Säker människa-robot-interaktion i fotorealistisk simuleringsmiljö

Abstract

The development of automated industrial environments increases the need for a safe and efficient transition toward fully automated factories where humans and robots can collaborate. One effective approach for enabling this is the creation of a digital twin that can replicate reality within a simulated environment for simulation, testing, and validation before implementation in real production environments. This bachelor’s thesis investigates how 3D Gaussian Splatting (3DGS) can be used to create a realistic digital twin of an industrial robot cell. The model is based on the RITA-cell, a kitting environment containing a collaborative robot arm (cobot), and is implemented in the Unity game engine for simulation and visualization. The project includes the development of a 3DGS-based simulation model integrated into Unity. In addition, an algorithm for generating synthetic training data and a YOLO-based segmentation model were implemented for identifying components within the RITA-cell. Motion planning for the cobot was developed through com munication between Unity and MoveIt using ROS 2, and a simplified sensor fusion solution was implemented to coordinate camera data in a shared coordinate system. These subsystems were primarily developed and tested as separate parts of the project. Furthermore, VR was chosen as the interaction method for visualization and manipulation of the simulated environment. The results show that 3DGS can be used in an efficient and practical way to create visually realistic digital environments, although with certain limitations related to mesh geometry, which resulted in challenges regarding collision handling and physical simulation. The developed environment showed potential as a platform for robotic application development, but the final demonstration should mainly be viewed as a proof-of-concept since motion planning and sensor fusion were not fully integrated into the final simulation environment. The conclusion is that 3DGS is well suited for creating visually realistic digital twins of industrial environments, but that the lack of straightforward mesh geometry integration makes it less suitable as a complete basis for physical simulation. The developed environment demonstrates potential as a platform for the development of robotic applications, provided that the visual model is combined with separate collision geometry and more robust integration of perception, sensor fusion, and robot control.