The development of autonomous driving (AD) algorithms and their testing would be very expensive when real-world vehicles and environments would be used from the very beginning of development. That is why in the initial development phases simulators are often used, in which it is possible to achieve realistic simulations of the real world. Therefore, many different simulators for AD have been developed. One of the main tasks of a fully autonomous vehicle is steering angle prediction (SAP). This paper explores SAP using a PilotNet CNN-based algorithm that recognizes essential features from input images to predict vehicle trajectory. After training on real-world images, the algorithm’s performance is evaluated in both CARLA and AirSim simulators, achieving 91.65% autonomy in CARLA and 71.87% in AirSim. The algorithm is further trained on the images obtained from the simulator, resulting in improved performance of 96.34% in CARLA and 78.99% in AirSim in terms of total autonomy. The study demonstrates the benefits of simulators for the development of AD algorithms and highlights the importance of testing in multiple simulators to achieve reliable performance across different environments and scenarios, including different lighting and weather conditions, types of road markings, and track configurations. The paper also highlights the advantages and disadvantages of individual simulators, which can be useful for the development and testing of future SAP algorithms.