---
language:
- en
tags:
- multimodal
- active-perception
- embodied-ai
- robotics
- tactile
- audio
- force
- qwen2.5-omni
pretty_name: ROMA
---
ROMA: LLM System for Real-World Object-Centric
Multi-Sensory Active Perception
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 **checkpoint** and **dataset** of [ROMA](https://github.com/GeWu-Lab/ROMA), an LLM-based system for Real-World Object-Centric Multi-Sensory Active Perception. ROMA integrates vision, audio, touch, and force into a *reasoning-interaction-feedback* loop: the LLM identifies the missing evidence and selects the target object, the interaction (`lift`, `press`, `collide`, `shake`, `rotate`, `squeeze`), and the sensory modalities, while a physical interface executes the interaction and returns the multi-sensory feedback.
## What's Inside
| Component | Description | Status |
| :--- | :--- | :---: |
| **ROMA-7B checkpoint** | Multi-sensory LLM built on [Qwen2.5-Omni](https://github.com/QwenLM/Qwen2.5-Omni) with action / modality tokens, an audio branch, and an [AnyTouch 2](https://github.com/GeWu-Lab/AnyTouch2) tactile branch, trained with multi-sensory alignment followed by active-perception SFT. | Available |
| **ROMI-2K** | Real-world multi-sensory object interaction dataset covering nearly 2,000 objects and 6 atomic interactions with synchronized visual, audio, tactile, and force feedback. | Coming soon |
| **ROMA Bench** | 2,100 scene-level tasks (single-chain, multi-chain, and intent-driven) for evaluating active perception. | Coming soon |
| **Demo example scene** | One recorded tabletop scene (`example_data/1`) used by the local web demo. | Available |
## Checkpoint
The ROMA-7B checkpoint directory (`ROMA-Qwen2.5-Omni-7B`) contains:
```
ROMA-Qwen2.5-Omni-7B/
├── xxx.safetensors # base Qwen2.5-Omni-7B model
├── anytouch2.pth # tactile encoder
├── audio.bin # fine-tuned audio adapter and encoder
├── tactile.bin # fine-tuned tactile adapter
└── ROMA-LLM.bin # ROMA LLM weights
```
The weights are loaded in this order: base Qwen2.5-Omni, tactile encoder, new action / modality tokens, audio adapter, tactile adapter, ROMA LLM. The loading code is in `demo.py` of the [GitHub repository](https://github.com/GeWu-Lab/ROMA).
## Dataset
**ROMI-2K** and **ROMA Bench** are **coming soon**. They will contain:
- **Handheld object collection:** 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.
- **Tabletop scene collection (Training Set):** 400 scenes recorded with a robotic arm, with visual, audio, tactile, and force feedback, object bounding boxes, and annotations.
- **ROMA Bench:** 2,100 scene-level active-perception tasks built from the held-out test scenes.
## Usage
```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
```
## 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}
}
```