{ "title": "Ropedia Xperience-10M Task Suite Evaluation Protocol", "status": "pass", "version": "2026-06-01", "generated_at_utc": "2026-06-06T23:26:13+00:00", "source_files": [ "docs/data/summary_metrics.json", "results/episode_task_suite/summary_report.json", "results/episode_task_suite/windows.csv", "results/episode_task_suite/feature_manifest.json" ], "scope": { "validated_episode_count": 1, "annotation": "data/sample/xperience-10m-sample/annotation.hdf5", "num_frames": 5821, "num_windows": 1161, "feature_dim": 8546, "window_frames": 20, "stride_frames": 5, "audio_featurized": true, "raw_data_redistributed": false }, "split_policy": { "name": "single_episode_chronological", "train_fraction": 0.7, "test_fraction": 0.3, "why": "The split preserves time order so future episode segments are not mixed randomly into the train set.", "limitation": "It is still one episode; cross-episode generalization is evaluated in the multi-episode stage." }, "feature_policy": { "input_contract": "8,546-dimensional current feature vector", "source_manifest": "results/episode_task_suite/feature_manifest.json", "normalization": "Scalers are fit on train windows only for the baseline heads.", "audio_status": "Audio is represented in the current feature vector." }, "baselines": [ { "name": "minimal", "heads": [ "softmax", "binary logistic", "multi-label logistic", "ridge regression", "ridge projection plus cosine ranking" ], "purpose": "Keep each task contract interpretable and easy to inspect." }, { "name": "neural_mlp", "heads": [ "PyTorch MLP classifier", "PyTorch MLP regressor", "PyTorch MLP multi-label head" ], "purpose": "Check nonlinear gains before larger omni-model fine-tuning.", "config": { "name": "neural_mlp", "type": "lightweight PyTorch MLP over shared window features", "epochs": 80, "hidden_dim": 128, "batch_size": 128, "learning_rate": 0.001, "weight_decay": 0.0001, "dropout": 0.1, "device": "auto" } } ], "task_protocols": [ { "task": "timeline_action", "task_display_name": "Action Recognition", "family": "supervised classification", "unit": "single window", "input": "current 20-frame all-feature window", "target": "current action label", "primary_metric": "macro_f1", "higher_is_better": true, "leakage_rule": "No future labels enter the input. Chronological split exposes unseen later action labels.", "counts": { "num_windows": 1144, "num_train_windows": 801, "num_test_windows": 343 }, "minimal_primary_metric": 0.05, "neural_primary_metric": 0.014814814814814814, "minimal_metric_source": "results/episode_task_suite/timeline_action/metrics.json", "neural_metric_source": "results/episode_task_suite/neural_mlp/timeline_action/metrics.json" }, { "task": "timeline_subtask", "task_display_name": "Procedure Step Recognition", "family": "supervised classification", "unit": "single window", "input": "current 20-frame all-feature window", "target": "current subtask label", "primary_metric": "macro_f1", "higher_is_better": true, "leakage_rule": "No future labels enter the input. Chronological split exposes unseen later subtask labels.", "counts": { "num_windows": 1147, "num_train_windows": 803, "num_test_windows": 344 }, "minimal_primary_metric": 0.05056355513846935, "neural_primary_metric": 0.02810810810810811, "minimal_metric_source": "results/episode_task_suite/timeline_subtask/metrics.json", "neural_metric_source": "results/episode_task_suite/neural_mlp/timeline_subtask/metrics.json" }, { "task": "transition_detection", "task_display_name": "Action Boundary Detection", "family": "temporal diagnostic", "unit": "single window", "input": "current 20-frame all-feature window", "target": "action boundary versus steady", "primary_metric": "macro_f1", "higher_is_better": true, "leakage_rule": "Boundary labels are targets only. Boundary timing is evaluated after prediction.", "counts": { "num_windows": 1161, "num_train_windows": 813, "num_test_windows": 348 }, "minimal_primary_metric": 0.6118237590630229, "neural_primary_metric": 0.5862068965517241, "minimal_metric_source": "results/episode_task_suite/transition_detection/metrics.json", "neural_metric_source": "results/episode_task_suite/neural_mlp/transition_detection/metrics.json" }, { "task": "next_action", "task_display_name": "Next-Action Prediction", "family": "short-horizon prediction", "unit": "single window", "input": "current 20-frame all-feature window at time t", "target": "action label at t + 20 frames", "primary_metric": "macro_f1", "higher_is_better": true, "leakage_rule": "Future labels are shifted into targets only; model inputs remain current-window features.", "counts": { "num_windows": 1161, "num_train_windows": 813, "num_test_windows": 348 }, "minimal_primary_metric": 0.05925925925925927, "neural_primary_metric": 0.04186046511627907, "minimal_metric_source": "results/episode_task_suite/next_action/metrics.json", "neural_metric_source": "results/episode_task_suite/neural_mlp/next_action/metrics.json" }, { "task": "hand_trajectory_forecast", "task_display_name": "Hand Trajectory Forecasting", "family": "trajectory regression", "unit": "single window", "input": "current all-feature window", "target": "future left/right hand 3D joints for 10 frames", "primary_metric": "mpjpe", "higher_is_better": false, "leakage_rule": "Future mocap coordinates are targets only, not inputs.", "counts": { "num_windows": 1159, "num_train_windows": 811, "num_test_windows": 348 }, "minimal_primary_metric": 0.8646570444107056, "neural_primary_metric": 0.10785018652677536, "minimal_metric_source": "results/episode_task_suite/hand_trajectory_forecast/metrics.json", "neural_metric_source": "results/episode_task_suite/neural_mlp/hand_trajectory_forecast/metrics.json" }, { "task": "contact_prediction", "task_display_name": "Contact State Prediction", "family": "binary classification", "unit": "single window", "input": "non-contact and non-caption feature blocks", "target": "any body contact", "primary_metric": "macro_f1", "higher_is_better": true, "leakage_rule": "Contact-derived fields and caption labels are excluded from inputs.", "counts": { "num_windows": 1161, "num_train_windows": 813, "num_test_windows": 348 }, "minimal_primary_metric": 1.0, "neural_primary_metric": 1.0, "minimal_metric_source": "results/episode_task_suite/contact_prediction/metrics.json", "neural_metric_source": "results/episode_task_suite/neural_mlp/contact_prediction/metrics.json" }, { "task": "object_relevance", "task_display_name": "Object Relevance Prediction", "family": "multi-label classification", "unit": "single window", "input": "non-caption feature blocks", "target": "current relevant object set", "primary_metric": "micro_f1", "higher_is_better": true, "leakage_rule": "Caption/object-label fields are excluded from inputs.", "counts": { "num_windows": 1161, "num_train_windows": 813, "num_test_windows": 348 }, "minimal_primary_metric": 0.18034382095361662, "neural_primary_metric": 0.1679279279279279, "minimal_metric_source": "results/episode_task_suite/object_relevance/metrics.json", "neural_metric_source": "results/episode_task_suite/neural_mlp/object_relevance/metrics.json" }, { "task": "caption_grounding", "task_display_name": "Language Grounding", "family": "retrieval", "unit": "caption query", "input": "caption object/interaction query plus candidate sensor windows", "target": "matching time window", "primary_metric": "mrr", "higher_is_better": true, "leakage_rule": "Queries are ranked against held-out candidate windows; reported ranks are computed after model scoring.", "counts": { "num_queries": 348, "num_train_windows": 813, "num_test_windows": 348 }, "minimal_primary_metric": 0.016023479050338015, "neural_primary_metric": 0.01684125567132316, "minimal_metric_source": "results/episode_task_suite/caption_grounding/metrics.json", "neural_metric_source": "results/episode_task_suite/neural_mlp/caption_grounding/metrics.json" }, { "task": "cross_modal_retrieval", "task_display_name": "Cross-Modal Retrieval", "family": "retrieval", "unit": "sensor query", "input": "motion, IMU, and camera query features", "target": "matching depth/video window", "primary_metric": "top5_accuracy", "higher_is_better": true, "leakage_rule": "Query-side and candidate-side feature blocks are split before projection/ranking.", "counts": { "num_queries": 348, "num_train_windows": 813, "num_test_windows": 348 }, "minimal_primary_metric": 0.367816091954023, "neural_primary_metric": 0.19827586206896552, "minimal_metric_source": "results/episode_task_suite/cross_modal_retrieval/metrics.json", "neural_metric_source": "results/episode_task_suite/neural_mlp/cross_modal_retrieval/metrics.json" }, { "task": "modality_reconstruction", "task_display_name": "Cross-Modal Reconstruction", "family": "cross-modal regression", "unit": "single window", "input": "motion, IMU, and camera features", "target": "depth/video feature vector", "primary_metric": "r2", "higher_is_better": true, "leakage_rule": "Target feature blocks are excluded from the input side.", "counts": { "num_train_windows": 813, "num_test_windows": 348 }, "minimal_primary_metric": -0.015271898913936655, "neural_primary_metric": -0.010171410134180991, "minimal_metric_source": "results/episode_task_suite/modality_reconstruction/metrics.json", "neural_metric_source": "results/episode_task_suite/neural_mlp/modality_reconstruction/metrics.json" }, { "task": "temporal_order", "task_display_name": "Temporal Order Verification", "family": "pairwise diagnostic", "unit": "adjacent window pair", "input": "two adjacent windows", "target": "correct versus reversed order", "primary_metric": "f1", "higher_is_better": true, "leakage_rule": "Pairs are built after windowing; labels are synthetic order labels, not input features.", "counts": { "num_samples": 2320, "num_train_samples": 1624, "num_test_samples": 696 }, "minimal_primary_metric": 0.5399515738498789, "neural_primary_metric": 0.8520179372197308, "minimal_metric_source": "results/episode_task_suite/temporal_order/metrics.json", "neural_metric_source": "results/episode_task_suite/neural_mlp/temporal_order/metrics.json" }, { "task": "misalignment_detection", "task_display_name": "Multimodal Synchronization Detection", "family": "pairwise diagnostic", "unit": "paired modality window", "input": "motion side plus visual/depth side", "target": "aligned versus shifted by 8 windows", "primary_metric": "f1", "higher_is_better": true, "leakage_rule": "Shift labels are synthetic targets; shifted visual/depth blocks are generated after feature splitting.", "counts": { "num_samples": 2306, "num_train_samples": 1614, "num_test_samples": 692 }, "minimal_primary_metric": 0.5051698670605613, "neural_primary_metric": 0.7152682255845944, "minimal_metric_source": "results/episode_task_suite/misalignment_detection/metrics.json", "neural_metric_source": "results/episode_task_suite/neural_mlp/misalignment_detection/metrics.json" } ], "global_leakage_controls": [ "Use chronological train/test splits instead of random window shuffling.", "Fit scalers and learned projections on train windows only.", "Keep future labels, future mocap, contact labels, object labels, and caption labels on the target side unless a task explicitly treats language as the query.", "For cross-modal tasks, split query-side and candidate-side feature blocks before training and ranking.", "Report unseen test classes when the chronological split exposes labels absent from the train segment." ], "current_limitations": [ "Cross-episode generalization for Qwen3-Omni has a first verified diagnostic pilot, but strong model quality is not yet shown.", "Feature-vector reconstruction is separate from pixel depth, mesh, NeRF, or Gaussian reconstruction.", "The final verified Qwen3-Omni diagnostic result meets the strict-JSON target, but action/subtask held-out quality remains weak and needs error analysis before larger model-quality claims.", "Full audio-visual representation learning still needs multi-episode training; the current report includes single-episode audio/no-audio ablations." ], "scale_up_gate": { "required_before_next_omni_quality_pilot": [ "selected prepared Xperience-10M episodes", "held-out episode split with no train/test episode leakage", "validation samples during training", "manifest, training metadata, progress logs, metrics, predictions, and run report", "held-out evaluation on test episodes rather than train windows" ], "current_status": "verified diagnostic result; strict-JSON quality target met, action/subtask quality still weak", "evidence": [ "docs/data/omni_finetune_verified_result.json", "results/omni_finetune/verified_public/" ] } }