video video 96.8 308 |
|---|
Humanola Egocentric Hand-Pose Dataset — Sample Delivery
Overview
Egocentric (head-mounted) manipulation video with synchronized 3D hand-pose tracking. Per frame: 3D hand keypoints in a gravity-aligned world frame, per-joint finger angles, a grip-closure scalar, the head camera's 6-DoF world pose, and a per-hand wrist pose. Four hardware-synchronized camera streams (head stereo + both wrists) and each camera's ~200 Hz IMU accompany every episode.
Format: LeRobot v2.1.
Dataset Statistics
| Metric | Value |
|---|---|
| Episodes | 9 |
| Total frames | 48,272 @ 30 fps |
| Total duration | ~27 min |
| State-vector dimension | 188 |
| Camera streams | 4 (head stereo L+R, wrist L+R), 1920×1200 |
| IMU | ~200 Hz × 3 (head + both wrists), gyro + accelerometer |
| Hand-tracking coverage | Both hands ≈100%; absences flagged per-frame (hand_present, tracking_quality) |
Head stereo: rectified, zero distortion, fx=fy≈754 px, baseline 0.120 m; wrist cameras mono; global shutter throughout. Exact per-episode camera intrinsics in meta/camera_intrinsics.json.
Episodes
| # | Duration | Frames | Task |
|---|---|---|---|
| 0 | 4:39 | 8,371 | ironing clothes |
| 1 | 2:11 | 3,931 | folding laundry |
| 2 | 4:00 | 7,201 | wiping |
| 3 | 5:08 | 9,241 | washing dishes |
| 4 | 1:36 | 2,904 | slicing carrots |
| 5 | 2:40 | 4,801 | peeling and slicing cucumbers |
| 6 | 1:47 | 3,211 | pouring cereal into a container |
| 7 | 2:00 | 3,601 | washing dishes |
| 8 | 2:47 | 5,011 | cutting an apple |
task_index → task lookup in meta/tasks.jsonl.
Directory Structure
humanola_sample_dataset/
├── README.md
├── meta/
│ ├── info.json # feature shapes + per-dim column names, collection info (v2.1)
│ ├── modality.json # observation.state field layout
│ ├── episodes.jsonl # per-episode: length, duration, coverage, task
│ ├── episodes_stats.jsonl # per-episode min/max/mean/std (v2.1 normalization stats)
│ ├── tasks.jsonl # task_index → task
│ └── camera_intrinsics.json # per-episode intrinsics for all 4 cameras
├── data/chunk-000/
│ └── episode_000000.parquet … # one row per timestep
├── videos/chunk-000/
│ ├── observation.images.main_stereo_left/episode_000000.mp4 … # head stereo, LEFT eye (rendered/primary view)
│ ├── observation.images.main_stereo_right/ … # head stereo, RIGHT eye
│ ├── observation.images.aux_left/ … # left-wrist camera
│ └── observation.images.aux_right/ … # right-wrist camera
├── imu/
│ ├── imu_info.json # per-sensor rate, units, camera↔IMU transform, video_key_map
│ └── chunk-000/
│ ├── main_stereo/episode_000000.parquet … # head-cam IMU
│ ├── aux_left/episode_000000.parquet … # left-wrist-cam IMU
│ └── aux_right/episode_000000.parquet … # right-wrist-cam IMU
└── extras/ # visualization aids, not loaded by LeRobot
├── overlays/episode_000000_overlay.mp4 … # 2D skeleton + box on the left eye
└── rrd/episode_000000.rrd … # rerun: 3D world hands + wrist-cam panels
All four camera streams are hardware-synchronized (same frame count, timestamps within one frame), cut to identical per-episode ranges from t=0.
State Vector Layout
observation.state — 188 float32 per frame (offsets in meta/modality.json):
| Field | Indices | Dims | Description |
|---|---|---|---|
left_hand_world_joints |
[0:63] | 21×3 | Left 21 joints xyz, world frame, meters |
left_finger_angles |
[63:93] | 15×2 | Left finger joints × [flexion, abduction], radians |
left_grip_closure |
[93:94] | 1 | Left grip-closure ∈ [0,1] (0 = open, 1 = fist) |
right_hand_world_joints |
[94:157] | 21×3 | Right 21 joints xyz, world frame, meters |
right_finger_angles |
[157:187] | 15×2 | Right finger joints × [flexion, abduction], radians |
right_grip_closure |
[187:188] | 1 | Right grip-closure ∈ [0,1] |
Joint order (21/hand — thumb cmc/mcp/ip/tip, fingers mcp/pip/dip/tip; per-dim names for every column in meta/info.json):
| Index | Joint | Index | Joint | Index | Joint |
|---|---|---|---|---|---|
| 0 | wrist | 7 | index_dip | 14 | ring_pip |
| 1 | thumb_cmc | 8 | index_tip | 15 | ring_dip |
| 2 | thumb_mcp | 9 | middle_mcp | 16 | ring_tip |
| 3 | thumb_ip | 10 | middle_pip | 17 | pinky_mcp |
| 4 | thumb_tip | 11 | middle_dip | 18 | pinky_pip |
| 5 | index_mcp | 12 | middle_tip | 19 | pinky_dip |
| 6 | index_pip | 13 | ring_mcp | 20 | pinky_tip |
World frame: right-handed and gravity-aligned, in meters, with up along −Y (gravity points toward +Y). It is fixed for the whole episode and does not move with the head; its horizontal orientation is arbitrary. observation.joints_cam.* holds the same joints in the camera frame.
Finger angles & grip-closure
Finger angles are 30 values, [flexion, abduction] per joint interleaved [j0_flex, j0_abd, …], ordered level-major (proximal → distal, thumb → pinky within each level):
| Slots (flex, abd) | Joint | Slots | Joint | Slots | Joint |
|---|---|---|---|---|---|
| 0–1 | thumb_cmc | 10–11 | thumb_mcp | 20–21 | thumb_ip |
| 2–3 | index_mcp | 12–13 | index_pip | 22–23 | index_dip |
| 4–5 | middle_mcp | 14–15 | middle_pip | 24–25 | middle_dip |
| 6–7 | ring_mcp | 16–17 | ring_pip | 26–27 | ring_dip |
| 8–9 | pinky_mcp | 18–19 | pinky_pip | 28–29 | pinky_dip |
- flexion: palm-curl, radians — 0 = straight, ≈1.5 (≈90°) = bent. abduction: sideways splay, signed, ≈±0.4. (Occasional ±π = angle-wrap.)
- grip-closure: mean flexion of the 5 MCP + 5 PIP joints ÷ 90°, clipped to [0,1]. 0 = open, 1 = fist; hand-size-independent.
Extra observation columns (not in observation.state)
| Column | Shape | Description |
|---|---|---|
observation.head_pose |
7 | Head-camera 6-DoF pose, world frame: [tx,ty,tz, qx,qy,qz,qw] (xyz meters, xyzw quaternion, camera→world) |
observation.wrist_pose.{left,right} |
7 | Per-hand wrist pose, world frame, same [tx,ty,tz, qx,qy,qz,qw] layout (wrist→world) |
observation.joints_cam.{left,right} |
63 | The 21 joints in the camera frame, meters |
observation.keypoints_2d.{left,right} |
42 | 21×2 pixel keypoints in the left-eye image |
observation.bbox.{left,right} |
4 | Hand detection box [x1,y1,x2,y2], left-eye image |
observation.hand_present.{left,right} |
1 | 1.0 tracked, 0.0 absent |
tracking_quality |
1 | 0 = both hands good, 1 = left absent, 2 = right absent, 3 = both |
When a hand is absent, its joints / keypoints / bbox / wrist_pose are zero-filled and hand_present = 0.
IMU
Each camera's ~200 Hz IMU is a separate parquet per episode at imu/chunk-000/<sensor>/episode_NNNNNN.parquet (main_stereo = head, aux_left/aux_right = wrists) — a sidecar, not a LeRobot feature, so read it directly. One row per sample:
| Column | Description |
|---|---|
sample_index |
sample index within the file |
timestamp_ns |
sensor timestamp on the shared camera clock (ns) — the cross-stream alignment key |
timestamp_s |
seconds from the episode's first frame |
frame |
fractional video-frame index; snap via round(frame) == frame_index |
angular_velocity_x, _y, _z |
gyroscope, rad/s |
linear_acceleration_x, _y, _z |
accelerometer, m/s² (gravity included) |
imu/imu_info.json has the per-sensor rate/units, the camera↔IMU transform (camera_imu_transform, 4×4, meters), and a video_key_map (sensor → video stream; main_stereo pairs with both head eyes).
Loading Data
import json, numpy as np, pyarrow.parquet as pq
t = pq.read_table("data/chunk-000/episode_000000.parquet")
state = np.array(t.column("observation.state").to_pylist(), dtype=np.float32) # (N, 188)
mod = json.load(open("meta/modality.json"))["state"]
def slice_(name): s = mod[name]; return state[:, s["start"]:s["end"]]
left_joints = slice_("left_hand_world_joints").reshape(-1, 21, 3) # (N, 21, 3) meters, world
right_joints = slice_("right_hand_world_joints").reshape(-1, 21, 3)
left_wrist = left_joints[:, 0] # (N, 3)
left_angles = slice_("left_finger_angles") # (N, 30) radians
left_grip = slice_("left_grip_closure") # (N, 1)
tq = np.array(t.column("tracking_quality").to_pylist()) # tq == 0 -> both hands good
# IMU: one file per sensor; align via the `frame` column
head_imu = pq.read_table("imu/chunk-000/main_stereo/episode_000000.parquet").to_pandas()
imu_at_frame_100 = head_imu[(head_imu.frame >= 99.5) & (head_imu.frame < 100.5)]
Visual checks (extras/)
overlays/episode_NNNNNN_overlay.mp4— 2D keypoints + boxes on the left eye (green = left, magenta = right).rrd/episode_NNNNNN.rrd— rerun: 3D world hand skeletons, head-camera frustum, wrist-camera panels.
Contact
- Downloads last month
- 110