Ultra-wideband technology enables highly precise indoor localization. However, its accuracy degrades significantly in non-line-of-sight conditions, where obstacles block the direct signal path. Such obstructions result in an overestimation of the transmitter-receiver distance due to signal reflections or a reduction in propagation speed as the signal passes through materials. Machine learning models have been successfully applied to mitigate such ranging errors. However, their ability to generalize to previously unseen environments, a crucial component for real-world deployment, remains a challenge. While model fine-tuning has been demonstrated to enhance predictive performance, it relies on labeled data collected within the target domain. This time-consuming process becomes increasingly impractical when deploying a large number of indoor localization systems across diverse environments. To address this limitation, this thesis evaluates the potential of self-supervised learning and domain-adversarial training to improve generalization and enable adaptation to new domains. The findings indicate that generalization across transmission channels can be accomplished with a purely supervised baseline, while cross-environment generalization poses a significant challenge, particularly across different datasets. Among the investigated approaches, domain-adversarial training improved robustness, while self-supervised learning produced mixed results. The integration of unlabeled target domain data into the training process was found to be largely ineffective.