Transformer-Based Object Detection in Unconstructed Environments: A ROS-Gazebo Simulation Framework

Abstract

This study introduces a transformer-based object detection framework at unstructured and cluttered environments using Robot Operating System 2 (ROS 2) and Gazebo simulation. We demonstrate robust object recognition capabilities Employing Detection with Transformers (DETR) architecture on a simulated DARPA-inspired environment using the Unitree Go2 quadruped robot by employing Detection with Transformers (DETR) architecture. The framework integrates a pretrained vision transformer model and clearly exhibits its object detection performance at unconstructed environments, comparing results with a YOLO-based baseline. Our experiments highlight the advantages of transformer models in spatial understanding and detection accuracy. The complete system is developed and tested within a ROS 2-Gazebo setup for transferring the developed model from simulation to real-world autonomous robotics applications. The results indicate that transformer-based models offer significant improvements in terms of generalization over convolution-based counterparts in unstructured, cluttered, and uncertain scenarios. © 2025 IEEE.