EgoSuite-Open100K

Lightwheel's egocentric human-demonstration release: stereo head video with synchronized 3D hand and body pose, plus wrist cameras on the EgoPro line. Announced as 100,000 hours; roughly 10,000 hours are published so far across ~362,000 episodes.

EgoSuite-Open100K is a large egocentric corpus of people performing everyday work tasks, captured on a head rig with stereo cameras and — on the EgoPro line — a camera on each wrist. Recordings carry 3D hand pose, full-body pose on the body subsets, and subtask-level semantic segments, and ship in LeRobot v3 and MCAP.

Read the hours claim carefully. The project page reports "Total Hours: 100,000", but every duration table in the release is headed Planned Duration:

The published column was measured, not quoted. Both Buckets were enumerated in full and 750 MCAP episodes sampled; each episode's exact span comes from its MCAP `Statistics` record. Two independent estimators — a ratio estimator over every file's known size, and mean duration × episode count — give ~10,019 h and ~9,672 h. With EgoDemo that is roughly 10,000–10,100 published hours, about a tenth of the announced total. The site itself calls this "the first batch".

One caveat on that table: the EgoPro Bucket has no `EgoProStandard-body` directory. It carries `EgoProStandard` (25,418 episodes) and a second directory named `EgoStandard-body` (5,892 episodes, ~200 h) — a name belonging to the other SKU, and distinct from the `EgoStandard-body` in the EgoStandard Bucket. The row above assumes that directory is the body subset of EgoPro; if it is instead misfiled EgoStandard data, EgoPro's published body hours are zero. The Bucket's own README tells you to `sync .../EgoPro/EgoStandard-body`, so the name is deliberate rather than a listing artefact.

Counting traps. Three of them, all of which inflate naive totals:

• `EgoStandard/rd/` is a complete duplicate — all 47,708 of its episodes already appear under `EgoStandard/mcap/`, adding 4.02 TiB and ~816 h of nothing new.
• LeRobot and MCAP are alternative encodings of the same episodes; the cards say explicitly they "must not be counted as additional hours".
• In the EgoDemo repo, the two raw-video subsets (`EgoRaw`, `EgoProRaw`) are entirely re-releases of episodes already in the annotated subsets, and a leftover `.hf-transfer/` staging tree mirrors all 1,960 episodes again. Summing its six subsets gives 2,712 episode-slots for 1,960 real episodes.

After de-duplication: 361,995 unique episodes — 360,035 across the two Buckets plus EgoDemo's 1,960, which are disjoint from them — and 6,184 distinct task names, against a claimed 15,000+ tasks and 15,000+ scenes for the full programme.

Where the data lives. This matters because the obvious download does not work. The `EgoStandard` and `EgoPro` dataset repos contain only a README — the bulk data is in S3-compatible HF Buckets (`hf://buckets/LightwheelAI/...`), 59.35 TiB across 407,949 objects. Only `EgoDemo` stores data in the repository itself. All three repos are manually gated; Bucket objects, in contrast, are currently readable without authentication.

Per episode. Four `foxglove.CompressedVideo` h264 channels on EgoPro (head_left, head_right, left_wrist, right_wrist) at 1920×1456 for the head pair and 1920×1536 for the wrists, 30 fps, alongside protobuf pose topics (`/pose/head`, `/pose/headcam`, `/pose/left_hand`, `/pose/right_hand`), camera intrinsics and extrinsics, `/annotation/semantic_segments`, and a `/annotation/bad_frame/pose/hand` quality channel. Task labels are directory names, with a mix of English and Chinese/Thai strings.

Recordings were de-identified with automated blurring of faces, licence plates and other PII plus human verification, and participants consented to controlled release. The licence is a custom `commercial-training-no-resale-v1.0`: open for academic research and commercial training, but not for resale.