language:
- en
license: cc-by-nc-4.0
size_categories:
- 1K<n<10K
task_categories:
- video-classification
- object-detection
tags:
- egocentric
- dexterous-manipulation
- hand-pose
- robotics
- embodied-ai
- first-person
- human-demonstration
pretty_name: OBayData Egocentric Dexterous Manipulation Demo
dataset_info:
config_name: default
splits:
- name: train
num_examples: 500
- name: test
num_examples: 100
OBayData Egocentric Dexterous Manipulation Demo
Dataset Description
- Homepage: obaydata.com
- Repository: obaydata/egocentric-data-collection
- Point of Contact: contact@obaydata.com
Dataset Summary
This is a demo dataset from OBayData, showcasing our egocentric dexterous manipulation data collection capabilities. It contains a curated subset of first-person human operation recordings captured with our multi-camera hardware setup across real-world scenarios.
This demo is intended for quality evaluation. Full-scale data collection is available as a managed service — see Production Data Service below.
Supported Tasks
- Action Recognition — classify human manipulation actions from egocentric video
- Hand Pose Estimation — predict 3D hand joint positions from first-person views
- Robot Learning from Human Demonstrations — use as training signal for imitation learning and sim-to-real transfer
- Video Understanding — temporal reasoning over long-horizon manipulation sequences
- Language-Grounded Manipulation — map atomic language annotations to visual actions
Languages
All annotations are in English.
Dataset Structure
Data Fields
| Field | Type | Description |
|---|---|---|
video |
Video (MP4) | 1080p egocentric video from head-mounted camera |
wrist_left |
Video (MP4) | 1080p video from left wrist camera |
wrist_right |
Video (MP4) | 1080p video from right wrist camera |
hand_pose |
JSON | Per-frame 3D hand joint positions (21 joints × 2 hands) |
camera_extrinsics |
JSON | Per-frame camera pose (rotation + translation) |
action_labels |
JSON | Temporal action segments with start/end timestamps and labels |
language_annotations |
JSON | Atomic-level natural language descriptions of sub-actions |
metadata |
JSON | Scenario type, session ID, operator ID, hardware config |
Data Splits
| Split | Examples | Purpose |
|---|---|---|
train |
500 | Model training and development |
test |
100 | Held-out evaluation |
Data Instances
Each instance represents a single continuous manipulation session (typically 30 seconds to 5 minutes) with synchronized multi-view video and annotations.
Dataset Creation
Collection Process
Data was collected using OBayData's standardized hardware rig:
- Head Camera — head-mounted, capturing the egocentric viewpoint at 1080p / 30fps
- Dual Wrist Cameras — left and right wrist-mounted cameras for close-up hand views
- Optional Tactile Gloves — Manus METAGLOVES PRO for finger-level force and contact data (not included in this demo)
- Synchronization — all streams are hardware-synchronized and temporally aligned
Operators performed real manipulation tasks in authentic environments (not staged lab settings). Scenarios in this demo include kitchen tasks, object rearrangement, and tool use.
Annotation Pipeline
- Camera Extrinsics — estimated via structure-from-motion on the multi-view streams
- Hand Pose Reconstruction — 3D hand mesh fitting using multi-view optimization
- Action Segmentation — temporal boundaries annotated by trained human annotators
- Language Annotations — atomic-level descriptions written by annotators following a structured protocol (e.g., "grasp the red mug handle with right hand", "pour water into the bowl")
All annotations undergo multi-stage quality review.
Source Data
Recorded in real-world environments including residential kitchens, hotel rooms, and office spaces. No synthetic or simulated data.
Personal and Sensitive Information
All operators provided informed consent. Faces are not visible in egocentric recordings. No personally identifiable information is included in the released annotations.
Considerations for Using the Data
Intended Uses
- Evaluating OBayData's data quality before commissioning large-scale collection
- Research in egocentric vision, hand-object interaction, and robot learning
- Benchmarking action recognition and hand pose estimation models
- Prototyping language-grounded manipulation systems
Out-of-Scope Uses
- This demo is not large enough for training production foundation models — contact us for scaled collection
- Not suitable for surveillance or biometric applications
- Not intended for commercial redistribution
Limitations
- Demo contains a limited subset of scenarios (full service covers 1,000+ scenario types)
- Tactile glove data is not included in this release
- Annotation density may vary across sessions
Production Data Service
This demo represents a small fraction of OBayData's collection capabilities.
Full Service Specifications
| Capability | Details |
|---|---|
| Monthly Capacity | 100,000 hours/month |
| Availability | Deliveries starting May 2026 |
| Scenario Coverage | 1,000+ real-world scenarios (hospitality, manufacturing, retail, logistics, food service, etc.) |
| Pricing | $5.50–$100/hour depending on volume, annotation complexity, and hardware config |
| Custom Scenarios | Define your own tasks, environments, and skill requirements |
| Benchmark Alignment | RoboTwin 2.0, PI Olympics, and custom benchmark protocols |
Contact Us
- Website: obaydata.com
- Email: simon.su@obaydata.com
- Company: New Oriental Bay Limited (香港新东湾有限公司)
Citation
@dataset{obaydata2026egocentric,
title={OBayData Egocentric Dexterous Manipulation Demo},
author={OBayData Team},
year={2026},
url={https://huggingface.co/datasets/obaydata/egocentric-dexterous-manipulation-demo},
publisher={Hugging Face}
}
License
This demo dataset is released under CC BY-NC 4.0. Commercial data collection available under custom licensing — contact us for details.