humyn-labs/LATAM-Egocentric-Residential-IMU

LATAM Egocentric Residential (with IMU)

Head-mounted, first-person video of everyday household chores recorded across Brazil, Argentina, Venezuela and Peru, each paired with a ~100 Hz accelerometer + gyroscope IMU stream.

The dataset targets embodied-AI and robotics research that needs real, unscripted human manipulation in cluttered domestic environments — not lab-staged demonstrations.

<video controls width="720" src="https://huggingface.co/datasets/humyn-labs/LATAM-Egocentric-Residential-IMU/resolve/main/preview/sampleLATAMODD16.mp4"></video>

Preview: `LATAM_OD_D_16` — gardening, outdoor, daytime, Argentina (30 s excerpt, downscaled to 720p). Full-resolution clips are in the viewer above and under [`videos/`](./tree/main/videos).

At a glance

Tasks covered

Folding Clothes (9) · Cleaning (3) · Organizing Wardrobe (2) · Gardening (2) · Sweeping The Floor · Organizing Cabinet · Washing Dishes · Organizing Desk

Repository layout

data/train-*.parquet   # video bytes + all metadata (powers the Dataset Viewer)
videos/*.mp4           # the same 20 clips as standalone files
imu/*_imu.jsonl        # per-sample IMU streams
preview/               # short downscaled clip used in this card
metadata.csv           # flat table of every column

data/ and videos/ hold the same footage — parquet for streaming with datasets, raw MP4s for direct download.

Columns

IMU format

One JSON object per line, ~100 Hz:

{"accelerometer":{"x":-0.0937,"y":0.0038,"z":-0.9875},
 "gyroscope":{"x":-0.0283,"y":-0.0013,"z":0.0256},
 "clock":"mach_absolute_time",
 "relativeMs":102.570792,
 "timestampEpochMs":1776729664005.492,
 "timestampNs":391427589938125}

Accelerometer is in g, gyroscope in rad/s. relativeMs counts from the start of the IMU capture; timestampEpochMs is Unix epoch milliseconds.

imu_alignment — read this before training on IMU

IMU was recorded per capture session, and a session sometimes spans more than one clip.

• —`clip` (13 samples) — the IMU file belongs to this clip alone.
• —`session` (7 samples) — the IMU file is shared with other clips from the same session, identified by imu_session_id. The signal is genuine, but the offset of each clip within the stream is not recoverable from the released files, because the MP4 creation_time records export time rather than capture start.

Three shared sessions exist: LATAM_ID_D_3+LATAM_ID_N_18, LATAM_OD_N_4+LATAM_OD_N_19, and LATAM_ID_N_11+LATAM_ID_N_20+LATAM_ID_N_21. In each case the IMU span slightly exceeds the summed clip durations, which is what identified them as session-level recordings.

If you need frame-accurate video–IMU sync, filter to `imu_alignment == "clip"`.

Usage

from datasets import load_dataset

ds = load_dataset("humyn-labs/LATAM-Egocentric-Residential-IMU", split="train")
print(ds[0]["country"], ds[0]["task"])

# frame-accurate IMU only
synced = ds.filter(lambda r: r["imu_alignment"] == "clip")

Load the IMU for a sample:

import json
from huggingface_hub import hf_hub_download

row = ds[0]
path = hf_hub_download(
    "humyn-labs/LATAM-Egocentric-Residential-IMU",
    row["imu_path"], repo_type="dataset",
)
imu = [json.loads(l) for l in open(path)]
print(len(imu), imu[0]["accelerometer"])

Grab the raw MP4s without the parquet:

from huggingface_hub import snapshot_download

snapshot_download(
    "humyn-labs/LATAM-Egocentric-Residential-IMU",
    repo_type="dataset", allow_patterns=["videos/*", "imu/*"],
)

Intended uses

Egocentric action recognition · hand–object interaction · activities of daily living · video + IMU sensor fusion · imitation learning for household robotics · long-horizon task understanding.

Limitations

• —Small. 20 clips totalling 39 minutes — a sample pack for evaluation and prototyping, not large-scale pretraining.
• —Skewed. Folding Clothes is 9 of 20 samples, and night recordings outnumber day 13:7. Low-light noise is common in the night clips.
• —No audio, and no action-boundary, hand-pose or depth annotations. Metadata is clip-level only.
• —Mixed resolution (1920×1440 and 1920×1080), so a resize or crop step is needed for fixed-size batching.
• —IMU sync is clip-accurate for only 13 of 20 samples — see above.

Provenance

Curated from the HumynLabs egocentric sample collection. One row of the source sheet (LATAM_OD_N_6) pointed at the same file as LATAM_OD_N_4 and was removed, leaving 20 unique clips — verified by SHA-256 over the full media. All technical fields (duration_seconds, width, height, fps) were probed from the media rather than copied from the sheet.

License

CC BY 4.0. Recorded with participant consent for research use. Faces are largely out of frame given the head-mounted viewpoint, but incidental bystanders and personal home interiors do appear — please handle accordingly and do not attempt to re-identify individuals.

Citation

@misc{humynlabs2026latamegocentric,
  title  = {LATAM Egocentric Residential (with IMU)},
  author = {HumynLabs},
  year   = {2026},
  url    = {https://huggingface.co/datasets/humyn-labs/LATAM-Egocentric-Residential-IMU}
}