1,000-hour dataset

Multi-camera egocentric human-demonstration video for robot learning · four-camera rig.

2026.07.02 · External

1,121 hours of multi-camera egocentric human-demonstration video. Every episode is captured on a synchronized four-camera rig — egocentric, exocentric, and both wrists — and ships with action-level annotations on a single 29.97 fps timeline.

At a glance

Headline figures for the 1,000-hour dataset. Hours are wall-clock; frames and annotations track the 29.97 fps capture rate; size is set by the aggregate bitrate.

1,121

Hours

Wall-clock, across all four camera streams

121M

Frames

29.97 fps · 960 × 720

655,192

Action annotations

Action-level labels across the dataset

15.7 TB

Total size

H.264 · ~32 Mbps aggregate (4 × 8 Mbps) over ~1,000 h

1,534 episodes (mean 43.9 minutes, continuous single-session recordings), each carrying action annotations on the same 29.97 fps timeline as the video. Captured 2026-05-12 → 2026-06-18 across 232 scripted tasks at 9 venues.

Sample diversity

Planned target hours across 8 environments × 35 action categories. Row = action; column = environment. Color scales with √(hours).

0 h

78 h (max)

Action	Home	Food & Bev.	Office	Hotel	Edu.	Farm	Retail	Construction	Total
Pick & Place	78	24	12	8	8	8	4	12	154
Open & Close	59	16	12	12			4		103
Grasp & Release	47	12	8	8			8		83
Push & Pull	51	8						8	67
Dig & Rake	12				12	8			32
Scan & Tag		8	8	12			4		32
Push Pull Drag	8				8	12			28
Wipe & Sanitize		8	8	8			4		28
Pour & Dispense		8	8	8			4		28
Pour & Transfer	16	8							24
Assemble & Dis.		4	8	8			4	12	36
Cut & Prepare		4	8	8			4	4	28
Fold & Wrap		4	8	8			4		24
Insert & Connect		4	8	8			4		24
Inspect & Measure		4	8	8			4	8	32
Load & Unload		4	8	8			4	4	28
Operate & Equip.		4	8	8			4		24
Sort & Organize		4	8	8			4		24
Rotate & Turn	12	8						8	28
Mix & Stir	12	8							20
Fold & Arrange	12	8							20
Wipe & Clean	12	8							20
Insert & Extract	12	8							20
Toggle & Switch	12	4							16
Press & Squeeze	12	4							16
Cut & Separate	12	4							16
Fill & Empty	12	4							16
Sweep & Collect	8				4	4			16
Adjust & Position	4				4	4			12
Carry & Transport	4				4	4		8	20
Cover & Uncover	4				4	4			12
Cut & Trim	4				4	4			12
Plant & Harvest	4				4	4			12
Spray & Water	4				4	4			12
Tie & Secure	4				4	4		4	16

Camera coverage

Four synchronized camera streams on every episode — head, third-person, and both wrists.

Every episode is captured on the full four-camera rig — egocentric (E), exocentric (X), left wrist (L), and right wrist (R) — for complete multi-view coverage on every recording.

E — Egocentric: Head-mounted

X — Exocentric: Fixed third-person

L — Left wrist: Wrist-mounted

R — Right wrist: Wrist-mounted

Technical specifications

Capture and annotation share one 29.97 fps timeline, so video and labels stay frame-aligned.

Dataset format: LeRobot v3.0

Resolution: 960 × 720

Frame rate: 29.97 fps

Bitrate (per camera): ~8 Mbps

Video codec: H.264 · yuv420p

Annotations: Action spans · free-text

Sample structure & metadata fields

How each sample is laid out on disk as a LeRobot v3.0 dataset, and the per-frame fields delivered with it.

Each sample is a LeRobot v3.0 dataset — a metadata folder, per-frame records in Parquet, and one H.264 video per camera, all on a shared 29.97 fps timeline.

AOn-disk layout

// LeRobot v3.0 dataset

dataset/
├── meta/
│   ├── info.json              # format · fps · features · camera roles
│   ├── tasks                  # task_index → task text
│   ├── episodes/              # per-episode metadata + stats
│   ├── custom_metadata.csv    # metadata
│   └── custom_annotation.json # annotation
├── data/
│   └── chunk-000/file-000000.parquet        # per-frame records
└── videos/
    └── {camera}/chunk-000/file-000000.mp4   # H.264, one per camera

BPer-frame fields · data/…parquet

**Table.** Columns in each per-frame Parquet record.
Field	Type	Description
index	int64	Global frame index across the dataset
episode_index	int64	Episode this frame belongs to
frame_index	int64	Frame position within the episode
timestamp	float64	Seconds from episode start (29.97 fps)
task_index	int64	Task label · resolves via meta/tasks

CVideo streams · videos/…mp4

Four synchronized RGB streams per episode, each H.264 at 960 × 720: base_0 (egocentric), base_1 (left wrist), base_2 (right wrist), base_3 (exocentric).

Sample views

The four camera views — one representative frame each, from sample episode 000001.

**Fig.** Egocentric view — episode 000001.

**Fig.** Exocentric view — episode 000001.

**Fig.** Left-wrist view — episode 000001.

**Fig.** Right-wrist view — episode 000001.

Sample annotation

Each episode is segmented into consecutive action spans, every span carrying a free-text label. Shown: the first ten spans of an actual annotation.

Annotations divide each episode into consecutive action spans, every span carrying a natural-language label. Below: the first ten spans of an actual episode annotation, verbatim.

{
  "episodes": [{
    "episode_id": "0",
    "spans": [
      { "start_time": 0.0,  "end_time": 2.3,  "label": "spray liquid from spray bottle over countertop and sink and target condition sprayed" },
      { "start_time": 2.3,  "end_time": 6.3,  "label": "pick knife handle from countertop and place knife on countertop" },
      { "start_time": 6.3,  "end_time": 8.6,  "label": "place knife blade into knife block and pick knife handle from countertop and place knife handle into knife block" },
      { "start_time": 8.6,  "end_time": 11.1, "label": "pick cleaver from knife holder and place cleaver on countertop" },
      { "start_time": 11.1, "end_time": 13.1, "label": "pick cleaver from knife block and place cleaver on countertop" },
      { "start_time": 13.1, "end_time": 16.1, "label": "pick knife handle from knife block and place knife handle on countertop" },
      { "start_time": 16.1, "end_time": 19.7, "label": "place knives on bottom of sink and pick knives from countertop and place knives on bottom of sink" },
      { "start_time": 19.7, "end_time": 23.7, "label": "pick up sponge from sink and place sponge on countertop" },
      { "start_time": 23.7, "end_time": 38.5, "label": "wring cloth over sink until dry" },
      { "start_time": 38.5, "end_time": 40.5, "label": "pick tissue paper from countertop and place tissue paper on countertop" },
      ...
    ]
  }]
}

Legal & licensing

Rights, consent, and permitted use for this delivery.

Rights & consent. All recordings were captured by consented collectors under written releases covering AI/ML training and commercial deployment. Synjuku holds all rights necessary to license this data; no third-party platform or scraped content is included.

License. Provided under the terms of the applicable Synjuku data delivery agreement. Use is limited to the permitted purposes defined there; the dataset may not be redistributed, resold, or made available to third parties on a standalone basis without Synjuku's prior written consent.

Confidentiality. This datasheet and the dataset it describes are confidential commercial materials. Provenance records (task card, capture site, recording date per episode) are retained by Synjuku and support audit on request.

Synjuku

1,000-hour dataset · 2026.07.02