Localizing sources of airborne chemicals with mobile sensing systems finds applications in various crucial and perilous situations, such as safety and security investigation for detecting explosives or illegal drugs, search and rescue operations to locate survivors in the aftermath of natural hazards, or environmental monitoring in unsafe sites, following harmful leaks. Using autonomous robots in such situations would eliminate or, at least, reduce human intervention and keep them from harm. Additionally, such operations would be more cost-effective and more time-efficient. That is why, in the past 30 years, gas source localization has been an attractive research topic in robotics and related areas, where different methods have been designed and evaluated for this purpose. However, the inherent complexity of gas dispersal phenomena, which is non-trivial to analyze and predict, especially in complex environments, is the main source of challenges in this field. Therefore, researchers tend to design and evaluate algorithms in simplistic environments before tackling more complex ones. In this thesis, we have designed and investigated a gas source localization algorithm based on source term estimation with a probabilistic approach. After validating the performance of the method in a baseline environment using a wheeled-robot, we gradually enhanced our method by enriching it with new features in order to be adaptable to more complex scenarios. In particular, the algorithm was shown to be successful in a simplified three-dimensional setup as well as in an unknown environment where no global map and positioning system is available. Furthermore, it was deployed on a homogeneous multi-robot system, where different coordination strategies between robots were designed and studied. Finally, designing a data-driven plume model and integrating it to the main framework of the method allowed for adaptation to cluttered environments. The method is systematically evaluated through high-fidelity simulations and in a wind tunnel emulating realistic and repeatable conditions. Lastly, the performance of our algorithm was compared with other state-of-the art methods to show its potentials and limits.