cy0307's picture
Add files using upload-large-folder tool
146ae33 verified
|
Raw
History Blame
16 kB

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 selected-128 Qwen3-Omni and Cosmos3 diagnostic rows 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 selected-128 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 plan records how to push the selected 128-episode setup harder before asking for more raw storage.

Two Evidence Lines

Evidence line Data unit Public result status Correct reading
Line 1: 1 sample episode One public sample episode: 5,821 frames and 1,161 aligned 20-frame windows 2 methods x 20 tasks = 40/40 direct scores Task construction, raw-sample inspection, reproducibility, and controlled single-episode baselines
Line 2: 128 selected episodes Selected 96/16/16 split over 128 official episode paths, with public-safe processed features 7 methods x 20 tasks = 140/140 scores: 134 direct + 6 compact-proxy Same-split metadata/raw baselines, Qwen3-Omni v6 diagnostics, Cosmos3 diagnostics, and scale-up planning

Qwen3-Omni v1-v6 are run versions inside Line 2, not six project evidence lines. v1-v4 are pipeline-hardening and ablation evidence, v5 is the pinned prior multiscale release, and v6 is the current public 20-task Qwen3-Omni row.

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 and share the same 20-frame windows, 5-frame stride, chronological split, and minimal/neural head pattern. The tier2_task_suite path is historical provenance inside the unified suite, 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 walkthrough-backed 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 tracks, 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 Qwen3-Omni/Cosmos3 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 diagnostic; Cosmos3-Super is represented by a base-weight Reasoner evaluation and a fine-tuned Forward-Dynamics LoRA adapter. 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 adapter; 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 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 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 rows/adapters 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 row verified; JSON target met QWEN3_OMNI_RUN_LINEAGE.md, docs/data/qwen3_omni_run_lineage.json, 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 Qwen3-Omni v1-v6 are one selected-128 run lineage: v1-v4 harden the pipeline and evaluate ablations, v5 is the pinned prior multiscale release, and v6 is the current public 20-task row. The v6 rank64/lr5e-5 package has 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 track.
  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 row.
  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 historical provenance paths inside the unified 20-task suite.
  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

  • Line 1 uses one public sample episode. Line 2 uses selected-128 public-safe processed artifacts linked to official gated episode paths. Do not merge those claims when reading scores.
  • 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 run 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 walkthrough-backed 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 track, 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 rows/adapters 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.