zeno-labs/egostation-catalog-v1

EgoStation Catalog v1

Curated egocentric manipulation dataset for Physical AI and robot imitation learning, distributed under a tiered access model.

• Built: 2026-05-12 20:22 UTC
• Episodes catalogued: 5 (5 sample, 0 private)
• Format: LeRobot v2.1 (Apache Parquet + MP4)
• Maintained by: ZenO Labs — https://zen-o.xyz

Dataset Summary

EgoStation is a growing catalog of first-person human demonstration videos annotated for robot imitation learning. Recordings are captured with consumer head-mounted cameras (GoPro Hero 9/11, ARKit-capable smartphones) and run through the ZenO Studio pipeline to produce:

• 6DoF head trajectory — ORB-SLAM3 mono-inertial with IMU gravity initialization and optional ArUco marker anchoring.
• 2D hand pose — MediaPipe Hands, 21 keypoints per hand, both hands, post-processed for L/R swap correction (Hungarian assignment with forearm skin-mask constraint) and short-gap interpolation.
• Per-frame relative depth — Apple Depth Pro (monocular).
• 3D hand keypoints in world frame — depth-fused, scaled via ArUco marker calibration when present.
• Action segments (subset) — manual labels for task boundaries.

Each sample-tier episode follows the LeRobot v2.1 spec for drop-in compatibility with the LeRobot training framework and the HuggingFace dataset viewer. private episodes are listed in catalog.parquet metadata only; raw data is available via Cloudflare R2 pre-signed URLs under a commercial agreement.

Access Model

Tier	Listed in catalog.parquet	Full data on HF	Access path
sample	Yes	Yes	Public download (this repo, CC-BY-NC 4.0)
private	Yes (metadata only)	No	R2 pre-signed URL after commercial agreement

The sample tier exposes 3–5 representative episodes per domain at full quality for evaluation. Production-grade access (entire catalog, R2 direct download) requires a signed agreement — see Contact below.

Repository Structure

zeno-labs/egostation-catalog-v1/
├── catalog.parquet              # All episodes metadata (sample + private)
├── README.md
└── gopro/                       # source = capture device family
    └── pick-and-place/          # domain = task family
        ├── data/chunk-000/episode_NNNNNN.parquet
        ├── videos/chunk-000/
        │   ├── observation.images.head/episode_NNNNNN.mp4
        │   └── observation.images.depth/episode_NNNNNN.mp4
        └── meta/
            ├── info.json
            ├── episodes.jsonl
            └── tasks.jsonl

• source ∈ {gopro, smartphone, smartphone-arkit}
• domain ∈ {pick-and-place, cooking, assembly, ...}

Each {source}/{domain} subfolder is a self-contained LeRobot v2.1 dataset and is registered as its own HuggingFace config. Config names use -- instead of / (HF identifier restriction): e.g. folder gopro/pick-and-place ↔ config gopro--pick-and-place.

Data Fields

catalog.parquet (one row per episode)

Field	Type	Description
episode_id	string	ep_NNNNNN, unique within this repo, assigned by publication order
session_id	string	ZenO Studio internal session UUID
source	string	Capture device family
title	string	Operator-supplied episode title
domain	string	Task family
duration_s	float	Episode duration in seconds
catalog_tier	string	sample or private
hf_path	string	Subfolder inside repo (empty for private)
has_head_trajectory	bool	6DoF camera pose available
has_hand_pose	bool	2D MediaPipe hand keypoints available
has_depth	bool	Per-frame depth maps available
has_hand_world	bool	3D hand keypoints in world frame
has_action_segment	bool	Action segmentation labels
completed_at	timestamp	Pipeline completion (UTC)
is_test	bool	Internal QA/test recording

Episode Parquet (LeRobot v2.1)

Per-frame columns under data/chunk-NNN/episode_NNNNNN.parquet:

Column	Shape	dtype	Description
observation.state	(7,)	float32	Head pose [x, y, z, qx, qy, qz, qw], world frame (RH, +X right, +Y up, +Z forward)
observation.hand_pose.left	(21, 3)	float32	Wearer's left hand. (u, v) in image pixels, z is MediaPipe relative depth
observation.hand_pose.right	(21, 3)	float32	Wearer's right hand
timestamp	scalar	float32	Seconds since episode start
frame_index	scalar	int64	0-indexed within episode
episode_index	scalar	int64	Episode index within subfolder dataset
index	scalar	int64	Global frame index across all episodes in subfolder
task_index	scalar	int64	Index into meta/tasks.jsonl

Videos

Under videos/chunk-NNN/:

• observation.images.head/episode_NNNNNN.mp4 — head camera, H.264, typically 1920×1080 @ 30 fps.
• observation.images.depth/episode_NNNNNN.mp4 — Depth Pro relative depth, colorized for visualization (not raw float).

meta/info.json highlights

{
  "codebase_version": "v2.1",
  "robot_type": "egocentric_head_camera",
  "fps": 30,
  "features": {
    "observation.state": {"dtype": "float32", "shape": [7]},
    "observation.hand_pose.left":  {"dtype": "float32", "shape": [21, 3]},
    "observation.hand_pose.right": {"dtype": "float32", "shape": [21, 3]},
    "observation.images.head":  {"dtype": "video", "shape": [1080, 1920, 3]},
    "observation.images.depth": {"dtype": "video", "shape": [1080, 1920, 3]}
  }
}

Auxiliary modalities (catalog flags vs LeRobot parquet)

Some flags in catalog.parquet indicate availability inside ZenO Studio but are not always materialized as columns in the LeRobot episode parquet for sample tier:

• has_hand_world — 3D hand keypoints in the world frame (depth-fused). Available as a separate hand_world.json file delivered via R2 for private tier; not currently a column in the LeRobot parquet.
• has_action_segment — manual task-boundary labels. Delivered as action_segments.json alongside the episode when present.

Capture Devices

Currently this catalog ships GoPro-source episodes only. Smartphone (source='smartphone') and ARKit smartphone (source='smartphone-arkit') support is partial / planned — when present in catalog.parquet, smartphone rows include video metadata only, with full SLAM-trajectory support landing in a later release.

Device	Resolution / fps	FOV mode	Calibration
GoPro Hero 9	1920×1080 @ 30	Wide (linearized)	Per-device intrinsics + IMU–camera extrinsics (T_bc) derived from UMI reference
GoPro Hero 11	1920×1080 @ 30	Wide / Linear	Same
Smartphone (planned)	1920×1080 @ 30	iPhone wide	Built-in IMU
Smartphone w/ ARKit (planned)	1920×1080 @ 60	iPhone wide	ARKit world pose + intrinsics

Coordinate Frames

• Image (u, v): pixel coordinates, origin top-left, +u right, +v down. Used for observation.hand_pose.* x and y components.

• MediaPipe z: normalized relative depth, ~0.0 at wrist, ±0.2 around finger extension. To convert to approximate metric, multiply by (video_width / focal_length_px).

• Robotics world frame (ROS REP 103): right-handed; +X forward, +Y left, +Z up. Used for observation.state head pose. Conversion from ORB-SLAM3's native (Y-down, +Z toward scene) frame is applied at publication time:

  x_ros =  z_slam
  y_ros = -x_slam
  z_ros = -y_slam
  q_ros =  q_conv ⊗ q_slam,  q_conv = (-0.5, 0.5, -0.5, 0.5)
  

Pipeline Quality

Each episode passes:

1. ORB-SLAM3 mono-inertial with IMU gravity init. Sessions that fail to initialize or lose tracking on >30% of frames are excluded.
2. MediaPipe Hands at min_detection_confidence=0.7, min_tracking_confidence=0.7, two-hand mode.
3. Pose post-processing: confidence filter (>0.6 hand-level), Hungarian L/R assignment with forearm skin-mask constraint, ≤3-frame linear interpolation, 1-Euro filter for jitter smoothing.
4. Depth Pro monocular relative depth.
5. Hand-world fusion: 2D keypoints back-projected through camera intrinsics, scaled with linear depth model, optionally anchored by ArUco markers when present in scene.

Aggregate quality metrics (mean detection rate, SLAM scale convergence flag, etc.) are recorded per episode in meta/info.json stats block.

Usage

Catalog only (metadata + discovery)

from datasets import load_dataset

catalog = load_dataset("zeno-labs/egostation-catalog-v1", "catalog", split="train")
print(catalog.to_pandas().head())

# Filter to a specific domain at sample quality
sample_pick = catalog.filter(
    lambda r: r["domain"] == "pick-and-place" and r["catalog_tier"] == "sample"
)

Full episode (sample tier)

# Config names use `--` instead of `/`:
ds = load_dataset("zeno-labs/egostation-catalog-v1", "gopro--pick-and-place", split="train")
frame0 = ds[0]
print(frame0["observation.state"])              # (7,) head pose
print(frame0["observation.hand_pose.left"].shape)  # (21, 3)

Via the LeRobot framework

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset(
    "zeno-labs/egostation-catalog-v1",
    subfolder="gopro/pick-and-place",
)

Private tier access

For full data access to catalog_tier='private' episodes, send your episode_id shortlist to support@zen-o.xyz. We issue Cloudflare R2 pre-signed URLs under signed agreement, typically valid 30 days with renewal on request.

Limitations & Known Issues

• Egocentric occlusion. Hands frequently leave the camera frame. Missing detections are interpolated only across short gaps (≤0.1 s); longer gaps are left as missing and marked quality='missing' in post-processed pose data.
• SLAM scale ambiguity. Mono-inertial SLAM scale converges with sufficient gravity excitation but is not always metric-accurate. ArUco marker recordings provide ground-truth scale anchoring; non-marker episodes carry a scale_uncertain flag in their meta/info.json.
• MediaPipe handedness convention. MediaPipe Hands labels handedness from a third-person observer perspective. We swap to the wearer's first-person frame at processing time, so observation.hand_pose.left in this dataset is the wearer's left hand throughout.
• No body pose ground truth. Only hands and head are tracked. Upper-body skeletons rendered in the ZenO Studio viewer are inferred from head pose with anthropometric heuristics and are not exported as data fields.
• HEVC ingest variability. Some GoPro recordings are HEVC; the pipeline operates on the original HEVC for IMU/PTS synchronization and produces an H.264 mp4 for web delivery.

Privacy & Ethics

Recordings are captured by ZenO operators or authorized contributors who provided informed consent. Incidental bystanders in sample-tier videos are blurred prior to publication. private-tier raw videos may contain unblurred footage and are only accessible under signed agreement.

Operator identifiers are pseudonymized; no personally identifying metadata is included in catalog.parquet.

Licensing

Component	License
catalog.parquet metadata	CC-BY 4.0
sample-tier videos, hand pose, trajectories	CC-BY-NC 4.0 (academic, non-commercial)
private-tier raw data	Commercial license required
ZenO Studio pipeline source	Proprietary (closed-source)

Commercial use of any component requires written agreement with ZenO Labs.

Citation

If you use EgoStation in research, please cite:

@dataset{egostation_2026,
  title    = {EgoStation: A Curated Egocentric Manipulation Catalog for Physical AI},
  author   = {ZenO Labs},
  year     = {2026},
  url      = {https://huggingface.co/datasets/zeno-labs/egostation-catalog-v1},
  version  = {v1}
}

Acknowledgments

The pipeline relies on:

• ORB-SLAM3 — mono-inertial head tracking
• MediaPipe Hands — 2D keypoints
• Apple Depth Pro — monocular relative depth
• LeRobot — dataset format
• Universal Manipulation Interface — GoPro calibration references

Contact

• Data access & partnerships: support@zen-o.xyz
• Technical issues: support@zen-o.xyz
• Website: https://zen-o.xyz
• Pipeline (ZenO Studio): https://studio.zen-o.xyz