PROJECT_STATUS.md

Project Status

This is the fastest way to understand the current research project state. It summarizes what has already been implemented from the public Xperience-10M sample, what the first multi-episode Qwen3-Omni and Cosmos3 diagnostic branches show, and which artifacts support the next development step.

Research Positioning

The project is a research-engineering study of Xperience-10M rather than a single demo result. It makes the public sample episode inspectable, defines embodied-AI tasks over synchronized modalities, records baseline behavior, and keeps the next multi-episode modeling stage explicit. The current evidence is useful for judging data understanding, task design, evaluation discipline, and scale-up readiness; it is not presented as final full-dataset model quality. The current no-new-episode enhancement layer records how to push the selected 128-episode setup harder before asking for more raw storage.

Area	Current state	Evidence	Research readout
Public-sample pipeline	Verified	results/episode_task_suite/summary_report.json, results/episode_task_suite/windows.csv, results/episode_task_suite/feature_manifest.json	One public Xperience-10M sample episode is converted into 5,821 frames, 1,161 aligned 20-frame windows, and an 8,546-dimensional current feature contract.
Unified 20-task suite	Verified	TASK_SUITE_20.md, docs/data/task_suite_20.json, results/episode_task_suite/, results/episode_task_suite/tier2_task_suite/	All 20 task contracts have committed minimal metrics; tasks 13-20 reuse the same 20-frame windows, 5-frame stride, chronological split, and minimal/neural head pattern. The tier2_task_suite path is historical and now stores tasks 13-20, not a separate public tier.
180-result method matrix	Verified complete	docs/data/task_method_20_result_matrix.json, TASK_METHOD_20_RESULT_MATRIX.md, docs/data/task_method_20_gap_audit.json, docs/assets/charts/unified_task_model_radar.svg	The public comparison matrix now has 9 methods x 20 tasks = 180/180 scored method-task records. Six rows are explicitly marked as compact-proxy scores where the public 128-episode export lacks the direct raw target.
Neural heads	Verified	scripts/neural_task_models.py, results/episode_task_suite/neural_mlp/	Each task also has a compact PyTorch MLP run over the same feature tensor and chronological split.
Audio contribution study	Verified	scripts/audio_ablation_and_raw_upgrade.py, results/audio_ablation/, docs/data/audio_ablation_summary.json	Audio variants are compared across the original task contracts; audio improves the primary metric on 6 of those contracts, and a 588-d audio-window representation improves over the baseline audio variant on 6 of those contracts.
Research takeaways	Verified	RESEARCH_TAKEAWAYS.md, docs/data/research_takeaways.json, scripts/build_research_takeaways.py	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.
Research roadmap	Current	RESEARCH_ROADMAP.md, docs/data/research_roadmap.json	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.
128-episode task-suite enhancement pack	Current no-new-episode plan	TASK_SUITE_ENHANCEMENT_128.md, docs/data/task_suite_enhancement_128.json, results/omni_finetune/task_suite_enhancement_128_v1_20260608/enhancement_plan.json, scripts/omni/build_task_suite_enhancement_128.py	The current 3,808-window selected split can be stressed without more episodes by exporting denser and multiscale windows. The recommended next export is multiscale_20s10_40s20_80s40, estimated at 106,095 windows from the observed frame spans; the pack also defines hierarchical action/subtask targets, raw-feature shard priorities for unsupported tasks, and Qwen/Cosmos follow-up run cards.
Foundation-model plan	Current	FOUNDATION_MODEL_PLAN.md, docs/data/foundation_model_plan.json	Qwen3-Omni remains the first structured JSON LoRA baseline; Cosmos3-Nano is verified as a future-window compatibility branch; Cosmos3-Super is represented by a base-weight Reasoner evaluation and a fine-tuned Forward-Dynamics LoRA branch. The Super LoRA target is camera-pose-conditioned future vision velocity, not supervised JSON action-token prediction. OpenVLA/openpi/GR00T remain policy candidates after robot-compatible action targets are explicit.
Cosmos3-Super action-target contract	Superseded by verified forward-dynamics LoRA	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	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 packer and contract audit are now supporting evidence for the trained forward-dynamics branch; they still do not supervise preds_action, so action-token prediction needs a separate policy or inverse-dynamics target export.
Cosmos3-Super Forward-Dynamics LoRA	Verified fine-tuned adapter branch	configs/omni_backbones/cosmos3_super_forward_dynamics.json, scripts/omni/train_cosmos3_super_forward_dynamics_lora.py, scripts/omni/eval_cosmos3_super_forward_dynamics_lora.py, results/omni_finetune/verified_public/xperience10m_cosmos3_super_forward_dynamics_lora_128ep_train1epoch_256_attn_full8gpu_20260608_eval_test_full_fsdp/verified_result_summary.json, results/omni_finetune/verified_public/xperience10m_cosmos3_super_forward_dynamics_lora_128ep_train1epoch_256_attn_full8gpu_20260608_eval_test_full_fsdp/package_audit.json	The first fine-tuned Cosmos3-Super adapter branch is locally verified as a public-safe package: 26.2M LoRA parameters, 2,848 train rows, 512 validation rows, 448 held-out test rows, validation MSE 4.0082, and test MSE 3.6853. The package excludes adapter safetensors; weights are published separately at cy0307/ropedia-cosmos3-super-forward-dynamics-lora-128ep.
Omni model extension contract	Current	OMNI_MODEL_EXTENSION_CONTRACT.md, configs/omni_backbones/, scripts/omni/backbone_registry.py, scripts/omni/smoke_test_backbone_packaging.py	Future model 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.
Xperience Embodied Foundation Model	Future goal	XPERIENCE_EMBODIED_FOUNDATION_MODEL_PRETRAINING.md	A future full-corpus pretraining plan describes target modules, objectives, staged scale-up, hardware ranges, and evaluation for a domain-specific embodied foundation model.
Evaluation protocol	Verified	EVALUATION_PROTOCOL.md, docs/data/evaluation_protocol.json, scripts/build_evaluation_protocol.py	Windowing, chronological split, per-task metrics, leakage controls, and current limitations are generated from committed metric artifacts.
Dataset context	Verified	XPERIENCE10M_DATASET_CARD_ALIGNMENT.md, official Xperience-10M and sample cards	The README and dashboard distinguish the public sample used here from the gated full dataset used for the selected multi-episode pilot.
Public dashboard and Hub pages	Verified	GitHub Pages, HF Space, artifact dataset, baseline model repo, Qwen3-Omni LoRA repo	Readers can move between the website, code, derived artifacts, baseline weights, and Qwen3-Omni pilot status without needing local infrastructure details.
Public package policy	Verified	DATA_NOTICE.md, REPRODUCIBILITY.md	Raw Xperience-10M data, private gated files, large archives, credentials, and full Qwen weights are not redistributed.
Reproducibility	Verified for the public sample	REPRODUCIBILITY.md, docs/data/reproducibility_matrix.json, notes/reproducibility_audit.md	The public sample workflow has explicit commands, expected outputs, and exact-match reproduction evidence.
128-episode aligned baselines	Verified companion result	results/omni_finetune/multi_episode_128_task_baselines/BASELINE_ALIGNMENT_REPORT.md, results/omni_finetune/a100_128_metadata_task_baselines_20260616_v2/, scripts/omni/run_128_task_baselines.py, scripts/omni/run_128_raw20_task_baselines.py	The earlier simple and neural baseline framing is aligned to the same selected 96/16/16 episode split used by the Qwen3-Omni pilot. Metadata/text simple and neural heads now have 20/20 rows, raw-feature simple and neural heads have 20/20 rows, and compact proxies remain marked for missing direct raw targets.
Qwen3-Omni fine-tuning	Latest v6 diagnostic branch verified; JSON target met	docs/data/omni_finetune_verified_result.json, docs/data/qwen3_v5_v6_comparison.json, results/omni_finetune/QWEN3_V5_V6_COMPARISON_20260614.md, results/omni_finetune/verified_public/xperience10m_qwen3_omni_128ep_multiscale_cap96_v6_rank64_lr5e5_full8gpu_lora_eval_test_full/, scripts/omni/package_verified_omni_result.py, scripts/omni/audit_verified_omni_package.py, scripts/omni/analyze_qwen3_omni_errors.py	The selected 96/16/16 episode split now has a current public-safe v6 rank64/lr5e-5 held-out package with 34,269 exported windows and 4,032 test predictions. JSON validity is 99.90%, meeting the 98% target; transition accuracy is 98.98%, contact accuracy is 81.77%, object micro-F1 is 30.65%, next-action accuracy is 4.31%, and action/subtask metrics remain weak. v6 improves action macro-F1 and contact accuracy versus v5, but v5 remains stronger on JSON validity, subtask, next-action, transition, and object metrics.
Raw Xperience-10M redistribution	Not included	DATA_NOTICE.md, docs/data/publication_audit.json	Raw MP4, HDF5, RRD files, private gated data, and full Qwen weights are intentionally excluded.

Fast Research Route

1. Read PUBLIC_READER_MAP.md if you need to choose the right public surface before opening detailed artifacts.
2. Read this status file to establish the current project scope.
3. Open the visual dashboard for the fastest overview of data, tasks, directions, and scale-up status.
4. Inspect RESEARCH_TAKEAWAYS.md and docs/data/research_takeaways.json for the generated result interpretation.
5. Inspect RESEARCH_ROADMAP.md and docs/data/research_roadmap.json for the path from public-sample task work to multi-episode modeling.
6. Inspect FOUNDATION_MODEL_PLAN.md and docs/data/foundation_model_plan.json before choosing a backbone branch.
7. 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.
8. Inspect XPERIENCE_EMBODIED_FOUNDATION_MODEL_PRETRAINING.md for the long-term full-corpus pretraining goal.
9. Inspect TASK_SUITE_20.md, docs/data/task_suite_20.json, docs/data/summary_metrics.json, and results/episode_task_suite/neural_mlp/ to check the unified 20-task outputs.
10. Inspect docs/data/tier2_task_suite.json and results/episode_task_suite/tier2_task_suite/TIER2_TASK_BASELINES.md only as the historical artifact path for tasks 13-20.
11. Inspect results/audio_ablation/AUDIO_ABLATION_SUMMARY.md before judging whether audio helps the current task suite.
12. Inspect EVALUATION_PROTOCOL.md before judging task metrics or leakage controls.
13. Inspect XPERIENCE10M_DATASET_CARD_ALIGNMENT.md only if you need the detailed upstream dataset-card context.
14. Inspect results/omni_finetune/multi_episode_128_task_baselines/BASELINE_ALIGNMENT_REPORT.md before comparing simple/NN baselines to the selected 128-episode setup.
15. Inspect TASK_SUITE_ENHANCEMENT_128.md and docs/data/task_suite_enhancement_128.json before deciding whether more episodes are needed; the current recommended no-new-episode export is multiscale_20s10_40s20_80s40.
16. Inspect docs/data/omni_finetune_verified_result.json before judging the Qwen3-Omni diagnostic pilot.

Current Reading Notes

• Cross-episode generalization is a later multi-episode evaluation target; the current results use one public sample episode.
• Public-facing fine-tuning results should come from the verified result package, not from live process logs or setup-only artifacts.
• The latest Qwen3-Omni v6 held-out package verifies the current dense multiscale branch and meets the strict-JSON target, but not strong action/subtask model quality: JSON validity is 99.90%, action macro-F1 is 0.0029, and subtask accuracy is 0.0037. v5 remains the pinned prior release row because it is still stronger on several metrics.
• The current 128-episode task suite can be pushed further without more raw episodes by using dense/multiscale windows, hierarchical action/subtask targets, stronger label-normalized scoring, and compact raw-feature shards for the tasks that are currently unsupported by metadata/text baselines.
• 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 current reconstruction task reconstructs feature vectors, not pixel depth, meshes, NeRF outputs, or Gaussian splats.
• Audio is part of the current 8,546-dimensional baseline feature vector.
• Audio contribution is evaluated across the original task contracts in results/audio_ablation/.
• Foundation-model selection is now explicit: Qwen3-Omni is the immediate trainable pilot, Cosmos 3 is the first world-model branch, and Cosmos3-Super has a camera-pose proxy forward-dynamics contract ready for trainer implementation; policy models such as OpenVLA/openpi/GR00T still wait for robot-compatible action-target conversion.
• Future model branches should be added through the backbone registry and verified package contract, not by creating 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.