UNISON-S2R

Dataset Summary

UNISON-S2R is a paired real-and-simulated RGB video dataset for sim-to-real visual adaptation in hand-object interaction.

Each record contains a real human-hand video clip and a corresponding simulated dexterous-hand video clip representing an aligned manipulation action. The paired clips share the same frame count, frame rate, and camera-view index.

The dataset is designed to reduce the visual domain gap between simulated and real hand-object interaction videos without requiring dense pose, contact, or pressure annotations for the real videos.

Dataset Statistics

Item	Value
Paired sequence-view records	4,820
Aligned sequence IDs	1,205
Source demonstrations	488
Camera views	4
Manipulation primitives	61
Total RGB frames	Approximately 3.74 million
Video frame rate	60 FPS
Video resolution	848 × 480

Data Contents

Each paired record includes:

• A real human-hand RGB video clip
• A corresponding simulated dexterous-hand RGB video clip
• Source sequence and camera-view information
• Frame-range metadata
• Hand-side information
• An action primitive label

The action primitives cover common manipulation behaviors such as holding, placing, pouring, scooping, and capping.

Intended Uses

UNISON-S2R can be used for research on:

• Sim-to-real visual domain adaptation
• Real-to-simulation video alignment
• Hand-object interaction representation learning
• Dexterous-hand and human-hand correspondence
• Robot-hand retargeting
• Action-conditioned manipulation simulation