Deep learning has made it possible to equip nano-drones with compact perception systems for autonomous navigation. However, these platforms have limited computing power, memory, energy, and time available for model updates. Their perception models are also generally trained offline and remain unchanged during operation. As a result, when a drone encounters conditions that differ from the original training data, its predictions may become unreliable, leading to unsafe navigation decisions and poor mission performance. This thesis considers gate classification for a swarm of nano-drones navigating through an indoor environment containing both target gates and obstacles. The multimodal classifier jointly processes a grayscale camera image and a depth map produced by a time-of-flight sensor.