A Cycle-Temporal Attention Network (CTAN) for audio-visual navigation in robots is proposed, using a bidirectional cycle-consistency constraint and a Temporal Cross-Modal Memory to enhance multimodal interaction and performance in complex environments.