Animals often make use of information from multiple sensory systems, such as vision, proprioception, olfaction, and audition to guide their behavior. They must thus integrate these different sensory inputs or select one of them in order to effectively make decisions in the environment. How exactly the brain accomplishes this sensory integration and to what extent each signal contributes under what circumstances remains to be fully understood. We use a deep reinforcement learning model to study sensory integration in a navigation task in two different scenarios. In the first scenario, there is environmental uncertainty that causes each signal to be lost intermittently. This requires the agent to learn to use each signal individually as well as combine them when needed. In the second scenario, neither signal is lost. The agent can thus freely choose which signal(s) to use, whether to integrate them and to what degree. We simulate these two scenarios using visual and goal-vector signals in our model, which has been studied extensively experimentally in the context of how visual and self-motion cues interact in navigation. In both scenarios, we added noise to each signal individually in order to manipulate its reliability. We found that when the agent is required to use both signals, i.e., in the first scenario, it learns to integrate the two. As expected, navigation with the noisy signal is less accurate, and integrating the two improves accuracy. However, in the second scenario, where integrating the signals is not strictly necessary, we find that the interaction between the two signals is more complex. At lower levels of noise, the agent still integrates the signals to different extents, but at higher levels of noise, the more reliable signal is chosen over the other and the noisy signal is ignored altogether.  In summary, we find that integrating both signals comes at the cost of accuracy when one signal is noisy, but with the benefit of increased robustness. On the other hand, choosing one signal over the other leads to faster learning and better accuracy, but sacrifices the ability to adapt to changing environments. Based on this finding, our model is able to replicate experimental results in the Morris Water Maze. Finally, we examine the representations that emerge in the network and find that these representations rely on both signals to drive their activity, albeit to different extents, and tie these findings back to the behavioral results.