File size: 16,375 Bytes
d4558b2 | 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 | {
"$schema": "../schemas/world-model-label-audit.schema.json",
"schema_version": "1.0.0",
"audit_id": "egobowwow-world-model-label-audit-v1",
"researched_at": "2026-08-25",
"boundary": "Primary papers and official dataset/toolkit pages define evidence. The audit records published supervision, not inferred capabilities, downloaded rows, or acceptance evidence. A UniversalLabeler translation may change language surfaces but must preserve all grounded dimensions unless a target profile reports explicit loss.",
"universal_dimensions": [
{"dimension_id": "environment_context", "graph_paths": ["environment.site_context_id", "environment.physical_environment_id", "environment.scene_type_id", "environment.station_type_id"], "rubric": "Label social site context, indoor/outdoor exposure, scene, and work station independently; do not infer one axis from another."},
{"dimension_id": "environment_conditions", "graph_paths": ["environment.lighting_id", "environment.weather_id", "environment.clutter_id", "environment.material_state_ids"], "rubric": "Record only observable physical conditions; use explicit unknowns rather than title-based inference."},
{"dimension_id": "scene_geometry", "graph_paths": ["environment.map_artifact_id", "assertions[kind=mask_ref]", "assertions[kind=keypoints_3d]"], "rubric": "Bind maps, depth-derived geometry, or point clouds by content-addressed reference and named coordinate frame."},
{"dimension_id": "coordinate_calibration", "graph_paths": ["coordinate_frames"], "rubric": "Every metric 2D/3D value names its clock and coordinate frame; frame transforms bind a calibration digest."},
{"dimension_id": "time_sync", "graph_paths": ["modalities.clock_id", "assertions.interval", "assertions.timestamp_ns"], "rubric": "Use named synchronized clocks and integer nanoseconds with half-open intervals; frame indices are not time authority."},
{"dimension_id": "language", "graph_paths": ["speech", "modalities[kind=language]"], "rubric": "Keep verbatim text, language tag, speech act, register, manner, purpose, and grammatical features separate from grounded event semantics."},
{"dimension_id": "temporal_hierarchy", "graph_paths": ["event.temporal", "procedure.activity_category_id", "procedure.task_id", "procedure.step_id", "assertions[kind=temporal_span]"], "rubric": "Keep activity, task, step, atomic action, and sensor sample boundaries distinct and linked."},
{"dimension_id": "agent_identity_role", "graph_paths": ["agents.agent_id", "agents.agent_kind", "agents.role_ids"], "rubric": "Give each human or robot a stable event-local identity and explicit performer, instructor, observer, or platform role."},
{"dimension_id": "hand_laterality_role", "graph_paths": ["agents.hands.laterality", "agents.hands.role_ids"], "rubric": "Label left and right independently and distinguish acting, stabilizing, supporting, idle, and unknown roles."},
{"dimension_id": "hand_pose", "graph_paths": ["agents.hands.pose_label_keys", "assertions[kind=keypoints_2d]", "assertions[kind=keypoints_3d]", "assertions[kind=pose_3d]"], "rubric": "Pose claims require a handedness identity, model/layout, visibility, coordinate frame, timestamp, confidence, and uncertainty."},
{"dimension_id": "body_head_pose", "graph_paths": ["agents.body_pose_label_keys", "assertions[kind=keypoints_3d]", "assertions[kind=pose_3d]", "assertions[kind=trajectory_3d]"], "rubric": "Keep body/head geometry separate from action classes and preserve the source skeleton convention."},
{"dimension_id": "gaze", "graph_paths": ["agents.gaze_label_keys", "relations[predicate_id=gaze_at]", "modalities[kind=eye_gaze]"], "rubric": "Represent gaze as a timed geometric signal in a named frame; gaze-at-object is a separate derived relation."},
{"dimension_id": "object_identity_role", "graph_paths": ["entities.entity_id", "entities.category_id", "entities.role_ids"], "rubric": "Track object instances separately from categories and record patient, instrument, source, destination, support, and container roles."},
{"dimension_id": "object_geometry_pose", "graph_paths": ["entities.geometry_label_keys", "entities.pose_label_keys", "entities.trajectory_label_keys", "assertions[kind=box_2d]", "assertions[kind=mask_ref]", "assertions[kind=pose_3d]", "assertions[kind=trajectory_3d]"], "rubric": "Do not flatten boxes, masks, meshes, 6DoF poses, or trajectories into an object noun."},
{"dimension_id": "contact_spatial_relations", "graph_paths": ["relations", "agents.hands.contact_entity_ids", "assertions[kind=object_relation]"], "rubric": "Relations identify subject, predicate, object, temporal phase, confidence, and supporting assertion."},
{"dimension_id": "object_state_transition", "graph_paths": ["entities.state_before_ids", "entities.state_after_ids", "assertions[kind=state_transition]"], "rubric": "State axes name the entity and observable before/after values; do not infer success from an action verb."},
{"dimension_id": "goal_outcome", "graph_paths": ["procedure.goal_state_ids", "procedure.precondition_ids", "procedure.outcome"], "rubric": "Instructions, goals, preconditions, observed terminal state, and outcome are distinct fields."},
{"dimension_id": "mistake_intervention", "graph_paths": ["procedure.correctness", "procedure.error_ids", "procedure.intervention_ids"], "rubric": "Annotate the observable mistake, recovery, and intervention independently from task identity."},
{"dimension_id": "proficiency", "graph_paths": ["procedure.proficiency"], "rubric": "Proficiency is separately assessed metadata and must not be inferred from identity or one local mistake."},
{"dimension_id": "multimodal_observation", "graph_paths": ["modalities"], "rubric": "Declare each available RGB, depth, audio, IMU, gaze, pose, segmentation, force, tactile, language, action, or proprioceptive stream with clock/frame bindings."},
{"dimension_id": "embodied_action_proprioception", "graph_paths": ["assertions[kind=action_chunk]", "assertions[kind=proprio_series]", "modalities[kind=robot_action]", "modalities[kind=proprioception]"], "rubric": "Robot command vectors and observed state vectors keep their layouts, sample periods, coordinate frames, and timestamps."},
{"dimension_id": "quality_uncertainty_provenance", "graph_paths": ["assertions.confidence", "assertions.uncertainty", "assertions.visibility", "provenance"], "rubric": "Every claim retains observability, confidence/uncertainty, producer, rubric, source profile, and review state; missing labels are not negative labels."}
],
"typed_label_kinds": [
"text_class",
"temporal_span",
"box_2d",
"mask_ref",
"keypoints_2d",
"keypoints_3d",
"pose_3d",
"trajectory_3d",
"audio_event",
"object_relation",
"state_transition",
"action_chunk",
"proprio_series"
],
"sources": [
{
"source_id": "ego_exo4d_v2",
"title": "Ego-Exo4D: Understanding Skilled Human Activity from First- and Third-Person Perspectives",
"paper_url": "https://arxiv.org/abs/2311.18259",
"official_url": "https://docs.ego-exo4d-data.org/",
"release_year": 2024,
"native_units": ["take", "synchronized ego/exo camera stream", "narration", "atomic action interval", "keystep interval", "3D hand/body pose sample", "proficiency assessment"],
"annotation_rubric": "Preserve synchronized multiview capture and task-specific layers: narration and expert commentary, fine-grained action/keystep timing, proficiency, gaze, camera pose, and 3D hand/body pose are parallel supervision rather than one caption.",
"supervision": ["multichannel audio", "eye gaze", "IMU", "3D point clouds", "camera poses", "language descriptions", "expert commentary", "fine-grained activity", "proficiency", "3D hand/body pose"],
"supported_dimensions": ["time_sync", "language", "temporal_hierarchy", "agent_identity_role", "hand_laterality_role", "hand_pose", "body_head_pose", "gaze", "proficiency", "multimodal_observation", "quality_uncertainty_provenance"],
"known_gaps": ["The activity label alone does not encode object state, contact, or exact environment conditions.", "Different benchmark annotation layers have different temporal and visibility coverage."]
},
{
"source_id": "hot3d_v1",
"title": "HOT3D: Hand and Object Tracking in 3D from Egocentric Multi-View Videos",
"paper_url": "https://arxiv.org/abs/2411.19167",
"official_url": "https://github.com/facebookresearch/hot3d",
"release_year": 2025,
"native_units": ["sequence", "timestamp_ns", "camera stream", "left/right MANO or UmeTrack hand pose", "object 6DoF pose", "headset pose", "visibility/QA mask"],
"annotation_rubric": "At each timestamp and stream, keep handedness, wrist transform, hand model parameters/joints, object instance pose, camera calibration, visibility, exposure, availability, and manual pose-QA masks distinct.",
"supervision": ["multi-view RGB/monochrome", "eye gaze", "scene point clouds", "camera poses", "left/right 3D hand poses", "object 6DoF poses", "hand/object 3D models", "per-stream quality masks"],
"supported_dimensions": ["scene_geometry", "coordinate_calibration", "time_sync", "hand_laterality_role", "hand_pose", "gaze", "object_identity_role", "object_geometry_pose", "multimodal_observation", "quality_uncertainty_provenance"],
"known_gaps": ["Pose availability is not pose accuracy; HOT3D exports separate availability and QA masks.", "The geometric labels do not by themselves state semantic goal, success, error, or hand functional role."]
},
{
"source_id": "aria_digital_twin_v1",
"title": "Aria Digital Twin: A New Benchmark Dataset for Egocentric 3D Machine Perception",
"paper_url": "https://arxiv.org/abs/2306.06362",
"official_url": "https://www.projectaria.com/datasets/adt/",
"release_year": 2023,
"native_units": ["sequence", "sensor calibration", "continuous device/object 6DoF pose", "3D gaze vector", "3D human pose", "2D segmentation", "depth map"],
"annotation_rubric": "Treat the real scene, calibrated sensor rig, static/dynamic object instances, device trajectory, gaze, human pose, segmentation, depth, and synthetic twin as synchronized but separately typed ground truth.",
"supervision": ["RGB/monochrome", "IMU", "complete calibration", "device and object 6DoF poses", "3D eye gaze", "3D human pose", "2D segmentations", "depth", "photorealistic twin"],
"supported_dimensions": ["environment_context", "scene_geometry", "coordinate_calibration", "time_sync", "body_head_pose", "gaze", "object_identity_role", "object_geometry_pose", "multimodal_observation", "quality_uncertainty_provenance"],
"known_gaps": ["The benchmark ground truth is primarily geometric and does not supply a universal manipulation-action or mistake ontology.", "The published release covers two indoor scenes and should not imply broad environment diversity."]
},
{
"source_id": "holoassist_v1",
"title": "HoloAssist: an Egocentric Human Interaction Dataset for Interactive AI Assistants in the Real World",
"paper_url": "https://arxiv.org/abs/2309.17024",
"official_url": "https://holoassist.github.io/",
"release_year": 2023,
"native_units": ["collaborative session", "synchronized sensor stream", "coarse/fine action interval", "conversation turn", "mistake", "intervention", "hand forecast target"],
"annotation_rubric": "Keep performer and instructor roles, grounded verbal guidance, coarse/fine physical actions, mistakes, recovery/intervention behavior, gaze/head/hand signals, and future-hand targets as linked layers.",
"supervision": ["seven synchronized streams", "action annotations", "conversation annotations", "mistakes", "intervention types", "hand forecasting"],
"supported_dimensions": ["time_sync", "language", "temporal_hierarchy", "agent_identity_role", "hand_laterality_role", "hand_pose", "gaze", "goal_outcome", "mistake_intervention", "multimodal_observation", "quality_uncertainty_provenance"],
"known_gaps": ["Conversation does not replace physical action or state evidence.", "A mistake label is task-relative and cannot be projected safely without the enclosing procedure context."]
},
{
"source_id": "nymeria_v1",
"title": "Nymeria: A Massive Collection of Multimodal Egocentric Daily Motion in the Wild",
"paper_url": "https://arxiv.org/abs/2406.09905",
"official_url": "https://www.projectaria.com/datasets/nymeria/",
"release_year": 2024,
"native_units": ["recording", "localized synchronized device stream", "full-body motion sample", "fine motion narration", "atomic action", "activity summary"],
"annotation_rubric": "Preserve metric-world localization, synchronized multi-device streams, full-body motion, gaze/IMU, and the hierarchy from fine narration through atomic action to high-level activity summary.",
"supervision": ["full-body ground-truth motion", "egocentric video", "eye tracking", "IMU", "third-person view", "metric-world localization", "hierarchical motion language"],
"supported_dimensions": ["environment_context", "coordinate_calibration", "time_sync", "language", "temporal_hierarchy", "agent_identity_role", "body_head_pose", "gaze", "multimodal_observation", "quality_uncertainty_provenance"],
"known_gaps": ["Motion-language hierarchy does not by itself provide object instance state, contact, or manipulation success.", "Natural-language summaries and metric motion have different annotation granularity."]
},
{
"source_id": "droid_rlds_v1",
"title": "DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset",
"paper_url": "https://arxiv.org/abs/2403.12945",
"official_url": "https://github.com/droid-dataset/droid",
"release_year": 2024,
"native_units": ["episode", "RLDS step", "three language instructions", "three camera observations", "robot Cartesian/joint/gripper state", "commanded action"],
"annotation_rubric": "At every ordered step, keep language instruction variants, camera observations, measured Cartesian/joint/gripper state, commanded Cartesian/joint/gripper action, reward, discount, and terminal flags separate.",
"supervision": ["wrist and exterior RGB", "language instruction variants", "Cartesian state/action", "joint state/action", "gripper state/action", "reward and terminal flags", "scene diversity"],
"supported_dimensions": ["environment_context", "time_sync", "language", "agent_identity_role", "goal_outcome", "multimodal_observation", "embodied_action_proprioception", "quality_uncertainty_provenance"],
"known_gaps": ["The released RLDS schema does not universally provide human hand pose, object instance pose, contact, or explicit before/after object state.", "Vector meaning depends on robot/action layout and coordinate conventions and must never be translated as untyped numbers."]
},
{
"source_id": "open_x_embodiment_rlds_v1",
"title": "Open X-Embodiment: Robotic Learning Datasets and RT-X Models",
"paper_url": "https://arxiv.org/abs/2310.08864",
"official_url": "https://robotics-transformer-x.github.io/",
"release_year": 2024,
"native_units": ["dataset", "episode", "RLDS step", "observation", "action", "language instruction", "robot-specific metadata"],
"annotation_rubric": "Normalize container structure while preserving each contributor's robot morphology, observation/action fields, units, coordinate frames, success semantics, language, and missing modalities as dataset-specific contracts.",
"supervision": ["standardized RLDS episodes", "multi-robot observations", "robot actions", "language instructions", "527 skills across heterogeneous platforms"],
"supported_dimensions": ["environment_context", "time_sync", "language", "agent_identity_role", "goal_outcome", "multimodal_observation", "embodied_action_proprioception", "quality_uncertainty_provenance"],
"known_gaps": ["A common container is not a common semantic action space; robot-specific layouts and units remain essential.", "Missing modalities and heterogeneous success labels require explicit loss/availability reporting."]
}
]
}
|