Access to the Advanced Edition is reviewed manually, usually within 3 business days. The data is licensed CC BY-NC 4.0: free for non-commercial research, commercial use requires a separate license from OpenElephant Intelligence (jiuchen@openelephant.ai).

Pico Egocentric Dataset · Advanced Edition

First-person stereo video, depth, head pose, pose-aligned point clouds and time-stamped action captions of real shelf-work and facility tasks, captured on a Pico VR headset.

By OpenElephant Intelligence (公象智能), Beijing.

The Advanced Edition adds 3D geometry and language to the Basic Edition: every episode comes with world-aligned point clouds, a ready-to-view MCAP recording, and a sequence of time-stamped captions describing what the operator is doing. It is a sample of a 10,000+ hour commercial collection.

🔒 Gated. Requests are reviewed manually, usually within 3 business days.

At a glance


Episodes	1,203 across 21 tasks in 2 environments
Device	Pico VR headset, head-mounted, operator's natural viewpoint
Video	Stereo RGB, 1920 × 1456, 30 fps, undistorted (H.264 MP4)
Depth	Per-frame depth from the headset depth sensor (320 × 240)
Head pose	6-DoF headset pose at ~1 kHz
Point clouds	Per-frame clouds aligned to head pose (`aligned_pointclouds.npz`)
MCAP	Head pose and point clouds in one file, viewable in Foxglove
Captions	Time-stamped natural-language action captions (Chinese)
Audio	Stereo AAC + 4-channel Opus
Calibration	Per-episode intrinsics, camera-to-IMU extrinsics, stereo baseline (~64 mm)
License	CC BY-NC 4.0 (commercial license available)

Editions

Edition	Contents	Access
Basic	Stereo video + depth + 6-DoF head pose + audio + calibration	🟢 Instant approval
Advanced (this repo)	Everything in Basic + pose-aligned point clouds + MCAP recordings + time-stamped action captions	🔒 Reviewed
Commercial	10,000+ hours, MANO hand reconstruction, action segmentation with skill labels, LeRobot export, custom collection	✉️ jiuchen@openelephant.ai

Environments and tasks

`hengxuan-commercial-bank-retail-area` — retail shelves · 879 episodes

Real store shelving with beverage coolers, snack gondolas and hanging snack racks. Most tasks are "error correction": the shelf starts out mis-stocked and the operator finds and fixes the misplaced items.

Task folder	What the operator does	Episodes
`supplement-hanging-melon-seed-snacks`	Restock a hanging rack of melon-seed snacks	145
`organize-and-replenish-snacks`	Tidy and restock a snack shelf	126
`snack-correction-mixed-pickled-mustard-and-kelp-shreds`	Separate mixed pickled-mustard and kelp-shred packs	121
`correction-of-shelf-snacks-different-categories-of-duck-products-are-mixed`	Sort mixed duck-snack products back to their slots	120
`整理袋-桶装薯片`	Arrange bagged and canned potato chips	103
`correcting-shelf-snacks-mixing-different-categories-of-meatballs`	Sort mixed meatball-snack categories	81
`correcting-shelf-snacks-mixing-different-types-of-pickled-mustard`	Sort mixed pickled-mustard varieties	58
`beverage-error-correction-mixing-different-tea-drinks`	Sort mixed tea drinks in a cooler	45
`correction-of-hanging-snacks-mixed-peanut-flavors`	Sort hanging peanut snacks by flavor	23
`correction-of-hanging-snacks-mixed-hanging-of-different-brands-of-melon-seeds`	Sort hanging melon seeds by brand	20
`beverage-error-correction-mixing-different-brands-of-beer`	Sort mixed beer brands	11
`beverage-error-correction-mixing-different-brands-of-soda-water`	Sort mixed soda-water brands	10
`clean-up-and-replenish-mineral-water`	Clear and restock bottled water	9
`organize-hanging-melon-seed-snacks`	Tidy a hanging melon-seed rack	5
`correction-of-hanging-snacks-mixed-rice-crackers-with-different-flavors`	Sort hanging rice crackers by flavor	1
`snack-correction-mixing-different-types-of-pickled-mustard`	Sort mixed pickled-mustard varieties	1

`一楼体育馆自采` — indoor gymnasium · 324 episodes

A large indoor sports hall with fitness equipment, mats and event cabling.

Task folder	What the operator does	Episodes
`整理收拾一半的电线`	Finish coiling partially tidied cables	107
`收纳瑜伽砖`	Collect and stack yoga blocks	84
`整理电线`	Coil and tidy loose cables	78
`整理收纳瑜伽垫`	Roll up and store yoga mats	39
`摆好普拉提圈`	Lay out Pilates rings	16

File layout

<environment>/<task>/case_<id>/
├── head_left_camera_undistorted.mp4     # left RGB, 1920×1456 @ 30 fps, undistorted
├── head_right_camera_undistorted.mp4    # right RGB, same
├── video_index.json                     # per-frame timestamps for both videos
├── undistort_camera.json                # original + rectified intrinsics (K, D)
├── spatial_calibration.json             # intrinsics, extrinsics, depth format, time base
├── video_stream_depth_frames.npz        # depth frames
├── video_stream_head_pose_6dof.npz      # 6-DoF head pose, ~1 kHz
├── video_stream_audio_aac_stereo.m4a    # stereo audio
├── video_stream_audio_opus_4ch.ogg      # 4-channel audio
├── aligned_pointclouds.npz              # ← Advanced: per-frame point clouds aligned to head pose
├── head_pose_point_cloud.mcap           # ← Advanced: head pose + point clouds, Foxglove schemas
└── segments.json                        # ← Advanced: time-stamped action captions

Earlier captures. 218 retail episodes recorded with an earlier pipeline (case IDs starting 15171–15190) do not include spatial_calibration.json, depth, the head-pose .npz or audio. They do include the two videos, undistort_camera.json, video_index.json, the point clouds, the MCAP recording and the captions.

`segments.json` — action captions

A list of contiguous, time-stamped segments covering the whole episode. Times are in seconds from the start of the video; captions are free-form Chinese descriptions of the operator's action.

[
  { "start": 2.2,  "end": 4.3,  "caption": "从纸箱中拿起黑色桶装薯片" },
  { "start": 4.3,  "end": 8.6,  "caption": "将黑色桶装薯片摆上货架" },
  { "start": 25.0, "end": 32.0, "caption": "调整三罐薯片排列使其对齐" }
]

("Pick up the black canned chips from the carton" → "Place the black canned chips on the shelf" → "Align the three cans of chips.")

In later recordings the operator makes a two-handed "OK" gesture at the start and end as a recording marker; those segments are captioned as such and can be trimmed.

Calibration

spatial_calibration.json gives, for the left and right RGB cameras and the depth camera:

intrinsics (fx, fy, cx, cy) and distortion (equiDis62 fisheye for RGB, Brown for depth)
extrinsics_4x4: T_IMU←camera, mapping the OpenCV camera frame to the OpenXR head / IMU frame
depth.format: encoding of the depth values
timebase_us: common time base (µs) shared by RGB, depth and pose streams
baseline_mm: stereo baseline

undistort_camera.json holds the original (K_original, D) and rectified (K_new) intrinsics used to produce the undistorted videos. Use K_new with the MP4 files.

Quick start

from huggingface_hub import snapshot_download

# Download one task (roughly 0.5–2.5 GB per episode in this edition)
path = snapshot_download(
    repo_id="skycn110/pico-robotics-advanced",
    repo_type="dataset",
    allow_patterns="一楼体育馆自采/摆好普拉提圈/*",
)

Read the captions alongside the video:

import json, glob, cv2

case = sorted(glob.glob(f"{path}/一楼体育馆自采/摆好普拉提圈/case_*"))[0]
segments = json.load(open(f"{case}/segments.json"))

cap = cv2.VideoCapture(f"{case}/head_left_camera_undistorted.mp4")
fps = cap.get(cv2.CAP_PROP_FPS)

for seg in segments:
    print(f"{seg['start']:6.1f}–{seg['end']:6.1f}s  {seg['caption']}")
    cap.set(cv2.CAP_PROP_POS_FRAMES, int(seg["start"] * fps))
    ok, frame = cap.read()                      # first frame of the segment

Open head_pose_point_cloud.mcap directly in Foxglove to scrub through the head trajectory and point clouds together.

Intended uses

Egocentric world models and action-conditioned video prediction
Video–language grounding and dense captioning of first-person manipulation
3D scene understanding and egocentric reconstruction from a moving head
Visual odometry, head-motion forecasting, and stereo / monocular depth estimation
Studying human shelf-work and tidying behaviour as a source for robot imitation

Limitations

Two environments only; task counts are uneven (several tasks have fewer than 10 episodes).
Captions are free-form Chinese text, not a closed label set; segment boundaries are approximate.
No hand pose in this edition — MANO hand reconstruction is part of the Commercial edition.
218 earlier retail episodes lack depth, the head-pose .npz and audio (see Earlier captures above).

License

CC BY-NC 4.0. Free for research and other non-commercial use with attribution. Training or evaluating models for commercial products requires a commercial license — contact jiuchen@openelephant.ai.

Academic groups working toward a publication can ask us about a research license for larger samples.

Citation

@misc{openelephant_pico_egocentric_2026,
  title        = {Pico Egocentric Dataset},
  author       = {{OpenElephant Intelligence}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/skycn110/pico-robotics-advanced}}
}

Contact

Jiuchen, OpenElephant Intelligence · jiuchen@openelephant.ai

Commercial licensing, larger samples, custom collection and research collaborations are all welcome. Questions about the data can also go in this repository's Discussions tab.

Downloads last month: 4,752

Total file size:

1.95 TB

You need to agree to share your Hugging Face information to access this dataset