UNISON-Sim

Dataset Summary

UNISON-Sim is a large-scale raw simulation dataset of dexterous hand-object interactions.

The dataset records complete grasp episodes with synchronized visual observations and physical hand states. Each episode covers multiple stages of an interaction, including hand approach, contact formation, object lifting and holding, and object return.

At each recorded timestep, UNISON-Sim provides:

• RGB observations of the hand-object interaction
• A 34-dimensional compact articulated hand pose
• Binary contact states over 162 hand-surface taxels
• Pressure values over the same 162 taxels
• Object properties and interaction context, when available

This repository provides the original generated episodes rather than a predefined model-training split. Users can construct their own temporal clips, samples, and dataset partitions according to their downstream tasks.

Repository Structure

The dataset files are organized as follows:

unison_sim/
├── tar_shards/
└── objects/

• unison_sim/tar_shards/ contains the compressed archive shards of the complete UNISON-Sim dataset. Users who need the full dataset should download and extract the archives in this directory.
• unison_sim/objects/ contains a small subset of example data for quickly inspecting the dataset structure and file contents. This directory does not contain the complete dataset.

Dataset Construction

UNISON-Sim is constructed from dexterous grasp trajectories provided by BODex.

The original trajectories contain object properties and staged hand-arm configurations for a UR10e robotic arm equipped with a Shadow Hand. These trajectories are replayed in the MuJoCo physics simulator and extended to form complete interaction episodes.

The construction process consists of the following main steps:

1. Trajectory replay Existing dexterous grasp trajectories are replayed in MuJoCo using the corresponding object and hand configurations.

2. Episode extension Successful grasp sequences are extended with additional holding and return motions. This produces complete episodes covering approach, contact formation, load bearing, and object return.

3. Physical validation Episodes are retained only when they satisfy physical-validity and data-completeness checks.

4. Tactile-state extraction Simulator contact forces are assigned to 162 virtual taxels calibrated on the Shadow Hand surface. Contact and pressure are represented on a consistent hand-surface topology.

5. Visual rendering RGB observations are rendered from multiple viewpoints. Camera placement, object appearance, textures, and illumination are varied while preserving the underlying physical interaction state.

6. Synchronized export RGB observations, hand pose, contact, pressure, and object context are exported on the same timeline.

Dataset Scale

Item	Value
Candidate grasp episodes	583,305
Retained complete episodes	45,124
Object instances	2,356
Object categories	15
Contact taxels	162
Pressure taxels	162

The object categories include common graspable objects such as bottles, jars, cans, mugs, and bowls.

Data Modalities

Depending on the available metadata, an episode may contain:

• Multi-view RGB observations
• Compact articulated hand states
• Taxel-level binary contact
• Taxel-level continuous pressure
• Object category and material
• Object mass, size, and friction
• Object geometry or shape information
• Textual interaction descriptions

All state modalities are temporally aligned within each episode.

Intended Uses

UNISON-Sim can support research on:

• Hand-object interaction understanding
• Dexterous grasp analysis
• Hand pose recovery
• Hand contact estimation
• Vision-based pressure estimation
• Unified hand-state modeling
• Temporal interaction modeling
• Tactile inference from visual observations
• Sim-to-real adaptation for hand interaction models

It can also serve as a source dataset for constructing customized training samples, temporal clips, camera-view subsets, and task-specific dataset partitions.

Limitations

UNISON-Sim is generated in a physics simulator and uses a robotic hand embodiment. Although the rendering process introduces variations in viewpoint, appearance, texture, and illumination, visual and embodiment gaps remain between the simulated data and real human-hand interactions.

Models trained using this dataset may therefore require domain adaptation or additional real-world data before being applied to natural videos.