Head-mounted iPhone (ultra-wide RGB + LiDAR) recordings of real household manipulation tasks in Chinese homes, built for robot-learning teams that need more than "loose video": every episode ships with metric depth, head IMU, dual-hand 2D keypoints, per-frame camera metadata, dual-clock sync anchors, and human-confirmed action segments with handedness labels.

▶ Watch the 4-panel demo video (demo/ActTrace_demo_M05.mp4): RGB + hand skeletons | LiDAR depth | 100 Hz IMU | live action-segment timeline. (The keypoint overlay is drawn directly from the shipped raw keypoints — no smoothing, no post-processing. A second demo of task M25 is also in demo/.)

Why this exists

The community has large corpora of unlabeled or lightly-labeled egocentric video (great pretraining fuel), and small corpora of teleoperated robot data (great finetuning fuel). The middle of the data pyramid — human video that is dense enough to supervise skill boundaries, handedness, and metric geometry — is thin. ActTrace episodes are recorded and annotated on-site by a trained annotator ("team leader") who watched the task happen, not by a crowd worker guessing from pixels afterwards.

What's in every episode

Column	Spec	Notes
RGB video	1920×1080 @ 60 fps, H.265, ultra-wide	rolling-shutter readout time provided
LiDAR depth	320×180 @ ~29 Hz, 16-bit PNG, millimeters	own intrinsics + distortion LUT, timestamp in filename
Head IMU	100 Hz: attitude quaternion, gravity, user acceleration, rotation rate, magnetic field	constant 9.96 ms sampling interval
Hand keypoints	~29 Hz, 2×21 2D landmarks + per-point confidence (Apple Vision)	normalized image coords
Per-frame camera metadata	pts, exposure, ISO, white balance, lens position, intrinsic matrix	every RGB frame
Clock sync	host-clock ↔ monotonic-raw anchor file	all streams reduce to one timebase
Annotations	task instruction (natural speech), task-level fields (bimanual, success, materials, environment, lighting), N action segments with label + handedness, human-confirmed	151-code action vocabulary

Each episode ships as one zip at the repository root: EP{nnn}_{taskID}.zip. Inside:

video_ultrawide.mp4        calibration.json         imu_200hz.json (actual rate: 100 Hz)
depth_frames/*.png         depth_calibration.json   vision_keypoints.json
frame_metadata.json        clock_sync.json          annotations.json
audio_mute_log.json  (voice segments were muted at recording time for privacy)

docs/manifest_DLV-20260811-02.json maps EP numbers to source session IDs for full traceability.

Episode index

EP	Task	Dur (s)	Segs	Instruction (zh)	Instruction (en)
EP001	M01 wash cups	93.2	23	洗了一个塑料杯、一个陶瓷马克杯和两个玻璃杯，都是普通冲一下就可以的那种干净杯子	Washed a plastic cup, a ceramic mug and two glass cups — all only lightly used, needing just a rinse
EP002	M02 make a drink	57.6	8	在餐桌上用烧水壶里的水冲了一碗紫菜虾皮汤	Made a bowl of seaweed and dried-shrimp soup at the dining table with hot water from the kettle
EP003	M04 fold laundry	73.3	17	叠了一条毛巾、一条白色裤子和一件T恤；白色裤子自带裤线，需按裤线的走向对齐折叠，因此动作有反复 ★	Folded a towel, a pair of white trousers and a T-shirt; the trousers have pressed creases, so folds had to be aligned along the crease lines, requiring repeated adjustments
EP004	M05 cut an apple	110.9	23	在水果菜板上将一个苹果切成四份，然后摆盘，然后洗水果刀，再把水果菜板洗干净放回去	Cut an apple into four pieces on a cutting board, plated them, then washed the fruit knife and board and put them back
EP005	M07 beat eggs	77.0	15	从冰箱取出两个鸡蛋，然后用筷子在小碗当中打发	Took two eggs from the fridge and beat them in a small bowl with chopsticks
EP006	M22 wash dishes	81.4	17	洗了一个碗和一双筷子	Washed one bowl and one pair of chopsticks
EP007	M25 boil & pour	133.8	13	用水壶在水龙头接了一些水之后再把它烧开，然后倒在一个白色的杯子里 ★	Filled the kettle at the tap, boiled the water, then poured it into a white cup
EP008	M28 fold T-shirts	61.6	15	叠三件T恤后放回柜子里	Folded three T-shirts and put them back into the wardrobe
EP009	M36 open parcel	55.2	12	从门口取到一个新的快递鞋子，拿到乒乓球桌上拆开	Picked up a newly delivered shoe parcel at the door, carried it to the ping-pong table and opened it
EP010	M36 open parcel	86.7	17	用剪刀割开纸盒上的胶带，取出里面的葡萄，然后把纸盒压扁 ★	Cut the tape on the box with scissors, took out the grapes inside, then flattened the box
EP011	M37 change trash bag	39.5	10	在客厅把还没装满的蓝色垃圾袋换下来，换上另一个蓝色垃圾袋 ★	Replaced the not-yet-full blue trash bag in the living room with a fresh blue one

★ = instruction proofread by the collector: the raw annotations.json preserves the original voice-input transcription (which contains truncations/typos for these episodes); the table above is the authoritative corrected text. English translations by the collector.

Quality: measured, not promised

Every released episode passes a 7-gate automated QC pipeline plus human review. Sample health-check figures for episode ...C5A260C8 (task M25, electric-kettle: fetch → fill at tap → boil → pour into cup, 133.8 s, 13 segments):

Video: 8022 frames, 59.98 fps, 0 dropped frames (max inter-frame gap 16.70 ms)
IMU: 13435 samples, 0 gaps (median = p99 = max interval 9.96 ms)
Depth: 3851 frames covering full duration, 0% invalid pixels — all 3851 frames exhaustively checked, values 0.14–5.2 m
Hand detection: 99.8% of manipulation-phase frames have ≥1 hand, 78.3% both hands
Cross-stream alignment: all four streams within 68 ms at start on one timebase, no drift
The full health-check report (docs/health_report_C5A260C8.html) is included

Cross-check of all 11 episodes: constant 60.0 fps throughout; max video inter-frame gap = 16.7 ms (exactly one frame period, i.e. zero drops) in 10 of 11 episodes — EP006 contains a single 33.3 ms gap, i.e. exactly 1 dropped frame across 46,834 total frames; IMU max sampling interval 10.0 ms in every episode; depth and keypoint streams complete in every episode. We report the one dropped frame because "measured, not promised" has to cut both ways.

Known limitations (read before training)

imu_200hz.json is a legacy filename — actual rate is 100 Hz (plenty for manipulation; robot control loops run 30–50 Hz).
Apple Vision hand chirality flips occasionally. 2D landmark positions are reliable; per-segment handUsed in annotations.json is human ground truth — trust that.
Four episodes' raw voice-input instructions contain truncations/typos — corrected text is in the Episode index above; raw files keep the original transcription untouched (traceability over cosmetics).
2D keypoints only in v0.1. 3D hand recovery (e.g. via HaMeR-class models + our metric depth) is on the roadmap.
Audio is intentionally muted wherever speech was detected (privacy by design); ambient task sounds (water, appliance clicks) are preserved outside muted intervals. Mute intervals are logged in audio_mute_log.json.
Camera extrinsics (6-DoF head trajectory) are not shipped in v0.1; IMU + rolling-shutter + per-frame intrinsics support offline recovery.

Relation to large free corpora

Large open egocentric corpora (Ego4D, Ego-Exo4D, EgoDex, EgoLive, …) are excellent pretraining fuel, and we assume you are already using them. Episodes like ours occupy a different slot in the stack:

Measured, not reconstructed: depth here is LiDAR time-of-flight in millimeters with its own calibration — not model-estimated from stereo or monocular video. Metric scale cannot be recovered post-hoc; it either was measured at capture time or it does not exist.
Verified, not batch-generated: every action segment was confirmed by a human annotator who watched the task happen, with per-segment handedness; every released episode carries an exhaustive health check. We publish the failure counts (see the single dropped frame above) — that is the level of guarantee this format is built for.
Commissionable to spec: the same capture-and-QC stack that produced these episodes takes commissions — your task list, your object/environment constraints, fresh recordings with full documented consent and rights chain, delivered with per-episode health reports. That, rather than generic volume, is what this pipeline is for. Contact: contact@acttrace.cn.

LeRobot

tools/actrace_to_lerobot.py converts an episode to LeRobot v2.1 (parquet + meta + mp4). The 159-dim observation.state packs head quaternion/gravity/acceleration/rotation-rate, both hands' 21×(u,v,conf), detected-hand count, and the live segment index. Load-tested: 4013 frames @30 Hz, random batch fetch works out of the box.

Privacy & consent

All v0.1 episodes were recorded by the dataset author in their own home. The capture app performs on-device real-time face detection (recording halts if a face enters frame), speech is muted at source, and server-side QC re-checks faces with human review. Operator IDs are pseudonyms. Signed data-collection agreements (copyright assignment + cross-border transfer consent, reviewed by counsel) are on file for all contributors; the same paperwork governs future multi-contributor releases.

License & commercial use

Released under CC BY-NC 4.0 — free for research and evaluation.

For commercial licensing (training commercial models, redistribution, larger volumes, custom task lists, exclusive cells): contact contact@acttrace.cn. Custom collection across many distinct homes/operators in China, with this same 7-column format and per-episode health reports, is available.

中文说明

行迹所至（ActTrace）：头戴 iPhone 采集的家庭操作任务第一视角数据。每条包含 RGB 视频、LiDAR 毫米级深度、头部 IMU、双手 21 点、逐帧相机元数据、双时钟对齐锚点、以及现场标注员人工确认的动作分段（含左右手标记）。全部通过七道自动质检 + 人工复核。科研免费（CC BY-NC 4.0），商用请联系 contact@acttrace.cn。可承接指定任务清单、多家庭多操作员的定制采集。

Citation

@misc{acttrace2026egohome,
  title  = {ActTrace Ego-Home: Annotated Egocentric Household Manipulation with Metric Depth},
  author = {ActTrace},
  year   = {2026},
  url    = {https://huggingface.co/datasets/acttrace/acttrace-ego-home}
}

Downloads last month: 79

Total file size:

3.44 GB