Datasets:

CaryxAI
/

everyday-manipulation-3d-raw

license: cc-by-4.0
pretty_name: Everyday Manipulation 3D (raw RGB-D)
size_categories:
  - 1K<n<10K
task_categories:
  - robotics
  - video-classification
tags:
  - egocentric
  - robotics
  - manipulation
  - rgbd
  - lidar
  - depth
  - hand-object-interaction
  - embodied-ai
  - arkit
  - record3d
configs:
  - config_name: default
    data_files:
      - split: train
        path: index.csv

Everyday Manipulation 3D (raw RGB-D)

1,513 clips · 10.28 hours · 279 GiB · 4 participants · 10 manipulation tasks · 42 recording sittings

Chest-mounted iPhone Pro capture of everyday two-handed manipulation by CaryX AI. Clips were recorded with Record3D, an iOS app that captures the iPhone's LiDAR RGB-D stream. Each clip is the app's .r3d recording with the audio track removed; the sensor streams are unmodified: synchronised RGB, metric LiDAR depth, per-frame ARKit 6-DoF camera pose, and per-frame camera intrinsics.

Why this exists

An open sample of raw 3D egocentric manipulation data for training and building models, released under CC BY 4.0, including commercial use with attribution. For labels, larger volumes, or questions, reach out to founders@caryx.ai.

What's in a clip

Each .r3d is a ZIP containing:

member	what it is
`metadata`	JSON: `K` (3×3 intrinsics, column-major), `poses` (per-frame `[qx,qy,qz,qw,tx,ty,tz]`), `frameTimestamps`, `perFrameIntrinsicCoeffs`, `initPose`, `cameraType`, `w`/`h`, `dw`/`dh`, `fps`
`rgbd/<i>.jpg`	RGB frame
`rgbd/<i>.depth`	metric depth, float32 metres, LZFSE-compressed
`rgbd/<i>.conf`	ARKit per-pixel confidence, uint8 (0=low, 1=med, 2=high), also LZFSE-compressed
`icon`	a small JPEG thumbnail written by the app

Both orientations are present. 1,027 clips are portrait and 486 are landscape, and the frame dimensions swap accordingly:

orientation	clips	RGB	depth
portrait	1,027	1440×1920	192×256
landscape	486	1920×1440	256×192

Every clip's actual w/h/dw/dh are in its own metadata, and in index.csv as rgb_w, rgb_h, depth_w, depth_h, orientation. Read them per clip. A loader that assumes one orientation will transpose every frame of the other.

K is in RGB pixel units, so it cannot be applied to depth pixel coordinates directly. Depth shares the camera's optical geometry but is sampled on a much smaller grid, so before back-projecting you must either map the depth pixel into RGB pixel space, or scale the intrinsics by the resolution ratio:

# map a depth pixel (px, py) into RGB pixel space, then use K unchanged
u = px / (dw - 1) * w
v = py / (dh - 1) * h
X = (u - cx) / fx * Z;  Y = (v - cy) / fy * Z          # Z = depth_m[py, px]

# equivalently, scale the intrinsics once into depth pixel space
fx_d, cx_d = fx * dw / w, cx * dw / w
fy_d, cy_d = fy * dh / h, cy * dh / h

Do not assume depth[y, x] corresponds to rgb[y, x].

Reading depth needs LZFSE (pip install pyliblzfse).

Tasks

sweep 351 clips (1.27 h) · fold_laundry 290 clips (2.62 h) · pick_and_place 230 clips (0.89 h) · scoop_grain 212 clips (1.39 h) · jar_open_close 139 clips (0.83 h) · wash_dishes 107 clips (2.28 h) · plug_charger 68 clips (0.24 h) · tie_shoelaces 63 clips (0.30 h) · pour_liquid 28 clips (0.25 h) · wipe_surface 25 clips (0.21 h)

Recorded by 4 adults, each wearing a chest-mounted iPhone. The participant_id column in index.csv groups clips by wearer.

Devices and capture conditions

device	clips	median depth conf	frame rates present
iPhone 12 Pro	867	87.8%	15 fps (514), 30 fps (323), 60 fps (13), plus 17 at irregular rates
iPhone 17 Pro	646	69.2%	60 fps (646)

Clip length: min 3.4 s · median 16.3 s · max 181.9 s.

Identifiers

id	example	meaning	stable?
`clip_id`	`EM1-0254`	the clip. Assigned once in capture order	permanent
`episode_id`	`EM1-D2-E06`	the continuous recording sitting it belongs to (dataset, shoot day, episode within that day)	grouping only

Split on episode_id, shoot_day, or participant_id rather than on clips. Clips from one sitting are correlated: they share participants, objects, and capture conditions.

Using it

import json, zipfile
import numpy as np, liblzfse          # pip install pyliblzfse
from huggingface_hub import hf_hub_download

p = hf_hub_download("CaryxAI/everyday-manipulation-3d-raw", "clips/D2/EM1-0254.r3d", repo_type="dataset")
z = zipfile.ZipFile(p)
meta = json.loads(z.read("metadata"))

K      = np.array(meta["K"], dtype=np.float32).reshape(3, 3).T   # column-major!
poses  = np.array(meta["poses"], dtype=np.float32)               # (N,7) qx qy qz qw tx ty tz
stamps = np.array(meta["frameTimestamps"], dtype=np.float64)

raw   = liblzfse.decompress(z.read("rgbd/0.depth"))
depth = np.frombuffer(raw, dtype=np.float32).reshape(meta["dh"], meta["dw"])  # metres
conf  = np.frombuffer(liblzfse.decompress(z.read("rgbd/0.conf")), dtype=np.uint8
                      ).reshape(meta["dh"], meta["dw"])

index.csv carries one row per clip with task, participant, episode, duration, frame rate, depth confidence, detected objects, and a description. Use it to select a subset before downloading anything.

Honest limitations

Read this section before you trust a column.

The semantic columns are machine-generated and unreviewed. task, task_phrase, primary_objects, secondary_objects, environment, hands_visible, lighting, and description all come from a vision-language model and have not been checked by a person. Treat them as a searchable first pass rather than ground truth. label_source=vlm marks every row.

Use the fps column, not the fps field inside the bundle. Record3D stamps the configured rate into metadata, and an iPhone 12 Pro will stamp 60 while actually delivering a hard-capped 30. We resolve the true rate from frameTimestamps and publish it in index.csv, so the column is correct even where the bundle's own stamp is not. If you parse metadata["fps"] yourself you get the stamp. On the iPhone 12 Pro clips the stamp is very often exactly double the achieved rate (verified on 532 of 867). Frame rate also varies per clip, so read it per clip and never assume a corpus-wide constant.

Depth confidence is device-relative. The 12 Pro and 17 Pro report substantially different high-confidence fractions on visually similar scenes, so depth_conf_high compares clips within a device, not across devices. It is not a scene-quality score.

Privacy and consent

All 4 participants are project contributors who consented to the recording and to open release. Capture was in private homes. Audio is removed from every published clip.

Every clip in which the automated pass detected a visible face, another person, or a screen or document was then reviewed by a person. 6 clips were withheld from this release as a result; the rest were cleared and are included.

Annotations

This drop has no annotations. A densely annotated sample built on these same clips, keyed by the same clip_id, is published separately at CaryxAI/everyday-manipulation-3d.

That sample aims to be the most densely annotated egocentric manipulation data available. Every frame carries metric depth, 6-DoF camera pose, MANO hand pose, per-hand object contact, hand and object segmentation, and time-aligned language, so a model can be tested against any one of those channels or all of them at once without collecting or labelling anything first. It is packaged as a single multi-episode LeRobot v3.0 dataset.

Want this corpus annotated, or a larger volume captured to spec? Reach out to founders@caryx.ai.

Licence

CC-BY-4.0. Commercial use permitted with attribution. Full terms in LICENSE.

Citation

@misc{em1_2026,
  title        = {Everyday Manipulation 3D (raw RGB-D)},
  author       = {CaryX AI},
  year         = {2026},
  publisher    = {CaryX AI},
  howpublished = {\url{https://caryx.ai}},
  url          = {https://huggingface.co/datasets/CaryxAI/everyday-manipulation-3d-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 the annotated episodes in your browser → · caryx.ai · Reach out founders@caryx.ai