README.md

metadata

language:
  - en
license: mit
pretty_name: ESARBench
task_categories:
  - robotics
tags:
  - uav
  - embodied-ai
  - search-and-rescue
  - mllm

ESARBench: A Benchmark for Agentic UAV Embodied Search and Rescue

Project Page | Paper | GitHub

ESARBench is a comprehensive benchmark designed to evaluate Multimodal Large Language Model (MLLM) driven UAV agents in highly realistic Embodied Search and Rescue (ESAR) scenarios. It requires aerial agents to autonomously explore complex environments, identify rescue clues, and reason about victim locations to execute informed decision-making.

Key Features

• High Fidelity: 4 large-scale environments built with UE5 + AirSim using real-world GIS data.
• Realistic Dynamics: Integrated simulations of weather, lighting, and diverse rescue clues.
• Task Diversity: 600 tasks modeled after real-world rescue cases.
• Evaluation: A robust set of metrics to evaluate spatial memory, aerial adaptation, and search efficiency.

Quick Start

Installation

1. Create a Python 3.9 environment using Conda:

   conda create -n esar_env python=3.9
   conda activate esar_env
   

2. Install the required Python packages:

   pip install numpy msgpack-rpc-python pandas scipy dashscope opencv-python scikit-image scikit-fmm ultralytics matplotlib transformers pillow
   

3. Clone and install the AirSim-Colosseum client:

   git clone https://github.com/CodexLabsLLC/Colosseum.git
   cd Colosseum/PythonClient
   pip install -e . --no-build-isolation
   

Running the Benchmark

The platform scripts load sample tasks by default. To run the benchmark, navigate to the source directory and run the platform script corresponding to your OS:

• Linux

cd source
python main_platform_linux.py

• Windows

cd source
python main_platform_win.py

Analyzing Results

Each task writes a log file to log/task_<task_id>_log.json. You can analyze the results using:

cd source
python result_analysis.py

Citation

If you use ESARBench in your research, please cite:

@article{esarbench2024,
  title={ESARBench: A Benchmark for Agentic UAV Embodied Search and Rescue},
  author={Authors list here},
  journal={arXiv preprint arXiv:2605.01371},
  year={2024}
}

License

This project is licensed under the MIT License.