Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Visual Embodied Brain: Let Multimodal Large Language Models See, Think, and Control in Spaces

Gen Luo, Ganlin Yang, Ziyang Gong, Guanzhou Chen, Haonan Duan, Erfei Cui, Ronglei Tong, Zhi Hou, Tianyi Zhang, Zhe Chen, Shenglong Ye, Lewei Lu, Jingbo Wang, Wenhai Wang, Jifeng Dai, Yu Qiao, Rongrong Ji, Xizhou Zhu

TL;DR

VeBrain introduces a unified Visual Embodied Brain that merges perception, reasoning, and real-world robot control by formulating control tasks as 2D text-based MLLM problems and coupling them with a robotic adapter for action execution. A semi-automated VeBrain-600k data engine provides multimodal, visual-spatial, and robotic-control data, augmented with multimodal chain-of-thought prompts to foster compositional reasoning. Through extensive benchmarks and real-robot tests, VeBrain consistently outperforms existing MLLMs and vision-language-action models, especially in legged-robot and robotic-arm tasks, while preserving multimodal understanding. This work demonstrates a practical pathway to deploy versatile, capable embodied agents that can see, think, and act in dynamic environments.

Abstract

The remarkable progress of Multimodal Large Language Models (MLLMs) has attracted increasing attention to extend them to physical entities like legged robot. This typically requires MLLMs to not only grasp multimodal understanding abilities, but also integrate visual-spatial reasoning and physical interaction capabilities. Nevertheless,existing methods struggle to unify these capabilities due to their fundamental differences.In this paper, we present the Visual Embodied Brain (VeBrain), a unified framework for perception, reasoning, and control in real world. VeBrain reformulates robotic control into common text-based MLLM tasks in the 2D visual space, thus unifying the objectives and mapping spaces of different tasks. Then, a novel robotic adapter is proposed to convert textual control signals from MLLMs to motion policies of real robots. From the data perspective, we further introduce VeBrain-600k, a high-quality instruction dataset encompassing various capabilities of VeBrain. In VeBrain-600k, we take hundreds of hours to collect, curate and annotate the data, and adopt multimodal chain-of-thought(CoT) to mix the different capabilities into a single conversation. Extensive experiments on 13 multimodal benchmarks and 5 spatial intelligence benchmarks demonstrate the superior performance of VeBrain to existing MLLMs like Qwen2.5-VL. When deployed to legged robots and robotic arms, VeBrain shows strong adaptability, flexibility, and compositional capabilities compared to existing methods. For example, compared to Qwen2.5-VL, VeBrain not only achieves substantial gains on MMVet by +5.6%, but also excels in legged robot tasks with +50% average gains.

Visual Embodied Brain: Let Multimodal Large Language Models See, Think, and Control in Spaces

TL;DR

VeBrain introduces a unified Visual Embodied Brain that merges perception, reasoning, and real-world robot control by formulating control tasks as 2D text-based MLLM problems and coupling them with a robotic adapter for action execution. A semi-automated VeBrain-600k data engine provides multimodal, visual-spatial, and robotic-control data, augmented with multimodal chain-of-thought prompts to foster compositional reasoning. Through extensive benchmarks and real-robot tests, VeBrain consistently outperforms existing MLLMs and vision-language-action models, especially in legged-robot and robotic-arm tasks, while preserving multimodal understanding. This work demonstrates a practical pathway to deploy versatile, capable embodied agents that can see, think, and act in dynamic environments.

Abstract

The remarkable progress of Multimodal Large Language Models (MLLMs) has attracted increasing attention to extend them to physical entities like legged robot. This typically requires MLLMs to not only grasp multimodal understanding abilities, but also integrate visual-spatial reasoning and physical interaction capabilities. Nevertheless,existing methods struggle to unify these capabilities due to their fundamental differences.In this paper, we present the Visual Embodied Brain (VeBrain), a unified framework for perception, reasoning, and control in real world. VeBrain reformulates robotic control into common text-based MLLM tasks in the 2D visual space, thus unifying the objectives and mapping spaces of different tasks. Then, a novel robotic adapter is proposed to convert textual control signals from MLLMs to motion policies of real robots. From the data perspective, we further introduce VeBrain-600k, a high-quality instruction dataset encompassing various capabilities of VeBrain. In VeBrain-600k, we take hundreds of hours to collect, curate and annotate the data, and adopt multimodal chain-of-thought(CoT) to mix the different capabilities into a single conversation. Extensive experiments on 13 multimodal benchmarks and 5 spatial intelligence benchmarks demonstrate the superior performance of VeBrain to existing MLLMs like Qwen2.5-VL. When deployed to legged robots and robotic arms, VeBrain shows strong adaptability, flexibility, and compositional capabilities compared to existing methods. For example, compared to Qwen2.5-VL, VeBrain not only achieves substantial gains on MMVet by +5.6%, but also excels in legged robot tasks with +50% average gains.

Paper Structure

This paper contains 31 sections, 1 equation, 3 figures, 11 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Multimodal Large Language Models
  4. Vision-Language-Action Models
  5. Method
  6. Task Formulations
  7. VeBrain Framework
  8. MLLM Architecture
  9. Robotic Adapter
  10. VeBrain-600k Data Engine
  11. Experiments
  12. Evaluation Datasets
  13. Implementation Details
  14. Quantitative Results
  15. Visualizations
  16. ...and 16 more sections

Figures (3)

  • Figure 1: Overview of VeBrain and VeBrain-600k. Compared to existing MLLMs, VeBrain achieves the best trade-off performance on benchmarks of multimodal understanding, visual-spatial reasoning, and robot control into one MLLM. To support the unified training of VeBrain, VeBrain-600k is built with a semi-automated data engine covering a variety of data sources and tasks.
  • Figure 2: Illustration of VeBrain architecture and robotic adapter. In VeBrain, the MLLM is capable of perception, thinking, and decision-making in common MLLM tasks. For robot control, an additional adapter is combined with the MLLM to achieve closed-loop control of the real robot.
  • Figure 3: Visualization of VeBrain on robotic arm and legged robot. VeBrain demonstrates compositional capabilities in handling complex robotic tasks. Due to space limitations, most of the text regarding the thinking process is omitted in the figure.