humanarchive/HA-Multi-Samples-v2

HA-Multi-Samples v2

A multimodal human activity dataset organized per-episode, per-modality with face-blurred RGB video (Meta EgoBlur Gen2), raw stereo pair, full-hand tactile sensing, and distributed body IMUs captured during everyday household tasks.

This is a reformatted and re-blurred release of the original HA-Multi-Samples dataset. The episode set, frame counts, durations, task labels, and environment labels are bit-identical to the v1 release, so any annotations or analyses keyed off (episode_index, frame_index) continue to work unchanged. What's new:

• EgoBlur Gen2 face blurring on all 4 RGB streams (replaces the v1 face blur model)
• Per-episode, per-modality folders for easier browsing and partial loading — no LeRobot dependency required
• .npy files for tactile, glove IMU, and body IMUs (one file per modality per episode)
• Raw stereo pair preserved (left + right rectified). A self-computed depth map (S²M² Large) will land in v2.1.

---

Dataset Summary

Metric	Value
Total episodes	36
Total frames	420,630
Frame rate	30 fps
Total duration	3 hours 54 minutes
Video streams	4 synchronized RGB cameras + computed depth
Sensor modalities	Tactile (512 taxels), Hand IMU (2), Body IMU (8)
Unique tasks	11
Unique environments	10
Cross-modal alignment	< 33 ms (< 1 frame at 30 fps)
Video data size	~64 GB (4 RGB streams)
Stereo data size	~26 GB (raw L+R, 1280×720, rectified)
Sensor data size	~66 MB
Average episode length	6.5 minutes
Median episode length	5.1 minutes
Shortest episode	22.3 seconds
Longest episode	21.3 minutes

---

Task and Environment Breakdown

By Task

Task	Episodes	Duration
Cleaning	19	118.9 min
Cooking	4	43.0 min
Ironing	4	31.2 min
Folding and cleaning	2	25.4 min
Folding clothes	6	14.2 min
Placing shoes	1	1.0 min

By Environment

Environment labels describe the room type, not unique rooms. Multiple episodes labeled "Bedroom" or "Kitchen" may come from different physical locations.

Environment	Episodes	Duration
Bedroom	20	115.8 min
Kitchen	7	61.0 min
Living room	3	32.3 min
Bathroom	3	18.3 min
Office	1	3.8 min
Hallway	2	2.5 min

---

File Structure

HA-Multi-Samples-v2/
├── episodes/
│   ├── ep_000/
│   │   ├── videos/
│   │   │   ├── egocentric.mp4        # 1920×1080, 30 fps, fisheye, face-blurred
│   │   │   ├── chest.mp4             # 1920×1080, 30 fps, face-blurred
│   │   │   ├── left_wrist.mp4        # 1920×1080, 30 fps, face-blurred
│   │   │   └── right_wrist.mp4       # 1920×1080, 30 fps, face-blurred
│   │   ├── stereo/
│   │   │   ├── left.mp4              # 1280×720, 30 fps, rectified left
│   │   │   └── right.mp4             # 1280×720, 30 fps, rectified right
│   │   │                             # (Depth coming in v2.1 from S²M² Large)
│   │   ├── tactile/
│   │   │   ├── left.npy              # (N, 256) float32
│   │   │   └── right.npy             # (N, 256) float32
│   │   ├── glove_imu/
│   │   │   ├── left.npy              # (N, 12) float32 — quaternion + accel/gyro
│   │   │   └── right.npy             # (N, 12) float32
│   │   ├── imu/
│   │   │   ├── head.npy              # (N, 9) accel+gyro+mag
│   │   │   ├── chest.npy             # (N, 6) accel+gyro
│   │   │   ├── left_bicep.npy        # (N, 6)
│   │   │   ├── right_bicep.npy       # (N, 6)
│   │   │   ├── left_forearm.npy      # (N, 6)
│   │   │   ├── right_forearm.npy     # (N, 6)
│   │   │   ├── left_hand.npy         # (N, 4) quaternion
│   │   │   └── right_hand.npy        # (N, 4) quaternion
│   │   └── meta.json                 # task, environment, source_run, num_frames, ...
│   ├── ep_001/
│   └── ...
├── meta/
│   ├── episodes.csv                  # one row per episode: ep_id, task, env, num_frames, ...
│   ├── tasks.csv                     # task_index → task_label
│   ├── calibration.json              # all camera intrinsics + stereo baseline + depth formula
│   └── stats.json                    # per-modality min/max/mean/std across the dataset
├── README.md
└── LICENSE

For every episode, the first axis of every .npy file equals the frame count of the four RGB videos. Frame i of any .npy corresponds to frame i of any of the four MP4s. Depth disparity is at half resolution (640×360) — see "Depth" below for conversion.

---

Methodology

Face Blurring — EgoBlur Gen2

All four RGB streams (egocentric, chest, left wrist, right wrist) are processed frame-by-frame with the EgoBlur Gen2 face detector (Meta Project Aria, Nov 2025), the successor to the original EgoBlur model. Each detected face bounding box is expanded by a factor of 1.15 to provide a buffer, then blurred with an elliptical Gaussian kernel sized in proportion to the face bbox. Detection score threshold: 0.55; NMS IoU threshold: 0.5. Stereo views are not blurred (downward-facing, no faces).

Stereo Pair

Each episode contains the raw rectified stereo pair from the head-mounted OAK-D camera, at 1280×720, 30 fps, H.264 encoded. The left and right videos are frame-aligned: frame i of stereo/left.mp4 corresponds to frame i of stereo/right.mp4 and to frame i of the RGB streams. Calibration (intrinsics + baseline) is in meta/calibration.json — see "Stereo Pair" under Modalities below.

Depth coming in v2.1. A self-computed depth map (from S²M² Large joint disparity/occlusion/confidence model) will be added in a follow-up release without changing the existing files. Episode meta.json already includes a depth_available: false flag that will flip when depth lands.

Recovering Depth From the Stereo Pair (manual)

import cv2, numpy as np

# Load and rectify is already done — these are rectified frames.
left = cv2.VideoCapture("HA-Multi-Samples-v2/episodes/ep_000/stereo/left.mp4")
right = cv2.VideoCapture("HA-Multi-Samples-v2/episodes/ep_000/stereo/right.mp4")
_, L = left.read(); _, R = right.read()
L_gray = cv2.cvtColor(L, cv2.COLOR_BGR2GRAY)
R_gray = cv2.cvtColor(R, cv2.COLOR_BGR2GRAY)

# Classical SGBM (cheap baseline); for higher quality use S²M² / RAFT-Stereo / etc.
stereo = cv2.StereoSGBM_create(minDisparity=0, numDisparities=128, blockSize=7)
disparity = stereo.compute(L_gray, R_gray).astype(np.float32) / 16.0

# Convert to depth
fx = 566.06
baseline_mm = 74.95
with np.errstate(divide="ignore", invalid="ignore"):
    depth_mm = (fx * baseline_mm) / disparity

Temporal Alignment

All sensor streams are synchronized to the video frame clock at 30 fps. Cross-modal alignment error is less than 33 ms (less than 1 frame). Variable-rate sensors (tactile gloves, BLE IMUs) are resampled to 30 fps using sample-and-hold: each video frame carries the most recent sensor reading available at that timestamp. The per-camera frame offsets used for alignment are identical to the v1 release.

Parity with v1

Every episode in v2 has identical num_frames, duration_s, task, and environment as v1. Frame i of episode j in v2 corresponds to frame i of episode j in v1, even though the underlying pixels are re-blurred. This is enforced programmatically: the build pipeline asserts each output .npy and .mp4 matches the v1 frame count before publishing.

---

Modalities

1. Video Streams (4 cameras)

All videos are H.264 encoded, 30 fps, with yuv420p pixel format.

Stream	Resolution	Mounting Position	Notes
egocentric	1920×1080	Head-mounted, first-person	Fisheye lens, wide-angle forward view
chest	1920×1080	Chest-mounted, downward-angled	Captures hands and workspace
left_wrist	1920×1080	Left wrist/forearm	Left hand and nearby objects
right_wrist	1920×1080	Right wrist/forearm	Right hand and nearby objects

Egocentric Camera Intrinsics

Fisheye lens, intrinsics at 1920×1080:

fx = 1093.98    fy = 1093.39
cx = 953.05     cy = 536.30

Stereo Pair

The head-mounted stereo pair (1280×720, 30 fps) is shipped as stereo/left.mp4 and stereo/right.mp4 per episode. Already rectified at recording time; frame-aligned with the four RGB streams.

Left stereo camera (at 1280×720):

fx = 566.06    fy = 566.02
cx = 640.75    cy = 400.78
Distortion model: Rational polynomial (14 coefficients)

Right stereo camera (at 1280×720):

fx = 566.54    fy = 566.69
cx = 644.35    cy = 403.60
Distortion model: Rational polynomial (14 coefficients)

Stereo geometry:

Baseline: 74.95 mm

The center RGB camera (egocentric/chest) sits approximately centered between the stereo pair — 37.4 mm to the right of the left camera and 37.6 mm to the left of the right camera. All cameras share a common rigid mount.

2. Tactile Sensors (256 taxels per hand)

Each hand is equipped with a full-coverage tactile glove containing 256 fiber-optic pressure sensors (taxels). The sensors use fiber-optic technology — light intensity through flexible optical fibers changes under mechanical pressure, providing responsive and high-dynamic-range force sensing across the entire hand surface. Values are unsigned 8-bit integers (0–255), stored as float32.

File (per episode)	Shape	Description
tactile/left.npy	(N, 256)	Left hand tactile pressure
tactile/right.npy	(N, 256)	Right hand tactile pressure

Pressure value ranges:

Contact Type	Typical Range
No contact	0
Light touch	1–5
Moderate grip	10–35
Hard press/grip	40–105
Sensor maximum	255

Approximately 60 of the 256 taxels are active during a typical grip. Some taxels may read zero consistently due to sensor placement or contact geometry. The taxel layout (finger phalanges, palm grid, bridge sensors) and the hand-motion-capture-from-tactile recipe are described in detail in the appendix below — they are unchanged from v1.

3. Hand IMU (12 values per hand)

Each glove contains an inertial measurement unit on the back of the hand, providing orientation and motion data.

File (per episode)	Shape	Description
glove_imu/left.npy	(N, 12)	Left hand IMU
glove_imu/right.npy	(N, 12)	Right hand IMU

The first 4 values are quaternion components [qx, qy, qz, qw] representing hand orientation. The remaining 8 values are supplementary IMU channels (accelerometer and gyroscope).

4. Body IMUs (8 streams)

IMU sensors are distributed across the upper body, providing acceleration and angular velocity data.

File (per episode)	Shape	Channels	Placement
imu/head.npy	(N, 9)	accel(3) + gyro(3) + mag(3)	On the camera, ~2 inches in front of the forehead
imu/chest.npy	(N, 6)	accel(3) + gyro(3)	Center of the sternum
imu/left_bicep.npy	(N, 6)	accel(3) + gyro(3)	Outer surface of the left upper arm
imu/right_bicep.npy	(N, 6)	accel(3) + gyro(3)	Outer surface of the right upper arm
imu/left_forearm.npy	(N, 6)	accel(3) + gyro(3)	Outer surface of the left forearm
imu/right_forearm.npy	(N, 6)	accel(3) + gyro(3)	Outer surface of the right forearm
imu/left_hand.npy	(N, 4)	quaternion(4)	Back of the left hand (from glove)
imu/right_hand.npy	(N, 4)	quaternion(4)	Back of the right hand (from glove)

For the 6-axis IMUs, the channel layout is [accel_x, accel_y, accel_z, gyro_x, gyro_y, gyro_z]. The head IMU includes 3 additional magnetometer channels. The hand IMUs provide orientation quaternions [qx, qy, qz, qw].

All IMU data is resampled to 30 fps to align with video frames using sample-and-hold interpolation from the original variable-rate sensor streams.

---

Loading the Dataset

Prerequisites

pip install huggingface_hub numpy opencv-python

Download

huggingface-cli login --token YOUR_TOKEN
huggingface-cli download humanarchive/HA-Multi-Samples-v2 \
    --repo-type dataset \
    --local-dir ~/HA-Multi-Samples-v2

Or with snapshot_download:

from huggingface_hub import snapshot_download
local = snapshot_download(repo_id="humanarchive/HA-Multi-Samples-v2", repo_type="dataset")

Loading One Episode

import json
import numpy as np
import cv2
from pathlib import Path

DATASET = Path("~/HA-Multi-Samples-v2").expanduser()

ep = DATASET / "episodes" / "ep_000"
meta = json.loads((ep / "meta.json").read_text())
print(meta["task"], meta["environment"], meta["num_frames"])

# Sensor data — all aligned to 30 fps, one row per video frame
tactile_left = np.load(ep / "tactile" / "left.npy")              # (N, 256)
tactile_right = np.load(ep / "tactile" / "right.npy")            # (N, 256)
glove_imu_left = np.load(ep / "glove_imu" / "left.npy")          # (N, 12)
glove_imu_right = np.load(ep / "glove_imu" / "right.npy")        # (N, 12)
head_imu = np.load(ep / "imu" / "head.npy")                      # (N, 9)
chest_imu = np.load(ep / "imu" / "chest.npy")                    # (N, 6)
left_bicep = np.load(ep / "imu" / "left_bicep.npy")              # (N, 6)
right_bicep = np.load(ep / "imu" / "right_bicep.npy")            # (N, 6)
left_forearm = np.load(ep / "imu" / "left_forearm.npy")          # (N, 6)
right_forearm = np.load(ep / "imu" / "right_forearm.npy")        # (N, 6)
left_hand = np.load(ep / "imu" / "left_hand.npy")                # (N, 4)
right_hand = np.load(ep / "imu" / "right_hand.npy")              # (N, 4)

# Stereo pair (raw, rectified)
stereo_left  = cv2.VideoCapture(str(ep / "stereo" / "left.mp4"))
stereo_right = cv2.VideoCapture(str(ep / "stereo" / "right.mp4"))

# RGB videos — load with cv2 or decord
cap = cv2.VideoCapture(str(ep / "videos" / "egocentric.mp4"))
frames = []
while True:
    ok, frame = cap.read()
    if not ok: break
    frames.append(frame)  # BGR HxWx3 uint8
ego = np.stack(frames)                                            # (N, 1080, 1920, 3)

Loading All Episodes

import pandas as pd

episodes = pd.read_csv(DATASET / "meta" / "episodes.csv")
tasks = pd.read_csv(DATASET / "meta" / "tasks.csv")

for _, row in episodes.iterrows():
    ep = DATASET / "episodes" / f"ep_{row['episode_index']:03d}"
    # ... load whatever modalities you need

Playing a Video

open ~/HA-Multi-Samples-v2/episodes/ep_000/videos/egocentric.mp4
ffplay ~/HA-Multi-Samples-v2/episodes/ep_005/videos/chest.mp4
ffplay ~/HA-Multi-Samples-v2/episodes/ep_007/stereo/left.mp4

---

Per-Episode Reference

Episode	Task	Environment	Frames	Duration
0	Cooking	Kitchen	26,469	14.7 min
1	Cleaning	Living room	17,792	9.9 min
2	Cleaning	Living room	38,395	21.3 min
3	Folding and cleaning	Bedroom	36,747	20.4 min
4	Placing shoes	Hallway	1,855	1.0 min
5	Cleaning	Bathroom	9,147	5.1 min
6	Cleaning	Office	6,842	3.8 min
7	Cleaning	Bedroom	17,165	9.5 min
8	Folding and cleaning	Bedroom	9,061	5.0 min
9	Cleaning	Bathroom	10,974	6.1 min
10	Cleaning	Bedroom	9,378	5.2 min
11	Folding clothes	Bedroom	19,023	10.6 min
12	Cleaning	Kitchen	17,684	9.8 min
13	Cleaning	Living room	1,925	1.1 min
14	Cleaning	Bedroom	15,399	8.6 min
15	Folding clothes	Bedroom	902	0.5 min
16	Folding clothes	Bedroom	1,193	0.7 min
17	Folding clothes	Bedroom	3,001	1.7 min
18	Folding clothes	Bedroom	675	0.4 min
19	Folding clothes	Bedroom	706	0.4 min
20	Cleaning	Bedroom	670	0.4 min
21	Cleaning	Bathroom	12,838	7.1 min
22	Cleaning	Hallway	2,603	1.4 min
23	Cooking	Kitchen	36,822	20.5 min
24	Cooking	Kitchen	3,772	2.1 min
25	Cooking	Kitchen	10,386	5.8 min
26	Cleaning	Bedroom	5,622	3.1 min
27	Cleaning	Bedroom	10,024	5.6 min
28	Cleaning	Bedroom	1,656	0.9 min
29	Cleaning	Kitchen	8,522	4.7 min
30	Cleaning	Kitchen	6,173	3.4 min
31	Cleaning	Bedroom	21,041	11.7 min
32	Ironing	Bedroom	1,909	1.1 min
33	Ironing	Bedroom	27,703	15.4 min
34	Ironing	Bedroom	14,642	8.1 min
35	Ironing	Bedroom	11,914	6.6 min

---

License

Released under Creative Commons BY-NC 4.0 for research and non-commercial use. The original v1 dataset is licensed identically.

Citation

@misc{humanarchive2026hamulti,
  title  = {HA-Multi-Samples v2: A Multimodal Human Activity Dataset with Tactile and Computed Depth},
  author = {Human Archive Team},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/humanarchive/HA-Multi-Samples-v2}}
}

Face blurring uses the EgoBlur Gen2 model from Meta. Stereo depth uses S²M² (Junhong et al.). The original sensor recordings and alignment match HA-Multi-Samples v1.

---

Appendix: Taxel Layout and Hand Motion Capture

The taxel index mapping (per-finger, per-phalanx) and the bend-from-tactile recipe are unchanged from v1 and reproduced below for completeness.

See full taxel layout in the v1 dataset card — finger mapping, bridge sensors, palm grid, and compute_finger_bend() helper.