---
pretty_name: OWMM-Agent-Model
arxiv: 2506.04217
language:
- en
license: mit
tags:
- robotics
- multimodal
- open world mobile manipulation
- agent
base_model:
- OpenGVLab/InternVL2_5-8B
- OpenGVLab/InternVL2_5-38B
size_categories:
- 10B<n<100B
pipeline_tag: image-text-to-text
---
# OWMM-Agent: Open World Mobile Manipulation With Multi-modal Agentic Data Synthesis

[![arXiv](https://img.shields.io/badge/arXiv-2506.04217-B31B1B?logo=arxiv)](https://arxiv.org/abs/2506.04217)
[![Model](https://img.shields.io/badge/HuggingFace-Model-blue?logo=huggingface)](https://huggingface.co/hhyrhy/OWMM-Agent-Model)
[![Dataset](https://img.shields.io/badge/HuggingFace--Dataset-yellow?logo=huggingface)](https://huggingface.co/datasets/hhyrhy/OWMM-Agent-data)
[![GitHub](https://img.shields.io/badge/GitHub-OWMM--Agent-black?logo=github)](https://github.com/HHYHRHY/OWMM-Agent)

## 🚀 Introduction

The rapid progress of navigation, manipulation, and vision models has made mobile manipulators capable in many specialized tasks. 
However, the open-world mobile manipulation (OWMM) task remains a challenge due to the need for generalization to open-ended instructions and environments, as well as the systematic complexity to integrate high-level decision making with low-level robot control based on both global scene understanding and current agent state.
To address this complexity, we propose a novel multi-modal agent architecture that maintains multi-view scene frames and agent states for decision-making and controls the robot by function calling.
A second challenge is the hallucination from domain shift. To enhance the agent performance, we further introduce an agentic data synthesis pipeline for the OWMM task to adapt the VLM model to our task domain with instruction fine-tuning.
We highlight our fine-tuned OWMM-VLM as the first dedicated foundation model for mobile manipulators with global scene understanding, robot state tracking, and multi-modal action generation in a unified model. 
Through experiments, we demonstrate that our model achieves SOTA performance compared to other foundation models including GPT-4o and strong zero-shot generalization in real world.
<!-- insert banner video here -->
![OWMM-Agent Banner](docs/demo_banner.gif)

## 📖 Project Overview

The following repositories contain the implementation and reproduction of the method described in the paper “[OWMM-Agent: Open World Mobile Manipulation With Multi-modal Agentic Data Synthesis](https://arxiv.org/abs/2506.04217)”. 

- **Paper**: [arXiv:2506.04217](https://arxiv.org/abs/2506.04217)  
- **Model**: [`OWMM-Agent-Model`](https://huggingface.co/hhyrhy/OWMM-Agent-Model) — **current repo**, the Models we trained and used in OWMM tasks(both simulator and real world).
- **Dataset**: [`OWMM-Agent-data`](https://huggingface.co/datasets/hhyrhy/OWMM-Agent-data) — the training dataset of our OWMM Models.
- **GitHub**: [`OWMM-Agent-codebase`](https://github.com/HHYHRHY/OWMM-Agent) — the codebase of OWMM-Agent, including scripts for data collection and annotation in the simulator, as well as implementations for both step and episodic evaluations.