3_omni_validation_aware": true,
    "qwen3_omni_json_quality_target_met": true,
    "qwen3_omni_lora_adapter_repo": "https://huggingface.co/cy0307/ropedia-qwen3-omni-lora-128ep",
    "cosmos3_nano_future_window_compatibility_verified": true,
    "cosmos3_nano_future_window_test_predictions": 378,
    "cosmos3_super_reasoner_verified": true,
    "cosmos3_super_reasoner_test_predictions": 448,
    "cosmos3_super_reasoner_json_validity_rate": 0.5111607142857143,
    "omni_model_comparison_available": true,
    "multi_episode_128_aligned_baselines": true,
    "multi_episode_128_baseline_window_counts": {
      "train": 2848,
      "val": 512,
      "test": 448
    },
    "multi_episode_128_baseline_task_count": 12
  },
  "rows": [
    {
      "area": "Public-sample pipeline",
      "status": "verified",
      "evidence": [
        "results/episode_task_suite/summary_report.json",
        "results/episode_task_suite/windows.csv",
        "results/episode_task_suite/feature_manifest.json"
      ],
      "readout": "One public Xperience-10M sample episode is converted into 5,821 frames, 1,161 aligned 20-frame windows, and an 8,546-dimensional representation for repeatable task evaluation."
    },
    {
      "area": "Task suite",
      "status": "verified",
      "evidence": [
        "scripts/episode_task_suite.py",
        "results/episode_task_suite/",
        "docs/data/summary_metrics.json"
      ],
      "readout": "All 12 task contracts have committed metrics, predictions, and minimal baseline outputs."
    },
    {
      "area": "Neural heads",
      "status": "verified",
      "evidence": [
        "scripts/neural_task_models.py",
        "results/episode_task_suite/neural_mlp/"
      ],
      "readout": "Each task also has a compact PyTorch MLP run over the same feature tensor and chronological split."
    },
    {
      "area": "Audio contribution study",
      "status": "verified",
      "evidence": [
        "scripts/audio_ablation_and_raw_upgrade.py",
        "results/audio_ablation/",
        "docs/data/audio_ablation_summary.json"
      ],
      "readout": "Audio variants improve the primary metric on 6 of 12 task contracts in this single-episode setting."
    },
    {
      "area": "Evaluation protocol",
      "status": "verified",
      "evidence": [
        "EVALUATION_PROTOCOL.md",
        "docs/data/evaluation_protocol.json",
        "scripts/build_evaluation_protocol.py"
      ],
      "readout": "Windowing, chronological split, per-task metrics, leakage controls, and current limitations are generated from committed metric artifacts."
    },
    {
      "area": "Research takeaways",
      "status": "verified",
      "evidence": [
        "RESEARCH_TAKEAWAYS.md",
        "docs/data/research_takeaways.json",
        "scripts/build_research_takeaways.py"
      ],
      "readout": "The main result interpretation is generated from committed metrics: chronological class shift, neural gains on dynamics/order/alignment, open retrieval/reconstruction problems, and the need for held-out episodes."
    },
    {
      "area": "Research roadmap",
      "status": "current",
      "evidence": [
        "RESEARCH_ROADMAP.md",
        "docs/data/research_roadmap.json"
      ],
      "readout": "The roadmap connects public-sample task development to the final verified Qwen3-Omni diagnostic result, same-split baseline alignment, action/subtask error analysis, robustness runs, world/policy branches, and the future Xperience-native pretraining goal."
    },
    {
      "area": "Foundation-model plan",
      "status": "current",
      "evidence": [
        "FOUNDATION_MODEL_PLAN.md",
        "docs/data/foundation_model_plan.json"
      ],
      "readout": "Qwen3-Omni remains the first trainable held-out LoRA baseline; Cosmos 3 is now represented by a verified Cosmos3-Nano future-window compatibility package, a verified Cosmos3-Super base-weight Reasoner evaluation, and a Cosmos3-Super camera-pose proxy forward-dynamics contract audit plus schema-only packer smoke. The current target supports vision-velocity training under action conditioning, not supervised action-token prediction; OpenVLA/openpi/GR00T are policy candidates after robot-compatible action targets are explicit."
    },
    {
      "area": "Omni model extension contract",
      "status": "current",
      "evidence": [
        "OMNI_MODEL_EXTENSION_CONTRACT.md",
        "configs/omni_backbones/",
        "scripts/omni/backbone_registry.py",
        "scripts/omni/smoke_test_backbone_packaging.py"
      ],
      "readout": "Future Qwen, Cosmos-style, and VLA/policy branches must keep the same episode split discipline, held-out metrics, validation gate, public-safe package contract, and explicit forbidden-artifact policy before reporting results."
    },
    {
      "area": "Xperience Embodied Foundation Model",
      "status": "future_goal",
      "evidence": [
        "XPERIENCE_EMBODIED_FOUNDATION_MODEL_PRETRAINING.md"
      ],
      "readout": "A future full-corpus pretraining plan describes target modules, objectives, staged scale-up, hardware ranges, and evaluation for a domain-specific embodied foundation model."
    },
    {
      "area": "Official dataset wording",
      "status": "verified",
      "evidence": [
        "XPERIENCE10M_DATASET_CARD_ALIGNMENT.md",
        "docs/data/xperience10m_dataset_card_alignment.json"
      ],
      "readout": "Public wording is aligned to the official gated Xperience-10M dataset card, public sample card, and HF API metadata, including modalities, scale, access path, sample license/tooling, and current project coverage."
    },
    {
      "area": "Source alignment",
      "status": "verified",
      "evidence": [
        "SOURCE_ALIGNMENT_AUDIT.md",
        "docs/data/source_alignment_audit.json",
        "scripts/validate_source_alignment.py"
      ],
      "readout": "Source facts, sample details, API-listing notes, and project coverage are checked across repo docs, website, and HF cards."
    },
    {
      "area": "Website and HF mirrors",
      "status": "verified",
      "evidence": [
        "docs/data/website_integrity.json",
        "docs/data/mirror_parity.json",
        "docs/data/live_publication_status.json"
      ],
      "readout": "Local website links/assets pass, prepared mirrors match, and public GitHub/HF URLs have been checked after upload."
    },
    {
      "area": "Publication package",
      "status": "verified",
      "evidence": [
        "docs/data/publication_audit.json",
        "QUALITY_GATES.md",
        "docs/data/quality_gates.json"
      ],
      "readout": "Public bundles are checked for raw-data exclusion, cache exclusion, heavy-archive exclusion, credential-text checks, and current presentation assets."
    },
    {
      "area": "Reproducibility",
      "status": "verified_for_public_sample",
      "evidence": [
        "REPRODUCIBILITY.md",
        "docs/data/reproducibility_matrix.json",
        "notes/reproducibility_audit.md"
      ],
      "readout": "The public sample workflow has explicit commands, expected outputs, and exact-match reproduction evidence."
    },
    {
      "area": "128-episode aligned baselines",
      "status": "verified_companion_result",
      "evidence": [
        "results/omni_finetune/multi_episode_128_task_baselines/BASELINE_ALIGNMENT_REPORT.md",
        "results/omni_finetune/multi_episode_128_task_baselines/summary_report.json",
        "scripts/omni/run_128_task_baselines.py"
      ],
      "readout": "The earlier simple and neural baseline framing is aligned to the selected 96/16/16 episode split used by the Qwen3-Omni pilot. JSON-supported tasks have metadata/text simple and neural MLP metrics; raw-feature-only tasks are explicitly marked unsupported until 128-run sensor feature blocks are available."
    },
    {
      "area": "Current result comparison",
      "status": "verified_generated_summary",
      "evidence": [
        "docs/data/omni_model_comparison.json",
        "results/omni_finetune/OMNI_MODEL_COMPARISON.md",
        "scripts/omni/build_omni_model_comparison.py"
      ],
      "readout": "The public comparison now has two views: the three result layers and a model-family grouping. The model grouping pairs 1-episode and 128-episode entries for task-head baselines, separates Qwen3-Omni sensor-adapter smoke from 128-episode LoRA diagnostics, and separates Cosmos3-Nano future-window compatibility from Cosmos3-Super base-weight Reasoner evaluation."
    },
    {
      "area": "Qwen3-Omni fine-tuning",
      "status": "final_verified_diagnostic_result_json_target_met",
      "evidence": [
        "docs/data/omni_finetune_verified_result.json",
        "results/omni_finetune/verified_public/xperience10m_qwen3_omni_128ep_structured_json_v3_strict_label_prompt_reuse_lora_eval_test_full/",
        "https://huggingface.co/cy0307/ropedia-qwen3-omni-lora-128ep",
        "scripts/omni/package_verified_omni_result.py",
        "scripts/omni/audit_verified_omni_package.py",
        "scripts/omni/analyze_qwen3_omni_errors.py"
      ],
      "readout": "The selected 96/16/16 episode split now has a v3 strict-label public-safe held-out package with 3,808 exported windows, 512 validation windows, 448 test predictions, two training epochs reused from the same LoRA adapter, validation/audit summaries, and a public LoRA adapter repo. JSON validity is 100.00%, meeting the 98% target; transition accuracy is 97.32%, contact accuracy is 72.10%, object micro-F1 is 30.69%, and action/subtask metrics remain weak, so it is still a diagnostic baseline rather than a strong model-quality claim."
    },
    {
      "area": "Cosmos3-Nano future-window branch",
      "status": "verified_compatibility_result",
      "evidence": [
        "configs/omni_backbones/cosmos_world_model.json",
        "scripts/omni/export_cosmos3_future_window_dataset.py",
        "scripts/omni/eval_cosmos3_future_window_retrieval.py",
        "results/omni_finetune/verified_public/xperience10m_cosmos3_nano_128ep_future_window_h5_compat_adapter_eval_test_full/verified_result_summary.json"
      ],
      "readout": "The Cosmos3-Nano branch now has a public-safe verified future-window compatibility package with 3,213 future-window samples, 378 held-out test predictions, future retrieval MRR 0.0221, temporal consistency 0.0952, transition accuracy 0.9683, and contact accuracy 0.7434. It is a compatibility adapter result, not a full Cosmos diffusion-weight fine-tune."
    },
    {
      "area": "Cosmos3-Super Reasoner branch",
      "status": "verified_base_weight_result",
      "evidence": [
        "configs/omni_backbones/cosmos3_super_reasoner.json",
        "scripts/omni/eval_cosmos3_super_reasoner.py",
        "scripts/omni/run_cosmos3_super_reasoner_eval.sh",
        "results/omni_finetune/verified_public/xperience10m_cosmos3_super_reasoner_128ep_test_full_20260607/verified_result_summary.json"
      ],
      "readout": "Cosmos3-Super Reasoner now has a public-safe verified 448-window held-out evaluation on the same structured JSON task as Qwen3. It uses staged nv-community/Cosmos3-Super base weights through an 8-GPU vLLM server, not fine-tuned weights: JSON validity 0.5112, action macro-F1 0.0008, transition accuracy 0.3683, contact accuracy 0.3214, and object micro-F1 0.1370."
    },
    {
      "area": "Cosmos3-Super action-target contract",
      "status": "ready_for_forward_dynamics_trainer_implementation",
      "evidence": [
        "scripts/omni/export_cosmos3_camera_pose_targets.py",
        "scripts/omni/pack_cosmos3_super_action_batch.py",
        "results/omni_finetune/xperience10m_cosmos3_camera_pose_targets_20260608/target_manifest.json",
        "results/omni_finetune/xperience10m_cosmos3_super_training_contract_audit_camera_pose_20260608/training_contract_audit.json",
        "results/omni_finetune/xperience10m_cosmos3_super_action_packer_schema_smoke_20260608/packer_summary.json"
      ],
      "readout": "The selected 128-episode JSONL is augmented with 3,808/3,808 valid camera_pose proxy cosmos_action_target records from SLAM pose deltas. The schema-only packer smoke confirms the current forward_dynamics target should supervise noisy vision tokens under camera-pose conditioning; it does not supervise preds_action. Remaining work is a pipeline-loaded packer check, one-sample forward-dynamics overfit, and a separate policy/inverse target export before claiming action-token prediction."
    },
    {
      "area": "Raw Xperience-10M redistribution",
      "status": "not_included",
      "evidence": [
        "DATA_NOTICE.md",
        "docs/data/publication_audit.json"
      ],
      "readout": "Raw MP4, HDF5, RRD files, private gated data, and full Qwen weights are intentionally excluded."
    }
  ],
  "fast_research_route": [
    "Read PROJECT_STATUS.md and EVIDENCE_CONTRACT.md to establish what is implemented.",
    "Open docs/data/project_packet.json for the machine-readable project path.",
    "Inspect RESEARCH_TAKEAWAYS.md and docs/data/research_takeaways.json before interpreting model scores.",
    "Inspect RESEARCH_ROADMAP.md and docs/data/research_roadmap.json for the path from public-sample task work to multi-episode modeling.",
    "Inspect FOUNDATION_MODEL_PLAN.md and docs/data/foundation_model_plan.json before choosing a backbone branch.",
    "Inspect OMNI_MODEL_EXTENSION_CONTRACT.md and run python scripts/omni/backbone_registry.py --validate --json before adding a new Qwen, Cosmos-style, or VLA/policy branch.",
    "Inspect XPERIENCE_EMBODIED_FOUNDATION_MODEL_PRETRAINING.md for the long-term full-corpus pretraining goal.",
    "Inspect docs/data/summary_metrics.json and results/episode_task_suite/neural_mlp/ to check the 12-task outputs.",
    "Inspect results/audio_ablation/AUDIO_ABLATION_SUMMARY.md before judging whether audio helps the current task suite.",
    "Inspect EVALUATION_PROTOCOL.md before judging task metrics or leakage controls.",
    "Inspect SOURCE_ALIGNMENT_AUDIT.md before judging source-card consistency across public surfaces.",
    "Inspect XPERIENCE10M_DATASET_CARD_ALIGNMENT.md before judging dataset wording.",
    "Inspect results/omni_finetune/multi_episode_128_task_baselines/BASELINE_ALIGNMENT_REPORT.md before comparing simple/NN baselines to the selected 128-episode setup.",
    "Inspect docs/data/omni_model_comparison.json before comparing the current three result versions or the model-family 1-episode versus 128-episode groupings.",
    "Inspect docs/data/omni_finetune_verified_result.json before judging the Qwen3-Omni diagnostic pilot."
  ],
  "current_reading_notes": [
    "The final Qwen3-Omni diagnostic result is verified and meets the strict-JSON target, but action/subtask held-out quality is still weak.",
    "Use docs/data/omni_model_comparison.json to compare both views: the single-episode/128-baseline/model-branch result layers and the model-family grouping for task heads, Qwen3-Omni LoRA, Cosmos3-Nano, and Cosmos3-Super.",
    "Use docs/data/omni_finetune_verified_result.json and the latest verified_public final Qwen package for current held-out results.",
    "The 128-episode aligned simple/NN baselines use metadata/text features from the derived Qwen JSONL export; they align the split and task ids but do not replace raw-modality baselines for trajectory, retrieval, reconstruction, or misalignment tasks.",
    "The Cosmos3-Nano future-window branch is verified as a compatibility adapter result, Cosmos3-Super Reasoner is verified as a base-weight evaluation, and Cosmos3-Super camera-pose forward-dynamics targets now pass the contract audit plus a schema-only packer smoke; one-episode Cosmos fine-tuning and full Cosmos adapter/diffusion-weight fine-tuning remain pending, so no Cosmos weight repo should be published yet.",
    "The current reconstruction task reconstructs feature vectors, not pixel-depth, mesh, NeRF, or Gaussian reconstruction.",
    "Audio is one of the synchronized source modalities in the current task representation.",
    "The audio ablation report compares audio/no-audio variants across all 12 task contracts in results/audio_ablation/.",
    "Foundation-model selection is explicit: Qwen3-Omni is the immediate trainable pilot, Cosmos 3 is the first world-model branch, Cosmos3-Super has a camera-pose proxy forward-dynamics contract ready for trainer implementation, and policy models such as OpenVLA/openpi/GR00T wait for robot-compatible action-target conversion.",
    "Future model branches should be added through the backbone registry and verified package contract, not as one-off result folders with incompatible metrics or publication rules.",
    "The Xperience Embodied Foundation Model is a future native-pretraining goal, not a completed model or current benchmark."
  ]
}