{
  "basis": "Human-1M \u00b7 LeRobot v3.0",
  "tree": "release-staging/oss_100k_v1/\n\u2514\u2500\u2500 vendors/\n    \u2514\u2500\u2500 {vendor_id}/\n        \u2514\u2500\u2500 {batch_id}/\n            \u251c\u2500\u2500 versions.json\n            \u2514\u2500\u2500 v0/\n                \u2514\u2500\u2500 shard_00/\n                    \u251c\u2500\u2500 data/\n                    \u2502   \u2514\u2500\u2500 chunk-000/file-000.parquet\n                    \u251c\u2500\u2500 meta/\n                    \u2502   \u251c\u2500\u2500 info.json\n                    \u2502   \u251c\u2500\u2500 episodes/chunk-000/file-000.parquet\n                    \u2502   \u251c\u2500\u2500 tasks.parquet\n                    \u2502   \u251c\u2500\u2500 subtasks.parquet\n                    \u2502   \u251c\u2500\u2500 environments.parquet\n                    \u2502   \u2514\u2500\u2500 scenes.parquet\n                    \u2514\u2500\u2500 videos/\n                        \u2514\u2500\u2500 observation.images.{camera}/\n                            \u2514\u2500\u2500 chunk-000/file-000.mp4",
  "fields": [
    {
      "group": "Shard info",
      "field": "codebase_version",
      "shape": "string",
      "description": "LeRobot v3.0.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Shard info",
      "field": "data_format_version",
      "shape": "string",
      "description": "Human-1M format version declared by the delivery.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Shard info",
      "field": "total_episodes",
      "shape": "integer",
      "description": "Episode count within this shard.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Shard info",
      "field": "fps",
      "shape": "number > 0",
      "description": "Delivered frame rate; resolve timing from the episode and stream metadata.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Shard info",
      "field": "data_path",
      "shape": "path template",
      "description": "Relative path template for frame Parquet files.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Shard info",
      "field": "video_path",
      "shape": "path template",
      "description": "Relative path template for per-camera MP4 files.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Shard info",
      "field": "features",
      "shape": "mapping",
      "description": "Feature names, dtypes, and shapes, including the available camera streams.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Frame data",
      "field": "index",
      "shape": "int64",
      "description": "Index identifying a frame row in the dataset.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Frame data",
      "field": "episode_index",
      "shape": "int64",
      "description": "Shard-local episode reference; join to episode metadata.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Frame data",
      "field": "frame_index",
      "shape": "int64",
      "description": "Frame position inside the episode.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Frame data",
      "field": "timestamp",
      "shape": "float",
      "description": "Frame timestamp in seconds, using the delivery clock convention.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Frame data",
      "field": "task_index",
      "shape": "int64",
      "description": "Reference to the T1 episode-level task instruction.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Frame data",
      "field": "subtask_index",
      "shape": "int64",
      "description": "Reference to the T2 subtask label; preserve explicit segment boundaries separately.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Frame data",
      "field": "observation.state.hand_left_cam",
      "shape": "21 \u00d7 3 float32",
      "description": "Left-hand joint positions in the camera frame, in metres; stored as 63 values.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Frame data",
      "field": "observation.state.hand_right_cam",
      "shape": "21 \u00d7 3 float32",
      "description": "Right-hand joint positions in the camera frame, in metres; stored as 63 values.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Frame data",
      "field": "observation.state.hand_left_cam_rotation",
      "shape": "21 \u00d7 4 float32",
      "description": "Left-hand per-joint quaternions in (qx, qy, qz, qw) order; stored as 84 values.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Frame data",
      "field": "observation.state.hand_right_cam_rotation",
      "shape": "21 \u00d7 4 float32",
      "description": "Right-hand per-joint quaternions in (qx, qy, qz, qw) order; stored as 84 values.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Frame data",
      "field": "observation.state.camera_position",
      "shape": "3 float32",
      "description": "Camera translation in the world frame, in metres.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Frame data",
      "field": "observation.state.camera_rotation",
      "shape": "4 float32",
      "description": "Camera orientation in the world frame, (qx, qy, qz, qw).",
      "basis": "Human-1M schema"
    },
    {
      "group": "Frame data",
      "field": "environment_index",
      "shape": "int64",
      "description": "Reference to the environment dictionary.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Frame data",
      "field": "scene_index",
      "shape": "int64",
      "description": "Reference to the scene dictionary.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Frame data",
      "field": "scene_id",
      "shape": "identifier",
      "description": "Consistent anonymous scene identifier.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Frame data",
      "field": "collector_id",
      "shape": "identifier",
      "description": "Consistent anonymous collector identifier.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Frame data",
      "field": "left_hand_mask / right_hand_mask",
      "shape": "int64, 0 or 1",
      "description": "Separate per-hand presence indicators.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Frame data",
      "field": "left_hand_confidence / right_hand_confidence",
      "shape": "float",
      "description": "Per-hand tracking confidence; consult the supplied confidence convention.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Episode metadata",
      "field": "episode_index",
      "shape": "int64",
      "description": "Shard-local episode index; not the globally unique episode ID.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Episode metadata",
      "field": "episode_id",
      "shape": "string",
      "description": "Stable, globally unique episode identifier after consistent anonymization.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Episode metadata",
      "field": "length",
      "shape": "int64",
      "description": "Number of frames in the episode.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Episode metadata",
      "field": "data/chunk_index \u00b7 data/file_index",
      "shape": "int64",
      "description": "Locate the episode frame-data file through info.json.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Episode metadata",
      "field": "dataset_from_index \u00b7 dataset_to_index",
      "shape": "int64",
      "description": "Frame-row interval, with an exclusive upper bound.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Episode metadata",
      "field": "meta/episodes/chunk_index \u00b7 meta/episodes/file_index",
      "shape": "int64",
      "description": "Location of the episode metadata record.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Episode metadata",
      "field": "videos/observation.images.{cam}/chunk_index \u00b7 file_index",
      "shape": "int64",
      "description": "Locate the MP4 for each available camera using video_path.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Episode metadata",
      "field": "videos/observation.images.{cam}/from_timestamp \u00b7 to_timestamp",
      "shape": "seconds",
      "description": "Episode interval inside that camera\u2019s concatenated video. These are not subtask boundaries.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Episode metadata",
      "field": "camera_intrinsics/{cam}",
      "shape": "calibration parameters",
      "description": "Per-episode camera calibration. Use the camera model supplied for the rig; do not assume every camera uses the same projection.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Episode metadata",
      "field": "camera_extrinsics/T_{i}_{j}",
      "shape": "rigid transform",
      "description": "Camera j expressed in camera i\u2019s coordinate frame. Interpret the delivered transform convention.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Episode metadata",
      "field": "rig_id",
      "shape": "string",
      "description": "Recording-rig identifier, consistently anonymized.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Episode metadata",
      "field": "country",
      "shape": "string",
      "description": "ISO 3166-1 alpha-2 recording country code.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Episode metadata",
      "field": "capture_privacy",
      "shape": "public | private",
      "description": "Declared recording-environment privacy category. Do not infer it solely from the broad environment label.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Episode metadata",
      "field": "vendor_id",
      "shape": "string",
      "description": "Stable anonymous collection-source identifier on every episode row in meta/episodes/*.parquet, including nested chunk files.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Language & context",
      "field": "tasks.parquet: task + task_index",
      "shape": "text index + integer",
      "description": "T1: episode-level natural-language task instruction and its local dictionary index.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Language & context",
      "field": "subtasks.parquet: subtask + subtask_index",
      "shape": "text index + integer",
      "description": "T2: natural-language subtask description and its label reference.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Language & context",
      "field": "subtasks.parquet: episode_index",
      "shape": "integer",
      "description": "Episode containing the annotation occurrence.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Language & context",
      "field": "subtasks.parquet: timestamp_start \u00b7 timestamp_end",
      "shape": "seconds",
      "description": "Annotation occurrence boundaries in the episode timestamp convention. Adjacent identical descriptions can remain separate rows.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Language & context",
      "field": "subtasks.parquet: source_id",
      "shape": "string",
      "description": "Anonymous annotation-provenance identifier.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Language & context",
      "field": "environments.parquet: environment + environment_index",
      "shape": "text index + integer",
      "description": "Local environment description and broad environment category.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Language & context",
      "field": "scenes.parquet: scene + scene_index",
      "shape": "text index + integer",
      "description": "Scene description and the local scene lookup index.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Video streams",
      "field": "observation.images.main",
      "shape": "MP4",
      "description": "Main monocular egocentric video; a stereo delivery may also supply main as the left-view alias.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Video streams",
      "field": "observation.images.main_left / main_right",
      "shape": "MP4",
      "description": "Left and right egocentric streams for stereo capture.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Video streams",
      "field": "observation.images.wrist_left / wrist_right",
      "shape": "MP4, when present",
      "description": "Left and right wrist-camera streams.",
      "basis": "Human-1M schema"
    },
    {
      "group": "Video streams",
      "field": "observation.images.main_add{k}",
      "shape": "MP4, when present",
      "description": "Additional main camera streams.",
      "basis": "Human-1M schema"
    }
  ],
  "notes": [
    "Deliveries use Human-1M V1.7 and LeRobot v3.0. Read meta/info.json for feature definitions, frame rate, and path templates.",
    "Release paths follow vendors/{vendor_id}/{batch_id}/v0/{shard_name}/ under release-staging/oss_100k_v1. Resolve frame and video files with each shard\u2019s meta/info.json and episode metadata.",
    "Each vN is a complete full_clean snapshot. Read one version of a delivery rather than concatenating versions.",
    "Batch-level versions.json, alongside v0/, records version entries, changed episodes, blurred episodes, removed episodes, and creation information.",
    "Use each camera\u2019s own episode video offset.",
    "Subtask boundaries use timestamp_start and timestamp_end in subtasks.parquet. Video from_timestamp/to_timestamp locates the episode within a longer MP4.",
    "For each hand joint, p_world = R(camera_rotation) \u00d7 p_camera + camera_position. The world frame has negative Z toward gravity; projection follows the supplied camera calibration.",
    "Hand joint order: wrist (0); thumb CMC/MCP/IP/tip (1\u20134); index MCP/PIP/DIP/tip (5\u20138); middle (9\u201312); ring (13\u201316); little (17\u201320).",
    "Use per-hand masks and confidence to distinguish valid tracking. A zeroed failure record is a sentinel, not a valid rotation.",
    "Capture-privacy and full calibration coverage require episode-level metadata; they cannot be established from aggregate curation statistics.",
    "Gallery videos are silent H.264 excerpts. Read the full episode\u2019s timing and calibration metadata when working with geometry."
  ],
  "storage_url": "https://huggingface.co/buckets/nvidia/human-1m-100k-oss/tree/release-staging"
}