bigai/TongSIM-Asset

TongSIM

What is TongSIM-Asset?

Vist the TongSIM-Asset Homepage

TongSIM: A General Platform for Simulating Intelligent Machines

As artificial intelligence advances, particularly through Large Language Models (LLMs), the shift is moving toward more complex and dynamic domains such as multimodal and embodied AI. Embodied intelligence, which focuses on agents learning and acting within realistic simulated environments, presents a significant challenge compared to traditional AI models that rely on labeled data. However, current simulation platforms tend to specialize in specific tasks, with a lack of a general-purpose, high-consistency environment that supports both low-level single-agent tasks (like navigation) and high-level tasks (including multi-agent social simulations and human-AI collaboration).

TongSIM-Asset addresses this gap by offering an open-source, high-fidelity, and versatile platform for training and testing embodied agents in diverse environments. The TongSIM-Asset package features over 100 varied, multi-room indoor scenarios, alongside a rich, interactive outdoor simulated town. This platform is equipped with thousands of 3D interactive objects spanning more than 500 categories, allowing researchers to experiment in both familiar indoor spaces and complex outdoor settings.

In addition to these environments, TongSIM-Asset includes a comprehensive evaluation system with benchmarks that assess a wide range of agent capabilities, including perception, cognition, decision-making, learning, execution, and social collaboration. Features like customizable scene design, rich annotations, and parallel training support further enhance the platform’s utility.

By offering this open-source asset, TongSIM-Asset provides a unified and powerful tool for the development and testing of general embodied agents, enabling researchers from diverse AI fields to advance embodied intelligence research and accelerate development processes.

The TongSIM dataset is a asset project based on physical world indoor or outdoor environments. This dataset is intended for use in the TongSIM simulator for embodied in-home interaction tasks such as object re-arrangement.

This version positions the platform as an open-source asset, highlighting its versatility, ease of use, and comprehensive evaluation features for researchers across multiple domains.

All materials are licensed under the MIT License.

Dataset Contents (this repository):

TongSIM-Asset provides a project built on Unreal Engine version 5.6. The dataset contains 100 scenes are used as a test set for challenge assessment. All scenes can run above 60FPS.

• 3000+ objects covering 500+ categories.
• 100 scenes.
• 10+ agents, including adults, children, and robots.
• 100+ agent animations

Dataset Structure

├── Config
│   ├── */*.ini
├── Content
│   ├── Animations
|      ├── *.uasset
│   ├── Characters
|      ├── *.uasset
│   ├── Maps
|      ├── *.uasset
│   ├── Objects
|      ├── *.uasset
├── Docs
│   ├── *.md
│   ├── *.png
└── TongSimLite.uproject

Install Unreal Engine 5.6:

TongSIM-Asset project built on Unreal Engine version 5.6 and requires to be installed first.

Vist the unrealengine Homepage

Vist the unrealengine Documentation

---

Citing TongSIM-Asset

Using TongSIM-Asset in your research? Please cite the following paper: arxiv

@inproceedings{sun2025tongsim,
    title       =     {TongSIM: A General Platform for Simulating Intelligent Machines},
    author      =     {Sun, Zhe and Wu, Kunlun and ... Zhang, Zhenliang},
    journal     =     {Technical Report},
    year        =     {2025},
    institution =     {State Key Laboratory of General Artificial Intelligence, BIGAI}
}