Due to their capacity to capture global context and hierarchical features, transformer-based models have recently garnered significant attention over CNNs for improving perception in mobile robotics. In this work, we employ deep imitation learning in conjunction with attention mechanisms to enable autonomous robot navigation. Using RGB camera inputs with a limited field of view (66°) and no depth information, we trained and evaluated several model configurations with different numbers of attention blocks, heads, and latent feature sizes. Our experiments, conducted in a simulated environment, showed that the configuration (4, 4, 64) demonstrated the best performance, as evidenced by its higher success rate (80%) and lower RMSE (0.0496). Visualization of attention maps further highlighted the crucial role that attention mechanisms play in interpreting key scene elements and guiding decision-making. Despite its limited input data, the robot successfully navigated to its target while avoiding static obstacles.