Semantic segmentation is a computer vision task that aims to assign a class label to every pixel in an input image. This thesis is divided into three main parts, all connected by the goal of solving semantic segmentation for driving scenes. In the first part, we address the task of semantic segmentation for driving scenes under adverse weather conditions. We collected images from diverse datasets and applied pseudo-labeling to unlabeled images. The model trained on this dataset achieved the second-best results in the ACDC2023 challenge. In the second part of the thesis, the investigation revolves around the application of language embeddings for semantic segmentation. The primary goal is to explore whether incorporating language embeddings facilitates domain generalization and contributes to the development of more robust models. Synthetic images extracted from the GTA5 video game were utilized to train the models, while the evaluation was conducted on real-world driving scene images. The findings reveal that the integration of language embeddings has the potential to enhance domain generalization capabilities. Lastly, the thesis delves into solving semantic segmentation for driving scenes through unsupervised domain adaptation. We reproduced state-of-the-art results and conducted experiments using various convolutional architectures as backbones. However, the application of convolutional backbones did not yield improved results compared to using the Transformer architecture, despite the convolutional backbones being pretrained on much larger datasets.