Datasets:
File size: 2,978 Bytes
4bb2c3f | 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 | {
"id": "policy_vla_branch",
"display_name": "VLA / Policy Model Branch",
"status": "planned_adapter",
"model_family": "OpenVLA, openpi, GR00T, Octo, and related policy models",
"default_model_id": null,
"local_model_env": "POLICY_MODEL_DIR",
"dataset_contract": "xperience10m_observation_action_v0",
"training_objective": "observation_to_action_or_motion_policy",
"split_policy": {
"unit": "episode",
"default_counts": {
"train": 96,
"val": 16,
"test": 16
},
"leakage_guard": "action targets and normalization statistics must be fit on train episodes only"
},
"modalities": {
"observations": [
"egocentric video",
"language instruction or task context",
"optional depth/pose/mocap/IMU state"
],
"candidate_targets": [
"action label",
"next action",
"hand trajectory chunk",
"contact state",
"retargeted body or humanoid action",
"robot-compatible action token"
],
"excluded_inputs": [
"visualization.rrd"
]
},
"entrypoints": {
"selection_manifest": "scripts/omni/build_selection_episode_manifest.py",
"neutral_index": "scripts/omni/export_model_neutral_window_index.py",
"export": null,
"train": null,
"eval": null,
"launcher": null,
"validate": "scripts/omni/validate_omni_finetune_run.py"
},
"primary_metrics": [
"action_accuracy",
"next_action_accuracy",
"contact_accuracy",
"trajectory_mpjpe",
"object_affordance_f1",
"held_out_episode_count"
],
"artifact_contract": {
"checkpoint_gate": "policy_checkpoint_action_space_and_normalizer",
"required_eval_files": [
"metrics.json",
"policy_predictions.jsonl",
"trajectory_metrics.csv",
"action_confusion_matrix.csv",
"retargeting_audit.json",
"RUN_REPORT.md"
],
"required_training_files": [
"training_metadata.json",
"progress.jsonl",
"action_space.json",
"normalization_stats.json",
"checkpoint_manifest.json"
],
"public_package_allowed": [
"metrics",
"policy prediction summaries",
"trajectory metric tables",
"action confusion matrices",
"action-space definitions",
"normalization metadata",
"retargeting audit summaries",
"validation summaries"
],
"public_package_forbidden": [
"raw MP4",
"annotation HDF5",
"Rerun RRD",
"private retargeting source files",
"base-model weights",
"full checkpoints",
"large archives"
]
},
"extension_requirements": [
"Define an explicit action space before policy fine-tuning.",
"Implement target conversion from human egocentric motion to policy-compatible action tokens or trajectories.",
"Fit action normalizers on train episodes only and save them with the run manifest.",
"Add policy evaluation that separates classification, trajectory, and retargeting metrics."
]
}
|