humanola-inc/egocentric_labeled

Humanola Egocentric Hand-Pose Dataset

Overview

Human egocentric manipulation video with synchronized 3D hand-pose tracking. Each episode ships one synchronized camera stream plus per-frame 3D hand keypoints (gravity-aligned world frame), per-joint finger angles, a grip-closure scalar, and 6-DoF world poses for the head camera and both wrists. Each episode also ships the ~201 Hz IMU of the head camera, aligned to the video frames.

Format: LeRobot v2.1 (meta/episodes.jsonl + meta/tasks.jsonl + per-episode meta/episodes_stats.jsonl) with a meta/modality.json describing the state-vector field layout. Loads directly in the LeRobot library and any framework that consumes the LeRobot v2.1 format. The per-episode IMU is under imu/ (see IMU).

Dataset Statistics

Metric	Value
Episodes	38
Total frames	335,936 @ 29 fps
Total duration	~3.2 h
State-vector dimension	188 floats
Camera streams	1 (head camera), 3840×3360 each
IMU	~201 Hz, 1 stream (head), gyro + accelerometer, aligned to the frames
Hand-tracking coverage	Both hands ≈100% throughout. Absences are flagged per-frame (hand_present = 0, and per-hand hand_observed / hand_reliable).

Episodes

#	Duration	Frames	Left cov.	Right cov.	Task
0	5:01	8,752	100%	100%	mending garments
1	3:57	6,893	100%	100%	mending garments
2	9:50	17,113	99.5%	99.7%	mending garments
3	12:36	21,939	100%	100%	mending garments
4	4:12	7,309	100%	100%	mending garments
5	4:40	8,122	100%	100%	trimming threads
6	3:38	6,323	100%	100%	trimming threads
7	4:20	7,553	99.7%	100%	trimming threads
8	4:29	7,822	100%	99.9%	trimming threads
9	3:04	5,355	99.9%	99.7%	sewing panels
10	3:36	6,277	100%	100%	sewing panels
11	3:00	5,245	98.7%	99.2%	sewing panels
12	2:52	5,005	100%	100%	sewing panels
13	4:00	6,983	100%	100%	sewing panels
14	5:13	9,081	100%	100%	sewing panels
15	2:55	5,095	100%	100%	sewing pieces
16	5:22	9,350	100%	100%	sewing pieces
17	8:15	14,356	100%	99.7%	sewing pieces
18	3:02	5,304	100%	100%	sewing pieces
19	3:24	5,934	100%	100%	sewing pieces
20	5:06	8,900	100%	100%	sewing pieces
21	5:54	10,280	100%	99.0%	sewing pieces
22	3:23	5,904	100%	100%	sewing pieces
23	11:01	19,181	99.7%	99.6%	sewing pieces
24	3:49	6,654	100%	100%	sewing pieces
25	4:00	6,978	100%	99.9%	sewing pieces
26	5:00	8,700	100%	99.4%	sewing pieces
27	9:44	16,963	99.2%	99.6%	sewing pieces
28	3:07	5,425	100%	99.9%	sewing pieces
29	6:28	11,268	99.4%	100%	sewing pieces
30	5:17	9,201	100%	99.9%	sewing pieces
31	5:10	8,991	100%	100%	sewing pieces
32	4:56	8,602	100%	100%	sewing pieces
33	4:49	8,392	100%	98.8%	sewing pieces
34	2:52	5,005	100%	99.9%	sewing pieces
35	7:52	13,701	100%	98.8%	sewing pieces
36	4:42	8,181	100%	99.9%	sewing pieces
37	2:11	3,799	100%	100%	sewing pieces

Directory Structure

rerun_out/
├── README.md
├── meta/
│   ├── info.json                   # global metadata, feature shapes, collection info (v2.1)
│   ├── modality.json               # state-vector field layout + video stream specs
│   ├── episodes.jsonl              # per-episode: length, duration, coverage, task
│   ├── episodes_stats.jsonl        # per-episode min/max/mean/std (v2.1 normalization stats)
│   ├── tasks.jsonl                 # task-description lookup
│   └── camera_intrinsics.json      # per-episode intrinsics + calibration for all 1 camera
├── data/chunk-000/
│   └── episode_000000.parquet …    # one row per 29 fps timestep
├── videos/chunk-000/
│   └── observation.images.head/episode_000000.mp4 …    # head-mounted camera
├── imu/
│   ├── imu_info.json               # per-sensor rate, units, gravity, video_key_map
│   └── chunk-000/                  # one folder per sensor (named like the cameras: head)
│       └── head/episode_000000.parquet …   # head-cam IMU, ~200 Hz, aligned to the video frames
└── extras/                          # not loaded by LeRobot — visualization aids
    ├── overlays/episode_000000_overlay.mp4 …   # 2D skeleton + detection box + L/R tab, on the delivered video
    └── rrd/episode_000000.rrd …                # opens directly in rerun: 3D world hands + wrist-cam panels

State Vector Layout

Each frame has a 188-element float32 observation.state, laid out as (see meta/modality.json):

Field	Indices	Dims	Description
left_hand_world_joints	[0:63]	21×3	Left-hand 21 joints xyz, gravity-aligned metric world frame, meters
left_finger_angles	[63:93]	15×2	Left finger joints × [flexion, abduction], signed radians
left_grip_closure	[93:94]	1	Left grip-closure ∈ [0,1]: mean MCP+PIP finger flexion ÷ 90°. 0 = open hand, 1 = closed fist
right_hand_world_joints	[94:157]	21×3	Right-hand 21 joints xyz, world frame, meters
right_finger_angles	[157:187]	15×2	Right finger joints × [flexion, abduction], signed radians
right_grip_closure	[187:188]	1	Right grip-closure ∈ [0,1] (same definition)

Joint ordering (21 per hand, OpenPose hand convention): 0 wrist → 1–4 thumb (mcp, pip, dip, tip) → 5–8 index → 9–12 middle → 13–16 ring → 17–20 pinky.

Coordinate frame: gravity-aligned metric world frame, meters. Up axis = −Y (gravity points +Y); yaw is arbitrary, fixed at session start.

Finger angles & grip-closure — how they're defined

• Finger angles — anatomical joint angles in radians (not degrees). Each of the 15 finger joints (MCP, PIP, DIP × five fingers) carries a [flexion, abduction] pair: flexion = curl in the palm-curl plane (0 = straight, ≈1.5 rad ≈ 90° = bent), abduction = sideways splay (≈ ±0.4 rad). Values are signed; working range is flexion ≈ 0–1.5 and abduction ≈ ±0.4 rad (occasional ±π values are angle-wrap).
• Grip-closure (*_grip_closure) — a scalar in [0, 1] = mean flexion of the five MCP + five PIP knuckles ÷ 90°, clipped. 0 = open hand, 1 = closed fist (0.5 ≈ one knuckle row bent ~90°). Not a fingertip/pinch distance; hand-size-independent.

Extra observation columns (not part of observation.state)

Column	Shape	Description
observation.head_pose	7	Head-camera 6-DoF pose in the world frame: [tx,ty,tz, qx,qy,qz,qw] — xyz meters + xyzw quaternion (camera→world). Same world frame as the joints; valid every frame
observation.wrist_pose.{left,right}	7	Per-hand wrist 6-DoF pose in the world frame, same [tx,ty,tz, qx,qy,qz,qw] layout (wrist→world). Zero-filled where the hand is absent
observation.joints_cam.{left,right}	63	The same 21 joints in the camera frame, meters
observation.keypoints_2d.{left,right}	42	21×2 pixel keypoints in the delivered video's own pixels
observation.bbox.{left,right}	4	Hand detection box [x1,y1,x2,y2] in the delivered video's own pixels
observation.hand_present.{left,right}	1	1.0 if the hand is tracked this frame, else 0.0
observation.hand_observed.{left,right}	1	1.0 if a detector measured that hand on this frame (not inferred, coasted or filled). The single most useful filter for training
observation.hand_reliable.{left,right}	1	1.0 if the row is also unflagged: the fit was not doubted and no physical invariant fired. Train on hand_observed & hand_reliable
observation.keypoints_visible.{left,right}	21	Per-joint: 2 = inside the image and measured, 1 = inside the image but inferred, 0 = outside the image or the hand is absent. Coordinates are deliberately NOT clamped to the border

When a hand is absent, its joint / keypoint / bbox values are zero-filled and hand_present = 0.

IMU

Each camera's ~201 Hz IMU is its own file per episode — imu/chunk-000/<sensor>/episode_NNNNNN.parquet, one folder per sensor named like the cameras (head = head), mirroring how the videos are split per camera. One row per sample.

Column	Description
sample_index	0-based sample index within the file
timestamp_ns	sensor timestamp on the shared camera clock, nanoseconds — the canonical alignment key
timestamp_s	seconds from the episode's first frame
frame	fractional episode video-frame index (derived from timestamp_ns) — snap to a video frame by round(frame) == frame_index
angular_velocity_x, _y, _z	gyroscope, rad/s
linear_acceleration_x, _y, _z	accelerometer, m/s² (gravity included)

timestamp_ns is the same clock the video frames are captured on, so it's the key to align IMU with anything else; frame is a ready-made fractional frame index for snapping a sample to its video frame. imu/imu_info.json has the per-sensor rate, units, and a video_key_map tying each sensor to its video stream (head -> observation.images.head). The imu/ tree is a sidecar (not a LeRobot feature) — read it directly via these paths.

Collection Hardware

Component	Model	Specs
Head camera	monocular	3840×3360; fisheye_equidistant, NOT rectified; fx=fy=1698.35 px, cx=1920.0, cy=1680.0; D=[0.018617, 0.0699458, -0.0453681, 0.00948494]; FOV 122°×108° (157° diagonal)
IMU	1× (one per camera)	~201 Hz, gyro (rad/s) + accelerometer (m/s²); see imu/imu_info.json
Pose	21-joint 3D hand keypoints per hand in a gravity-aligned world frame; per-joint finger angles + a grip-closure scalar

Per-camera intrinsics are in meta/camera_intrinsics.json.

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)

# Which rows to trust. hand_observed: a detector measured that hand on THIS frame. hand_reliable:
# additionally unflagged -- no doubted fit, no physical invariant fired. Train on the conjunction.
obs_l = np.array(t.column("observation.hand_observed.left").to_pylist(), dtype=bool)
rel_l = np.array(t.column("observation.hand_reliable.left").to_pylist(), dtype=bool)
good_left = obs_l & rel_l                                            # (N,) bool

# Per-joint visibility: 2 = in frame and measured, 1 = in frame but inferred, 0 = out of frame/absent.
vis_l = np.array(t.column("observation.keypoints_visible.left").to_pylist(), dtype=np.float32)   # (N, 21)
# IMU — one file per sensor; align to a video frame via the `frame` column
head_imu = pq.read_table("imu/chunk-000/head/episode_000000.parquet").to_pandas()   # head cam
imu_at_frame_100 = head_imu[(head_imu.frame >= 99.5) & (head_imu.frame < 100.5)]   # head samples at frame 100