ActTrace/acttrace-ego-home

ActTrace Ego-Home — verified egocentric household-manipulation episodes, and new ones recorded to your task list

Head-mounted iPhone recordings of real household tasks in real Chinese homes. Each episode ships RGB video, LiDAR-measured depth in millimetres, 100 Hz IMU, hand keypoints, a metric-scale 6-DoF head trajectory that has been cross-examined against the gyroscope, verbatim narration and human-confirmed action marks — plus a per-episode verification report you can recompute yourself.

EP003, "fold laundry", 10 s of 71 s. Every panel is a file in the zip; the numbers in panel 6 are in verification/EP003.json. Full 4-panel demo video: demo/ActTrace_demo_M05.mp4; 8-second previews of every episode since v0.4 in previews/.

What you get per episode

One zip per episode (EP{nnn}_{taskID}.zip, 0.1–0.5 GB), containing:

• video_ultrawide.mp4 — 1920×1080 @ 60 fps ultra-wide RGB
• depth_frames/ — 320×180 LiDAR depth @ ~29 Hz, 16-bit PNG, millimetres, with its own intrinsics and distortion table
• imu_200hz.json — 100 Hz attitude, gravity, user acceleration, rotation rate (legacy filename, see limitations)
• vision_keypoints.json — 2×21 hand landmarks per frame with confidence and chirality
• annotations.json — instruction (ZH verbatim + EN), task fields, taskSuccess, human-confirmed action marks with handedness
• frame_metadata.json, clock_sync.json, calibration.json, depth_calibration.json — every stream on one time base, every camera parameter per frame

Alongside the zips, at repository level: trajectories/EPxxx_traj.txt (6-DoF head pose, 30 Hz, TUM format; multiply positions by the episode's s for metres), verification/EPxxx.json (recovered scale, residual curve, accelerometer referee, gyroscope cross-exam r and clock lag), and docs/manifest_*.json (every episode, every verification number, field-level provenance). The exact spec of each column is in the table further down.

Commission episodes to your own task list

The 25 released episodes are a sample and a verification record, not the inventory. What we sell is recording new episodes to a specification: your task list (the 151-code action vocabulary is extensible), real homes, the same column set and file layout, the same per-episode verification report, documented consent and rights chain for every recording. A pilot batch is tens of episodes; larger volumes are recruited through household collectors on the same protocol. Tell us the tasks, the rooms, the number of takes and what you need to verify, and we will come back with a scope and a quote: contact@acttrace.cn.

Also available under commercial terms: licensing of the released episodes for training commercial models, redistribution, exclusive task cells.

Try it in five minutes

1. Download one episode zip (EP003 is a good first pick: r = 0.995, scale grade A, 15 marks, 71 s) plus its trajectories/EP003_traj.txt and verification/EP003.json.
2. Open tools/quickstart.ipynb — it walks the zip, overlays hand keypoints, colour-maps a depth frame, and plots the metric head trajectory.
3. tools/actrace_to_lerobot.py exports LeRobot v2.1 (141-dim observation.state); tools/traj_qc.py and tools/scale_audit.py recompute the verification numbers on this page.

Free for research and evaluation under CC BY-NC 4.0. Cite as below.

---

Release status — v0.5 (2026-09-08)

25 episodes · 17 household tasks · 29.5 minutes · 2 collectors · 25/25 head trajectories IMU-cross-examined and released (r = 0.914–0.998, 5 of them flagged 0.90 ≤ r < 0.97) · per-episode metric scale verified by three independent referees (9 A / 13 B / 3 C)

Every quantity we publish for depth, metric scale and trajectory rotation is either measured by a sensor or verified against one; estimated columns are labelled as such; and every per-episode verification number can be recomputed from the shipped files.

What's new in v0.5

1. +2 episodes from the first collector, produced end-to-end by the new pipeline. EP024 (M36, open parcel; 91.5 s, 15 marks) and EP025 (M73, seal a bowl with cling film — a task new to the dataset; 63.2 s, 9 marks) are the first episodes to go through the production console from upload to package without a hand step: automated pre-check, batch QC, gyroscope cross-exam on the GPU, three-referee metric scale, three release gates. EP024 is a deliberately imperfect take: the collector marked it partial (a pomelo slipped at 49.6 s, recovered at 52.1 s — both moments are in the annotations as failureTimestampSec / recoveryTimestampSec), and its head trajectory ships flagged (r = 0.963). EP025 is clean (r = 0.991, scale grade B).
2. Ten episode packages re-issued (EP014–EP023). Their scene_ref.json — a file introduced by app 0.2 — carried two internal identifier fields (an operator code and a room code with a team prefix) that the v0.4 desensitization pass did not cover. v0.5 re-ships those ten zips with the fields removed or neutralized; nothing else in them changed. The manifest now records scrub_version per episode (2 = current rules) and rescrubbed_in for the re-issued ones. The first thirteen packages were already clean.
3. Annotation semantics stated in the manifest (annotation_provenance). Action marks are single-tap witness timestamps made live by the collector: each tap marks a moment at which the labelled action is in progress — not its onset and not an interval — with an accuracy of about ±1 s. Where an interval is needed, [tap_i, tap_i+1) is a usable approximation. The order of marks is the collector's own, not a template. taskSuccess is the collector's judgement at the end of the take and may be corrected by the admin in the console, with provenance recorded.
4. Still pending from v0.4: the wider-grid scale re-solve for the three low-confidence episodes (EP002, EP015, EP019) has not been done yet; their s remains flagged low-confidence.

What's new in v0.4 (kept for history)

1. +10 episodes from a second collector, all under the look-around protocol. EP014, EP015, EP016, EP017, EP018, EP019, EP020, EP021, EP022, EP023 (8.9 min) add 5 tasks not in v0.3 — wash broccoli (M42), pour milk (M47), stack books (M08), unwrap a parcel (M75), take a drink from the fridge (M16) — plus further takes of wash cups (M01, ×1), beat eggs (M07, ×2). They went through exactly the same chain as the first 13: gyroscope cross-exam, per-episode metric scale with three referees, human review. All 10 ship with a trajectory; 4 of them are flagged (0.90 ≤ r < 0.97: EP016 (r = 0.956), EP018 (r = 0.943), EP019 (r = 0.914), EP022 (r = 0.959)) — the trajectory and scale ship, status says released_with_caveat, and if you need the cleanest rotation, exclude them. The rest clear the clean bar at r = 0.977–0.998.
2. The gyroscope check now measures the clock offset instead of gating on it. The cross-correlation peak is refined to sub-sample precision and r is recomputed at the refined alignment; every episode publishes imu_crosscheck_r, alignment_lag_sec and a three-tier status (released / released_with_caveat / withheld). Observed lags are -0.026 to +0.079 s (median +0.025 s). An earlier internal build gated on |lag| ≤ 0.03 s against a 33.3 ms measurement grid — a rule that admitted only lag = 0 exactly and would have withheld three trajectories with r ≥ 0.99; it was retired before release and the story is in the technical report. The 13 v0.3 trajectories were re-scored the same way (r = 0.979–0.995; recordings unchanged).
3. Desensitized delivery. Collector identifiers and internal storage keys never appear in the release: annotations carry no operator/team fields, the manifest names collectors only by an opaque id (C01, C02), and verification/ records carry no internal object paths.
4. Three low-confidence scales, flagged, not hidden. EP002, EP015, EP019 solve, but with a residual curve that does not rise on one side — and for EP015 and EP019 the minimum sits at the upper edge of the search range (s ≈ 2.0), so their s reads "at least this" rather than a measurement. All are small-translation takes where scale is weakly observable. Their s ships flagged low-confidence; a wider-grid re-solve is planned for v0.5.
5. Technical report. The verification chain, its thresholds and its one documented failure-and-fix are written up in the accompanying technical report (link on this page once posted); the report's per-episode numbers are all recomputable from this repository.
6. Corrections to the data card: the LeRobot observation.state is 141-dimensional (not 159); the LiDAR depth column does not ship a per-pixel confidence map (only the 16-bit millimetre depth, its intrinsics and distortion LUT).

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; 0 = invalid, 65535 = saturated
Head IMU	100 Hz: attitude quaternion, gravity, user acceleration, rotation rate, magnetic field	constant 9.96 ms interval
Hand keypoints (2D)	~29 Hz, 2×21 landmarks + per-point confidence	raw on-device estimates, normalized coords
Head trajectory (6-DoF, metric scale)	TUM, 30 Hz, per-episode scale s, IMU-cross-examined	trajectories/EPxxx_traj.txt × s
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 on one timebase
Annotations	narration (verbatim) + instruction (machine EN, disclosed), task-level fields (incl. taskSuccess ∈ success/partial/fail with failure reason & timestamps when not success), human-confirmed action marks with handedness	151-code vocabulary

Each episode ships as one zip at the repository root: EP{nnn}_{taskID}.zip. Head trajectories live in trajectories/; multiply positions by the episode's s (manifest / verification/EPxxx.json) for meters. docs/manifest_REL-v05-20260908.json lists every episode with its verification results and per-field language provenance. verification/ holds one JSON per episode (recovered scale, residual curve, accelerometer statistics) plus mount_vibration_v04.json / mount_vibration_v05.json.

Episode index

EP	Task	Dur (s)	Marks	Head-traj r	Lag (ms)	Scale s (m per SLAM unit)	Scale grade	Collector
EP001	M01 wash cups	93.2	23	0.980	-26	0.47201	B	C01
EP002	M02 make a drink	57.6	8	0.994	+33	1.11368	C (low conf.)	C01
EP003	M04 fold laundry	71.3	15	0.995	+8	0.99461	A	C01
EP004	M05 cut an apple	110.9	23	0.995	+12	0.91952	B	C01
EP005	M07 beat eggs	70.9	16	0.981	+26	0.02178	A	C01
EP006	M22 wash bowl & chopsticks	81.4	17	0.986	+28	0.04343	B	C01
EP007	M25 boil & pour	133.8	13	0.995	+26	0.85659	B	C01
EP008	M28 fold T-shirts	61.6	15	0.990	+25	0.49444	A	C01
EP009	M36 open parcel	55.2	12	0.987	+14	0.56517	A	C01
EP010	M36 open parcel	86.7	17	0.979	+5	0.72888	B	C01
EP011	M37 change trash bag	56.5	9	0.991	+12	0.69458	B	C01
EP012	M22 wash bowl & chopsticks	108.6	23	0.982	+35	0.58612	B	C01
EP013	M06 clear table after meal	93.1	13	0.992	+4	0.50514	A	C01
EP014	M42 wash broccoli	119.8	11	0.988	+79	0.33832	B	C02
EP015	M07 beat eggs	54.3	11	0.998	+6	1.93017	C (low conf.)	C02
EP016	M47 pour milk	60.7	8	0.956 ⚑	+16	1.40152	A	C02
EP017	M07 beat eggs	60.0	12	0.997	+25	0.42702	B	C02
EP018	M08 stack books	42.1	15	0.943 ⚑	+30	0.12973	A	C02
EP019	M47 pour milk	54.3	9	0.914 ⚑	+34	1.99985	C (low conf.)	C02
EP020	M01 wash cups	36.4	8	0.996	+6	0.64175	B	C02
EP021	M75 unwrap a parcel	28.9	8	0.989	-2	0.61494	B	C02
EP022	M16 take a drink from the fridge	40.2	13	0.959 ⚑	+34	0.36455	B	C02
EP023	M08 stack books	35.2	8	0.977	+30	0.63592	A	C02
EP024 ★	M36 open parcel	91.5	15	0.963 ⚑	+30	0.62926	A	C01
EP025 ★	M73 seal a bowl with cling film	63.2	9	0.991	+0	0.74245	B	C01

★ new in v0.5. ⚑ trajectory shipped flagged (released_with_caveat, 0.90 ≤ r < 0.97). Lag = sub-sample alignment offset of the odometry time base relative to the gyroscope (positive = odometry trails). Chinese verbatim narrations and full per-episode metadata are in the manifest. Grades are defined in SCALE_VERIFICATION_v04.md: A all three referees green; B primary referee green, accelerometer within its own systematic band; C low confidence.

Quality: measured, not promised

Every released episode passes the automated pre-check plus human review; every head trajectory is cross-examined against the gyroscope and shipped under a three-tier rule — released (r ≥ 0.97, |lag| ≤ 0.25 s), released_with_caveat (0.90 ≤ r < 0.97), withheld (r < 0.90) — with its r and lag published whichever tier it lands in; and every shipped trajectory's metric scale goes through the three-referee audit — the same harness runs in the production pipeline, so future episodes ship with the same verification record by default. Across the 25 episodes with a released trajectory the depth-transfer residual is 1.0–4.8% (median), 17 of 25 at or below 2.5%, i.e. at the sensor noise floor; the recovered scales span 0.022–2.000 m per SLAM unit — a 92-fold range, which is why scale is recovered per episode and never assumed. Mount-vibration audit (verification/mount_vibration_v05.json covers the 2 episodes new in v0.5, mount_vibration_v04.json the 10 new in v0.4; the v0.3 episodes were audited with the same script at v0.3 time): gyro spectral energy above 20 Hz is a few percent of voluntary head-motion energy — the motion you see is the human's, not the mount's.

The check has found things before, and we publish them: in this release 5 trajectories ship flagged (EP016 (r = 0.956), EP018 (r = 0.943), EP019 (r = 0.914), EP022 (r = 0.959), EP024 (r = 0.963)) — four from the second collector, for which we do not offer a cause, and EP024 from the first, a fast and partly failed parcel-opening take where the odometry visibly struggles with the motion; earlier, three takes (fold laundry, beat eggs, change trash bag) recorded before the look-around protocol measured r = 0.918–0.951 under the single-bar rule of the time and were superseded by re-recordings that clear the clean bar; and the check caught a defect in its own lag gate before this release shipped. Details in the technical report.

Known limitations (read before training)

1. imu_200hz.json is a legacy filename — actual rate is 100 Hz.
2. Apple Vision hand chirality flips occasionally; per-mark human handUsed is ground truth.
3. Head-trajectory translation is metric only through the published per-episode s (absolute uncertainty ±5%, measured on one calibration session); EP002, EP015, EP019 are low-confidence. Translation has not been compared against motion capture.
4. Episodes begin and end with a look-around of the room (about 5–6 s, paced by a spoken five-count; the closing one precedes the stop). There is no dedicated label: identify these windows by convention from the first/last action marks.
5. Audio is muted wherever speech was detected (privacy by design); mute intervals logged.
6. English instructions are machine-generated (disclosed); narration is verbatim human speech.
7. Fingertip depth bleed (quantified). At 320×180 the depth map is coarse relative to a fingertip. On a 413-frame subset of EP003 (2,820 high-confidence fingertip samples), 16.0% of fingertips read more than 8 cm deeper than their knuckle — background bleed; a 3×3 minimum-pool cuts that to 13.3% and a 5×5 minimum-pool to 11.3%. Treat fingertip points with more caution than wrist and knuckle points; the audit script ships in tools/.
8. Sessions recorded before 2026-08-13 lack device_camera_calibration.json inside the zip (older app build); the device-constant extrinsics used for their verification are included in verification/.
9. 2 collectors, 2 homes, one device model. Sensor characteristics are not sampled across hardware.

Relation to large free corpora

Large open egocentric corpora (Ego4D, Ego-Exo4D, EgoDex, and the recent 100k-hour-scale RGB releases) are excellent pretraining fuel. Episodes like ours occupy a different slot: measured, not reconstructed (LiDAR ToF depth in millimeters; hand poses lifted through calibrated optics and measured depth — not estimated from RGB); verified, not batch-generated (human-confirmed marks; trajectories that must survive an independent sensor cross-exam and a three-referee metric-scale audit — failures are flagged or withheld, and we publish the story); commissionable to spec (your task list, real Chinese homes, same format, per-episode health and verification reports, full documented consent and rights chain). Train on the big corpora; calibrate, evaluate, and ground on measured data. Contact: contact@acttrace.cn.

Tools

tools/ ships with the dataset under the same license: quickstart.ipynb (five-minute tour), actrace_to_lerobot.py (LeRobot v2.1 export; instruction → language_instruction, a 141-dim observation.state = head quaternion 4 + gravity 3 + user acceleration 3 + rotation rate 3 + two hands × 21 × (u, v, conf) + detected-hand count + live mark index), traj_qc.py (gyroscope cross-exam), scale_audit.py (depth-transfer scale recovery + accelerometer referee), shake_qc.py (mount vibration), fingertip_audit.py. Note on LeRobot versions: the export is v2.1 format; lerobot ≥ 0.3 loads it after the official convert_dataset_v21_to_v30, or use lerobot < 0.3 directly.

Privacy & consent

No person other than the collector appears in any released episode, and the collector appears only as hands and forearms; an episode in which another person enters the frame is rejected outright. On-device real-time face detection halts recording if a face enters frame; speech is muted at source; server-side QC re-checks with human review. Each collector recorded in their own home under a signed data-collection agreement (copyright assignment + cross-border transfer consent, reviewed by counsel). The release carries no collector identifiers.

License & commercial use

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@acttrace.cn.

中文说明

行迹所至（ActTrace）v0.5：25 集、17 个家庭任务、29.5 分钟，2 位采集者各自在自己家中用头戴 iPhone 采集，多列同步交付。已发布的 25 集是样品和验收记录，不是库存：我们做的是按客户的任务清单在真实家庭新录、逐集验证、同一格式交付，试点几十集起，更大规模由家庭采集者按同一协议完成；说明任务、房间、条数和验收要求，联系 contact@acttrace.cn。深度是激光雷达实测的毫米值；头部旋转用手机陀螺仪逐集对账并分三档发布（r ≥ 0.97 正常发布；0.90 ≤ r < 0.97 带标记发布，status 为 released_with_caveat；r < 0.90 不发布轨迹），每集的 r 与时间对齐偏移都如实公布，本版 EP016、EP018、EP019、EP022、EP024 带标记发布；单目轨迹的米制尺度用每集自己的实测深度恢复，并经三名独立裁判验收（深度搬运残差、加速度计交叉验证、A4 实物标定链审计），逐集评级 9 A / 13 B / 3 C，低置信的三集如实标出。v0.4 起交付物脱敏：不含采集者身份与内部存储路径；v0.5 重新发布了 EP014–EP023 十个包（其 scene_ref.json 里此前残留两个内部编号字段，已去除），并在装箱单里写明打点的语义：打点是采集者当场轻点的"动作正在发生"的时刻，不是起点也不是区间，精度约 ±1 秒。逐集验收 JSON 与全部检查脚本随数据发布，欢迎复验。科研免费（CC BY-NC 4.0），商用与定制采集联系 contact@acttrace.cn。

Citation

@misc{acttrace2026egohome,
  title  = {ActTrace Ego-Home: A Verification-First Egocentric Household-Manipulation Dataset with Measured Depth, Metric-Scale Head Trajectories and Disclosed Annotation Provenance},
  author = {Tang, Casey},
  year   = {2026},
  url    = {https://huggingface.co/datasets/ActTrace/acttrace-ego-home}
}