Fotorealistisk simulering av mobila robotar i fabriksmiljöer

Abstract

Modern factories handle more components than ever before, creating major logistical challenges. As a solution, autonomous mobile robots (AMRs) are now being used. Training the behavior of these robots in real-world environments is expensive and time-consuming. For this reason, the possibility of using simulated environments built with 3D Gaussian Splatting (3DGS) is being investigated as an alternative to real-world training. This bachelor’s thesis evaluates whether the simulation environment can be used to collect synthetic data for a segmentation model for object localization, which in turn can enable the control of AMRs. The method involved filming industrial environments which could then be used to create a digital twin using a 3DGS model. From the 3D model, a simulation environment was created through synthetic dynamic and static objects that replicate the real industrial environment. Synthetic data was then collected from the simulation environment which could be used to train a segmentation model. The fine-tuned model could be used for object localization and identification, which was utilized for the control of AMRs. The results showed that it is possible to produce photorealistic simulation environments of industrial environments, from which it is possible to produce large quantities of high-quality annotated data automatically. The instance segmentation model YOLOv8-Seg Nano was able to localize objects in the simulation environment in real time. Consequently, AMRs could be controlled using the segmented instances. The conclusion is that although perfect photorealism was not achived in this project, 3DGS proves to be a viable tool in towards complete realism. The resulting environment is however good enough to create synthetic training data that results in a segmentation model capable of predicting objects in real world photos.