---
language:
- en
tags:
- multimodal
- active-perception
- embodied-ai
- robotics
- tactile
- audio
- force
pretty_name: ROMI-2K
---
ROMI-2K: Real-World Object-Centric
Multi-Sensory Interaction Dataset
I saw, I touched, I understood.
Ruoxuan Feng*, Yutong Chen*, Ruihua Song, Huan Yang, Zhongyuan Wang, Guocai Yao, Di Hu✉
*Equal contribution ✉Corresponding author
---
This repository hosts the **dataset** of [ROMA](https://github.com/GeWu-Lab/ROMA), an LLM-based system for Real-World Object-Centric Multi-Sensory Active Perception. The model checkpoint is released separately as [GeWu-Lab/ROMA-7B](https://huggingface.co/GeWu-Lab/ROMA-7B).
## Overview
**ROMI-2K** is a real-world multi-sensory object interaction dataset covering nearly 2,000 objects and 6 atomic interactions (`lift`, `press`, `collide`, `shake`, `rotate`, `squeeze`) with synchronized visual, audio, tactile, and force feedback. It also includes **ROMA Bench**, 2,100 scene-level tasks for evaluating active perception.
> **Status:** The ROMA-Bench QA tasks and labels are now available in [`roma_bench_tasks.json`](roma_bench_tasks.json). The remaining data (including the images, tactile, and other sensory files referenced by the tasks) is being prepared and will be uploaded to the directories below.
## Dataset Structure
```
ROMA/
├── Handheld/ # Handheld object collection (coming soon)
├── Tabletop-train/ # Tabletop scene collection, training set (coming soon)
└── ROMA-Bench/ # Scene-level evaluation tasks (✅ updated 2026-10-20)
roma_bench_tasks.json # ROMA-Bench QA tasks and labels (✅ updated 2026-10-20)
```
| Directory | Description | Status |
| :--- | :--- | :---: |
| `Handheld` | 1,657 object-content combinations, each interacted with at 3 grasp locations using 6 atomic interactions, with wrist and third-person views, audio, and tactile feedback. | Coming soon |
| `Tabletop-train` | 400 tabletop scenes recorded with a robotic arm, with visual, audio, tactile, and force feedback, object bounding boxes, and annotations. | Coming soon |
| `ROMA-Bench` | 2,100 scene-level active-perception tasks (single-chain, multi-chain, and intent-driven) built from the 100 tabletop scenes with unseen objects. QA tasks and labels are in `roma_bench_tasks.json`. | ✅ Updated (2026-10-20) |
## Usage
Download the dataset into the `resources/datasets` directory of the [GitHub repository](https://github.com/GeWu-Lab/ROMA):
```bash
git clone https://github.com/GeWu-Lab/ROMA.git
cd ROMA
# see the GitHub README for environment setup
hf download GeWu-Lab/ROMA --repo-type dataset --local-dir resources/datasets
```
The ROMA-7B checkpoint and the demo do not require this dataset. See the [GitHub repository](https://github.com/GeWu-Lab/ROMA) for the demo, and [GeWu-Lab/ROMA-7B](https://huggingface.co/GeWu-Lab/ROMA-7B) for the checkpoint.
## Citation
```bibtex
@article{feng2026roma,
title = {ROMA: LLM System for Real-World Object-Centric Multi-Sensory Active Perception},
author = {Feng, Ruoxuan and Chen, Yutong and Song, Ruihua and Yang, Huan and
Wang, Zhongyuan and Yao, Guocai and Hu, Di},
journal = {arXiv preprint arXiv:2610.06955},
url = {https://arxiv.org/abs/2610.06955},
year = {2026}
}
```