Datasets:

zeno-labs
/

sample-gopro-v1

ZenO Labs builds an egocentric data pipeline for training Physical AI and robot learning policies. Head-mounted GoPro / smartphone footage is processed into LeRobot v2.1 datasets with multiple modalities aligned frame-by-frame:

6-DoF head trajectory (visual-inertial SLAM)
Hand pose — 21 keypoints × 2 hands, with temporal post-processing
Depth — per-pixel depth maps as video
Action segments (where annotated)

This release is a free sample — a small preview of what we collect. For larger / customized datasets we offer:

Longer recordings + more episodes per domain (cooking, assembly, laundry, dishwashing, bartending, …)
Custom task scenarios on demand
Higher framerate (60 fps), 4K source, synchronized multi-camera
Stricter QA + curation tier

Quick use

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("zeno-labs/sample-gopro-v1")
# Or after local download:
# ds = LeRobotDataset.from_root("./sample-gopro-v1")

Modalities

head trajectory (6-DoF, x,y,z + quaternion)
hand pose (21 keypoints × 2 hands)
depth (per-pixel depth map video)

Hand pose — how to access

Each frame stores 21 MediaPipe keypoints per hand in two separate parquet columns. These are not embedded in the videos — they are numeric data.

Column	Shape	Meaning
`observation.hand_pose.left`	`[21, 3]` (float32)	21 keypoints, [x_pixel, y_pixel, z_relative]
`observation.hand_pose.right`	`[21, 3]` (float32)	same, right hand

x_pixel, y_pixel: pixel coordinates in the source video resolution (e.g. 1920×1080 for GoPro Hero 9 wide). Already aligned to observation.images.head frame.
z_relative: MediaPipe's relative depth (~ −1..+1, not metric). Smaller = closer to camera. Use observation.images.depth if you need metric depth.
Missing detection → all 63 values are 0.0 (not NaN). Detect with np.allclose(kp, 0) per frame.

Keypoint order follows the standard MediaPipe Hands topology (0=wrist, 4=thumb tip, 8=index tip, 12=middle tip, 16=ring tip, 20=pinky tip).

Quick visualize (per-frame overlay)

import cv2, numpy as np, pyarrow.parquet as pq

t = pq.read_table(
    "https://huggingface.co/datasets/zeno-labs/sample-gopro-v1/resolve/main/data/chunk-000/episode_000000.parquet"
)
df = t.to_pandas()

# pick a frame
i = 100
left  = np.asarray(df["observation.hand_pose.left"].iloc[i],  dtype=np.float32)  # (21, 3)
right = np.asarray(df["observation.hand_pose.right"].iloc[i], dtype=np.float32)

img = cv2.imread("frame_100.png")  # extract the same frame from the head video
for kp_set, color in [(left, (0, 255, 0)), (right, (0, 0, 255))]:
    if np.allclose(kp_set, 0): continue
    for x, y, _z in kp_set:
        cv2.circle(img, (int(x), int(y)), 4, color, -1)
cv2.imwrite("overlay.png", img)

Quick load (LeRobot)

ds = LeRobotDataset("zeno-labs/sample-gopro-v1")
sample = ds[0]
left  = sample["observation.hand_pose.left"]   # tensor [21, 3]
right = sample["observation.hand_pose.right"]

Stats

Metric	Value
Episodes	2
Frames	27,901
Total duration (~)	15.5 min
Videos	3

Schema (`meta/info.json`)

{
  "codebase_version": "v2.1",
  "robot_type": "egocentric_head_camera",
  "total_episodes": 2,
  "total_frames": 27901,
  "total_tasks": 1,
  "total_videos": 3,
  "total_chunks": 1,
  "chunks_size": 1000,
  "fps": 30,
  "splits": {
    "train": "0:2"
  },
  "data_path": "data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet",
  "video_path": "videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4",
  "features": {
    "observation.state": {
      "dtype": "float32",
      "shape": [
        7
      ],
      "names": [
        "x",
        "y",
        "z",
        "qx",
        "qy",
        "qz",
        "qw"
      ]
    },
    "timestamp": {
      "dtype": "float32",
      "shape": [
        1
      ],
      "names": null
    },
    "frame_index": {
      "dtype": "int64",
      "shape": [
        1
      ],
      "names": null
    },
    "episode_index": {
      "dtype": "int64",
      "shape": [
        1
      ],
      "names": null
    },
    "index": {
      "dtype": "int64",
      "shape": [
        1
      ],
      "names": null
    },
    "task_index": {
      "dtype": "int64",
      "shape": [
        1
      ],
      "names": null
    },
    "observation.images.head": {
      "dtype": "video",
      "shape": [
        720,
        1280,
        3
      ],
      "names": [
        "height",
        "width",
        "channel"
      ],
      "info": {
        "video.fps": 30.0,
        "video.height": 720,
        "video.width": 1280,
        "video.channels": 3,
        "video.codec": "h264",
        "video.pix_fmt": "yuv420p",
        "video.is_depth_map": false,
        "has_audio": false
      }
    },
    "observation.hand_pose.left": {
      "dtype": "float32",
      "shape": [
        21,
        3
      ],
      "names": [
        "keypoint",
        "xyz"
      ]
    },
    "observation.hand_pose.right": {
      "dtype": "float32",
      "shape": [
        21,
        3
      ],
      "names": [
        "keypoint",
        "xyz"
      ]
    },
    "observation.images.depth": {
      "dtype": "video",
      "shape": [
        540,
        1920,
        3
      ],
      "names": [
        "height",
        "width",
        "channel"
      ],
      "info": {
        "video.fps": 30.0,
        "video.height": 540,
        "video.width": 1920,
        "video.channels": 3,
        "video.codec": "h264",
        "video.pix_fmt": "yuv420p",
        "video.is_depth_map": true,
        "has_audio": false
      }
    }
  }
}

Interactive trajectory viewer

Each episode has an interactive 3D path viewer (head pose + original video + QA metrics). Click a link below — opens directly in your browser.

Episode	Task	View
000000	egocentric manipulation	view trajectory
000001	egocentric manipulation	view trajectory

⚠ Access codes are embedded in these links and may be rotated without prior notice. If a link returns an error, the code has been refreshed — please contact ZenO Labs for new codes.

Limitations

Hand pose detection rate varies per video (sideways / occluded hands have lower recall).
Body keypoints are not included — hands only.
Coordinate frame is ROS REP 103 (X=forward, Y=left, Z=up); remap if your stack uses a different convention.
Episodes may be re-published; pin to a specific commit SHA for strict reproducibility.

Reproducibility

This dataset may be updated with new commits. To pin a specific snapshot, reference an explicit HF revision (commit SHA):

ds = LeRobotDataset("zeno-labs/sample-gopro-v1", revision="<commit-sha>")

Contact

General support: support@zen-o.xyz
Business / commercial license: business@zen-o.xyz
Telegram: @zenoglasses

Citation

If you use this dataset in academic work, please cite:

@misc{zeno-sample-gopro-v1,
  title={ZenO Egocentric Dataset — Sample (gopro, v1, LeRobot v2.1)},
  author={ZenO Labs},
  year={2026},
  publisher={Hugging Face Datasets},
  howpublished={\url{https://huggingface.co/datasets/zeno-labs/sample-gopro-v1}}
}

Generated by ZenO Studio.