File size: 11,903 Bytes
540e67a a8124a8 540e67a df8f96e 540e67a a8124a8 540e67a df8f96e 540e67a df8f96e 540e67a df8f96e 540e67a df8f96e 540e67a df8f96e 540e67a df8f96e 540e67a df8f96e 540e67a df8f96e 540e67a a8124a8 540e67a df8f96e 540e67a df8f96e 540e67a df8f96e 540e67a a8124a8 540e67a df8f96e 540e67a df8f96e 540e67a df8f96e 540e67a | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 | {
"source": {
"shared_windows": "results/episode_task_suite/shared_windows.npz",
"windows_csv": "results/episode_task_suite/windows.csv",
"feature_manifest": "results/episode_task_suite/feature_manifest.json"
},
"dataset_scope": {
"sample_episode_count": 1,
"num_windows": 1161,
"feature_dim": 8546,
"first_start_frame": 0,
"last_end_frame": 5819,
"warning": "Single public sample episode; these extension probes validate task design and pipeline mechanics, not cross-episode generalization."
},
"baselines": {
"minimal": "Ridge classifiers/regressors/projections plus cosine retrieval on the committed feature tensor.",
"neural_mlp": "Small one-hidden-layer PyTorch MLP heads using the same inputs, targets, chronological split, and evaluator."
},
"run_config": {
"train_fraction": 0.7,
"ridge_l2": 10.0,
"seed": 7,
"future_windows": 4,
"neural_epochs": 25,
"neural_hidden_dim": 128,
"neural_batch_size": 128,
"skip_neural": false
},
"task_specs": {
"body_motion_intensity": {
"direction": "A",
"direction_name": "Human Modeling & Motion Understanding",
"name": "Body and Hand Motion Intensity",
"family": "classification",
"case_study": "A window with a fast reach or pour should be classified as high motion; a steady holding window should be low motion.",
"input": "Current non-mocap feature blocks: video, audio, depth, camera pose/rotation, IMU, SLAM, calibration, and language context.",
"middle_process": "Compute the target from hand/body joint changes between neighboring windows, hide the mocap blocks from the input, then classify high versus low motion using the train-set median as the threshold.",
"output": "Binary label: high_motion or low_motion.",
"minimal_baseline": "Ridge classifier on standardized non-mocap features.",
"neural_baseline": "One-hidden-layer MLP binary classifier on the same input features.",
"metric_name": "macro-F1",
"metric_key": "macro_f1",
"metric_direction": "higher",
"current_limit": "This is a motion-energy proxy, not a SMPL/MANO body model or a generative motion prior."
},
"multi_view_consistency_retrieval": {
"direction": "B",
"direction_name": "3D/4D Reconstruction & Neural Rendering",
"name": "Multi-View Consistency Retrieval",
"family": "retrieval",
"case_study": "Given the fisheye camera features for a pouring moment, retrieve the synchronized stereo-left view from the same time window.",
"input": "Query side: fisheye_cam0 video feature block. Candidate side: stereo_left video feature block from held-out windows.",
"middle_process": "Learn a projection from one camera-view feature space into another, then rank held-out candidate windows by cosine similarity.",
"output": "Ranked candidate windows; the correct synchronized view should rank near the top.",
"minimal_baseline": "Ridge projection followed by cosine nearest-neighbor retrieval.",
"neural_baseline": "One-hidden-layer MLP projection followed by the same cosine retrieval evaluator.",
"metric_name": "MRR",
"metric_key": "mrr",
"metric_direction": "higher",
"current_limit": "This checks calibrated multi-view signal, but it is still feature retrieval, not NeRF, Gaussian Splatting, or novel-view synthesis."
},
"action_phase_progress": {
"direction": "C",
"direction_name": "Egocentric Vision & Interaction",
"name": "Action Phase Progress Estimation",
"family": "regression",
"case_study": "Inside a Pour coffee action segment, estimate whether the current window is near the beginning, middle, or end of that action.",
"input": "Current non-caption multimodal feature vector, so the label text cannot be copied directly from the language block.",
"middle_process": "Convert contiguous action-label runs into a normalized 0-to-1 progress target, train on earlier windows, and regress progress for later windows.",
"output": "A scalar progress value between 0.0 and 1.0 for the current action segment.",
"minimal_baseline": "Ridge regressor on standardized non-caption features.",
"neural_baseline": "One-hidden-layer MLP regressor on the same input features.",
"metric_name": "MAE",
"metric_key": "mae",
"metric_direction": "lower",
"current_limit": "This is an action-structure probe inside one episode, not a general intent model across homes, people, or tasks."
},
"ego_motion_forecast": {
"direction": "D",
"direction_name": "Scene Reconstruction & World Modeling",
"name": "Short-Horizon Ego-Motion Forecasting",
"family": "forecast",
"case_study": "From the current sensors, predict how the camera translation will change over the next 20 frames while the wearer moves through the scene.",
"input": "Current multimodal features excluding the camera-translation block and caption text.",
"middle_process": "Build a future target from camera-translation difference at a four-window horizon, then regress that future ego-motion delta from current sensors.",
"output": "A future camera-translation delta vector.",
"minimal_baseline": "Ridge regressor with a 20-frame forecast horizon.",
"neural_baseline": "One-hidden-layer MLP regressor with the same horizon and split.",
"metric_name": "MAE",
"metric_key": "mae",
"metric_direction": "lower",
"current_limit": "This is a compact world-model proxy; it does not build a persistent map, scene graph, or object permanence model."
}
},
"tasks": {
"body_motion_intensity": {
"train_windows": 812,
"test_windows": 348,
"target_threshold_train_median": 0.476467490196228,
"input_dim": 6425,
"target_source": "hand/body joint delta between neighboring windows",
"minimal": {
"accuracy": 0.7758620689655172,
"macro_f1": 0.7658385093167701,
"positive_rate_true": 0.35919540229885055,
"positive_rate_pred": 0.4339080459770115,
"num_test": 348
},
"neural_mlp": {
"accuracy": 0.8304597701149425,
"macro_f1": 0.8254423029509534,
"positive_rate_true": 0.35919540229885055,
"positive_rate_pred": 0.47126436781609193,
"num_test": 348
},
"neural_training": {
"available": true,
"epochs": 25,
"hidden_dim": 128,
"loss_history": [
0.37322977610996794,
0.22245765099384515,
0.1382973729242832,
0.10363741681493562,
0.0795709453523159,
0.06539858697817244,
0.055655122610735776,
0.043255199022187385,
0.03558319240001035,
0.031631215764530776,
0.029465350402711796,
0.024383640274625695,
0.02020622924740972,
0.016222001351599624,
0.018758724778523587,
0.013199950316049196,
0.014794624612432689,
0.01013119441452505,
0.009688532855040554,
0.008956241283767622,
0.006733611014469761,
0.006677041435843618,
0.0067647325489761795,
0.005346325666556511,
0.004052691048084588
]
}
},
"multi_view_consistency_retrieval": {
"train_windows": 813,
"test_windows": 348,
"query_block": "video_fisheye_cam0",
"target_block": "video_stereo_left",
"query_dim": 686,
"target_dim": 686,
"minimal": {
"mrr": 0.5533982515335083,
"top1": 0.41954022988505746,
"top5": 0.7068965517241379,
"top10": 0.8304597701149425,
"median_rank": 2.0,
"num_test": 348
},
"neural_mlp": {
"mrr": 0.34691643714904785,
"top1": 0.23275862068965517,
"top5": 0.46264367816091956,
"top10": 0.5890804597701149,
"median_rank": 7.0,
"num_test": 348
},
"neural_training": {
"available": true,
"epochs": 25,
"hidden_dim": 128,
"loss_history": [
0.9800718805740094,
0.8296866191855803,
0.7029420470986126,
0.6089339927846948,
0.5426930648814268,
0.49323386093200644,
0.45315542230600214,
0.4240272395578551,
0.3964498403400806,
0.37567753094588696,
0.3599070675332021,
0.3417643405048023,
0.32952829051721577,
0.31516501450450657,
0.3070896395824639,
0.29752101269888553,
0.287490411878072,
0.2791558311654193,
0.2707079971921693,
0.2669465311998811,
0.2603630047442728,
0.2501040017656148,
0.24714160980920216,
0.24146720613060843,
0.23866056472173036
]
}
},
"action_phase_progress": {
"train_windows": 813,
"test_windows": 348,
"input_dim": 7650,
"target_source": "normalized position inside contiguous action-label runs",
"minimal": {
"mse": 0.16943013668060303,
"mae": 0.32674381136894226,
"r2": -1.0236103208433347,
"num_test": 348
},
"neural_mlp": {
"mse": 0.14226463437080383,
"mae": 0.301545649766922,
"r2": -0.6991557278041094,
"num_test": 348
},
"neural_training": {
"available": true,
"epochs": 25,
"hidden_dim": 128,
"loss_history": [
3.236165899632162,
1.2893786148831414,
0.8036823107014429,
0.5113777590120557,
0.42643894586909153,
0.37627028047965166,
0.30304253713524504,
0.25041163572526065,
0.20774717810409096,
0.18116216590630319,
0.16409150619165191,
0.14362229159397394,
0.1277421933001991,
0.1232111468672899,
0.11924667946013724,
0.11960234325310401,
0.09951645682988572,
0.08239271907825459,
0.08965909919102401,
0.07973466079503408,
0.07939992471154765,
0.07085797808340408,
0.08331163055269188,
0.068286436959313,
0.07575550977814241
]
}
},
"ego_motion_forecast": {
"train_windows": 810,
"test_windows": 347,
"forecast_horizon_windows": 4,
"forecast_horizon_frames": 20,
"input_dim": 7629,
"target_dim": 21,
"target_source": "future minus current camera_translation feature block",
"minimal": {
"mse": 3.3189830780029297,
"mae": 0.16999860107898712,
"r2": -5674.96718626448,
"num_test": 347
},
"neural_mlp": {
"mse": 0.547074019908905,
"mae": 0.0833469107747078,
"r2": -934.5800418396838,
"num_test": 347
},
"neural_training": {
"available": true,
"epochs": 25,
"hidden_dim": 128,
"loss_history": [
1.034793225188314,
0.6070329941349265,
0.459870958622591,
0.3616479166495947,
0.2902924293353234,
0.23913239262722158,
0.1907091121982645,
0.17297036942140556,
0.15034288657300265,
0.13807891327657817,
0.13084740807980666,
0.12387588925567675,
0.12085371225336451,
0.11271689225126195,
0.10728766140010622,
0.10044601888936243,
0.09365091108613544,
0.09090288755150489,
0.08592776887946658,
0.0805281208805096,
0.08014915316929051,
0.07900124887625376,
0.07897219589830917,
0.07575527565714754,
0.0774567014273302
]
}
}
}
} |