Autonomous Vehicle Guidance in Industrial Settings: Dataset Acquisitions via Unity-based Simulation Environments

This thesis proposes a comprehensive framework for acquiring large-scale, semantically annotated datasets in industrial contexts using a Unity-based simulation environment closely integrated with the Robot Operating System (ROS). The work addresses the growing demands of autonomous driving research, where scalable data generation and robust domain coverage are key to developing high-performing perception modules. By precisely reconstructing an industrial yard at 1:1 scale and combining random object spawning, the simulator can emulate realistic, variable conditions while maintaining full control over scene geometry and class labels. In parallel, a specialized ROS node autonomously selects and publishes navigation goals, enabling the simulated vehicle to traverse diverse routes without manual intervention, thus accelerating data collection. A custom editor script in Unity streamlines the assignment of semantic classes to each virtual asset, guaranteeing the immediate generation of labeled LiDAR point clouds following the SemanticKITTI format. Additionally, performance bottlenecks in Unity–ROS communication, common when transferring dense sensor data, are mitigated by using a TCP-based connector instead of WebSocket-based solutions, significantly improving throughput. Experimental evaluations demonstrate that combining real and synthetic datasets is a viable solution in order to train a segmentation network. While synthetic data alone achieves reliable results, the inclusion of real-world samples further refines performance, suggesting that large-scale simulation can effectively supplement or partially replace field-based acquisitions. The methodology ultimately points toward safer and cost-effective development cycles for autonomous vehicle guidance systems in both controlled and open environments.