CaryxAI/everyday-activities-ego-raw

CaryX Everyday Ego (sample dataset)

28 recordings · 10.21 hours · 64.1 GB · 16 tasks at home and at work · 616 labelled windows of about one minute · 2,134 labelled sub-steps

CaryX Everyday Ego: long-form, first-person video of everyday household chores and hands-on work, recorded by CaryX AI contributors on phones they wear. This dataset is a sample of the larger CaryX Everyday Ego corpus. Each file is one recording session of one task, 7 to 62 minutes long: washing dishes, folding laundry, cooking, cleaning, gardening, caring for animals, taking apart furniture, and work at shop benches: repairing watches, and vinyl decal work at a sign shop that includes packing orders. A recording does not always reach the end of its task. The montage above shows one moment of each recording.

What's in it

file	what it is
clips/<clip_id>.mp4	one recording: H.264 video and the phone's AAC audio (mono or stereo)
index.csv / index.jsonl	one row per recording: task, scene, device, duration, frame rate, frame gaps, rotation, and machine labels
windows.csv / windows.jsonl	a coarse timeline: one row per window of about one minute, with what the hands do in it
spans.csv / spans.jsonl	a fine timeline: one row per sub-step ("rinse pot lid under faucet"), grouped into keysteps

All recordings are landscape 1920×1080 at ~30 fps (measured 27.4–30.0 fps). There is no depth, no camera pose, and no calibrated intrinsics: this is plain RGB video from consumer phones.

Tasks

task	where	recordings	hours	clips	task value
Vinyl decal signs, including packing orders	work	1	1.03	EA1-0028	pack_a_shipping_box_dense_taped_labeled
Watch repair	work	1	0.66	EA1-0027	watch_repair
Washing dishes by hand	home	5	1.93	EA1-0002, EA1-0004, EA1-0017, EA1-0022, EA1-0023	hand_wash_dishes
Folding laundry	home	4	1.56	EA1-0001, EA1-0013, EA1-0018, EA1-0020	fold_shirts_pants_towels_shorts
Cooking lunch	home	4	1.54	EA1-0008, EA1-0015, EA1-0021, EA1-0024	make_lunch
Taking apart furniture	home	1	0.66	EA1-0005	assemble_disassemble_furniture
Machine laundry	home	1	0.59	EA1-0009	machine_laundry_load_transfer_unload_run_dryer
Cooking breakfast	home	2	0.41	EA1-0019, EA1-0026	make_breakfast
Setting and clearing a table	home	1	0.34	EA1-0011	set_and_clear_the_table
Weeding plants	home	2	0.34	EA1-0010, EA1-0025	weed_beds_and_pots
Cooking dinner	home	1	0.27	EA1-0016	make_dinner
Making a drink	home	1	0.20	EA1-0003	prepare_drinks_coffee_juice
Cleaning a bathroom	home	1	0.19	EA1-0006	clean_the_bathroom
Feeding chickens	home	1	0.18	EA1-0007	feed_chickens_collect_eggs
Bathing a dog	home	1	0.17	EA1-0014	bathe_and_dry_a_dog
Watering plants	home	1	0.15	EA1-0012	water_plants

task is the task the contributor chose before recording. Each recording also has a short scene note from the collection (for example where the chore happens and which lens was used).

Devices and capture conditions

Recorded on 16 Android phone models, usually worn on the head and angled down at the work surface. Placement varies: EA1-0006 films from above the wearer's head, and EA1-0012 looks forward and shows mostly the right hand. Lens: main 22, wide 6 (wide is the phone's ultra-wide camera, so its field of view is larger and more distorted).

device	recordings
Xiaomi 24117RN76G	4
realme RMX3760	3
samsung SM-A356E	3
HONOR REA-NX9	2
TECNO CM7	2
Xiaomi 2602BPC18G	2
realme RMX3941	2
vivo V2352	2
Xiaomi 2201117SG	1
Xiaomi 24094RAD4G	1
Xiaomi 25080RABDG	1
samsung SM-A075F	1
samsung SM-A166P	1
samsung SM-A245F	1
samsung SM-A576B	1
vivo V2319	1

Recording length: min 7.2 min · median 16.7 min · max 62.0 min.

The timeline (windows.csv)

Every recording is cut into 60-second windows; the last window runs 22 to 73 seconds (a remainder under 20 s joins the window before it). For each window a vision-language model looked at 12 evenly spaced frames and wrote what the wearer's hands do, the objects they handle, and whether people, faces, or screens appear. Windows are a fixed grid, not action boundaries: one window can hold the end of one step and the start of the next. Use them to find and skim the part of a recording you want. 9 of 616 windows are labelled unclear: most hold several steps in one window or fiddly small-object handling, and a few are walking or pausing.

column	meaning
clip_id, window	the recording and the window's index within it
start_s, end_s	the window, in seconds on the video's own timestamps
task_phrase	a short verb phrase for what the hands do, or unclear
description	one or two sentences on the window
primary_objects / secondary_objects	objects handled / tools and containers, ; -separated, empty when the model listed none
environment, hands_visible, lighting	the space, which of the wearer's hands appear, light level
faces_visible, other_people_visible, screens_or_documents_visible	what else appears
label_source	vlm, or vlm_corrected where the review changed at least one field (the other fields may still be unchecked)
label_model	the model that wrote the labels

Sub-steps (spans.csv)

Every recording is also cut into sub-steps: short actions written as verb + object, such as "scrub bowl with sponge" or "brush small part with brush", grouped under a keystep. 2,134 sub-steps, about 3.5 per minute; median length 8 s; 193 (9%) are confidence = low. An example from the watch repair recording:

clip_id	time	keystep	label
EA1-0027	0:02:50–0:02:54	place tweezers on table	place tweezers on table
EA1-0027	0:02:54–0:03:00	open drawer	open drawer
EA1-0027	0:03:00–0:03:02	retrieve white wire from drawer	retrieve white wire from drawer
EA1-0027	0:03:02–0:03:08	manipulate white wire	manipulate white wire

column	meaning
clip_id, span	the recording and the sub-step's index within it
start_s, end_s	the sub-step, in seconds on the video's own timestamps
keystep	the group the sub-step belongs to (often the same text when a keystep has one sub-step)
label	verb + object (+ place)
object	the main object handled, empty when the model named none
confidence	high, or low when the check pass could not confirm the label
label_model	the model that wrote the labels

The sub-steps are machine-generated and unreviewed. To get the timeline of one recording, or every moment a tool is used:

steps = spans[spans.clip_id == "EA1-0027"].sort_values("start_s")
tweezers = spans[spans.label.str.contains("tweezers")]

index.csv carries the same label columns for the whole recording, plus:

column	meaning
file, size_bytes	the clip's path in this repository and its size
task, scene	the chore chosen before recording, and the collection's scene note
device, lens	phone model, and main or wide camera
duration_s, n_frames, fps	length, frame count, and average frame rate (frames ÷ duration)
width, height, rotation	stored frame size, and 180 where the file is stored upside down
max_frame_gap_ms, n_frame_gaps, missing_s	longest step between frames, number of skipped-frame gaps, and seconds missing
n_windows	rows this recording has in windows.csv

Identifiers

id	example	meaning	stable?
clip_id	EA1-0001	the recording. Assigned once, in submission order	permanent
window	0	a window within a recording	with its clip_id
span	0	a sub-step within a recording	with its clip_id

Split on device, not on clip_id. Recordings with the same phone model are likely, but not certainly, the same person and home (8 models appear in 2–4 recordings each). Windows from one recording share the person, the room, and the objects.

Using it

import pandas as pd, av                      # pip install av pandas huggingface_hub
from huggingface_hub import hf_hub_download

REPO = "CaryxAI/everyday-activities-ego-raw"
read = lambda f: pd.read_csv(hf_hub_download(REPO, f, repo_type="dataset"), keep_default_na=False)
index, windows, spans = read("index.csv"), read("windows.csv"), read("spans.csv")

# find every window where the hands rinse something, then open that recording
hits = windows[windows.task_phrase.str.contains("rins", case=False)]
row  = hits.iloc[0]
rot  = index.set_index("clip_id").at[row.clip_id, "rotation"]
path = hf_hub_download(REPO, f"clips/{row.clip_id}.mp4", repo_type="dataset")

with av.open(path) as c:
    s = c.streams.video[0]
    s.thread_type = "AUTO"
    c.seek(int(row.start_s / s.time_base), stream=s)
    for frame in c.decode(s):
        t = float(frame.pts * s.time_base)
        if t < row.start_s:
            continue
        if t >= row.end_s:
            break
        img = frame.to_ndarray(format="rgb24")      # (1080, 1920, 3)
        if rot == 180:
            img = img[::-1, ::-1].copy()            # stored upside down

Each recording is 0.8–6.5 GB, so pick from index.csv and windows.csv before you download.

Honest limitations

Read this section before you trust a column.

The labels are machine-generated and mostly unreviewed. Treat windows.csv, spans.csv and the label columns of index.csv as a searchable first pass, not ground truth; label_source = vlm_corrected marks rows a person changed, and label_model names the model.

The lenses differ. Phones and lenses vary between recordings, and none are calibrated. Read device and lens per recording.

Privacy and consent

Every recording was made by an adult contributor (18 or older) who signed a content agreement that permits commercial use and open release of their recordings. Every tag from the delivered file (including any location, creation-time, and encoder tags) is stripped. The files keep only the MP4 writer's fixed defaults (brand, language, handler_name) and the display-rotation flag. The audio is kept as recorded: it can carry speech, other people's voices, television or radio, and no label describes it. The video itself is unedited: faces are not blurred. EA1-0027 and EA1-0028 were recorded at workbenches in a watch repair shop and a sign shop. The visibility flags show where the model saw other people (1 recording), faces (1 recording), or screens and documents (3 recordings) in at least one window.

Licence

CC-BY-4.0. Commercial use permitted with attribution. A summary and the link to the full legal code are in LICENSE.

Citation

@misc{ea1_2026,
  title        = {CaryX Everyday Ego (sample dataset)},
  author       = {CaryX AI},
  year         = {2026},
  publisher    = {CaryX AI},
  howpublished = {\url{https://caryx.ai}},
  url          = {https://huggingface.co/datasets/CaryxAI/everyday-activities-ego-raw},
  note         = {Dataset. Available at \url{https://caryx.ai/data}}
}

---

Built by CaryX AI. We record everyday human manipulation and turn it into densely annotated, machine readable episodes for physical AI.

Browse our annotated episodes in your browser → · caryx.ai · Reach out founders@caryx.ai