Samba: Hybrid Mamba model boosts audio-visual navigation by 11.3%

For over five years, audio-visual navigation models have relied on CNN and recurrent architectures (e.g., GRUs) that fail to efficiently handle dynamic multimodal sequences. The paper Samba (A Hybrid Mamba for Audio-Visual Navigation) proposes a complete overhaul: it replaces conventional GRUs with a Mamba State Encoder (M-SE) that leverages adaptive selection for better temporal aggregation. Additionally, an Audio Mamba Encoder (AME) replaces convolutional operators to capture global time-frequency dependencies in spectrograms, overcoming limitations of local receptive fields.

Experimental results demonstrate Samba's exceptional generalization to unheard sound sources and unseen scenes. On the Matterport3D dataset, Samba boosts navigation success rate (SR) by 11.3% compared to existing state-of-the-art models. Performance gains are even larger on the Replica dataset, which features finer scene structures. The modernized architecture achieves stronger embodied representation at lower computational cost, providing a robust pathway for advancing audio-visual navigation research. The paper has been accepted for publication at IEEE SMC 2026.