Samba Swaps GRUs for Mamba to Break a Five-Year Stall in Audio-Visual Navigation

A navigation agent that listens as much as it looks just posted an 11.3% jump in success rate on the Matterport3D benchmark, and the architecture behind it borrows its core idea from state-space language models rather than the convolutional networks that have dominated this niche for years.

The system, called Samba, has been accepted for publication at the IEEE International Conference on Systems, Man, and Cybernetics 2026 (IEEE SMC 2026). Its authors point to a specific stagnation problem: audio-visual navigation backbones, the neural networks that let a simulated agent move toward a sound source while interpreting camera input, have seen no meaningful architectural update in more than five years. Most existing systems still lean on convolutional operators paired with GRUs (gated recurrent units) for tracking information over time, a combination that has become the de facto standard despite its known ceiling on modeling long-range dependencies.

Where Mamba fits in

Samba's answer is to bring the Mamba family of state-space models into this pipeline. The paper introduces two components built around that idea. The first is an Audio Mamba Encoder (AME), which processes spectrograms, the time-frequency representations of sound, in place of standard convolutional layers. Convolutions are effective at picking up local patterns but struggle to connect distant points in a spectrogram, which matters when an agent needs to reason about how a sound evolves or reflects across a large, complex space. The AME is designed specifically to capture those global time-frequency relationships that convolutional operators tend to miss.

The second component is the Mamba State Encoder (M-SE), which takes over the role GRUs have traditionally played: aggregating information across time as the agent moves through an environment. Rather than processing every input in a rigid recurrent step, the M-SE uses an adaptive selection mechanism, a hallmark of Mamba-style state-space models, to decide which information is most relevant to retain as the agent's trajectory unfolds. This selective approach is the same principle that has made Mamba an attractive alternative to transformers and RNNs in sequence modeling more broadly, now redirected toward spatial audio-visual reasoning.

What the results show

On Matterport3D, a widely used dataset of photorealistic indoor 3D environments for embodied AI research, Samba's combination of AME and M-SE produced an 11.3% improvement in success rate over prior state-of-the-art models. Success rate in this context measures how often the agent actually reaches the target sound source within a navigation episode, making it one of the most direct indicators of practical performance rather than a proxy metric.

The result is notable less for the raw percentage gain and more for what it implies about the ceiling that convolution-plus-GRU designs had quietly imposed on the field. If a swap in backbone architecture, without new training data or task reformulation, can unlock a double-digit improvement, it suggests the audio-visual navigation subfield has room the community had not fully explored simply because the standard toolkit went unchallenged for so long. Whether Mamba-based backbones become the new default for embodied audio-visual agents, or whether hybrid designs like Samba represent a transitional step toward something else, will depend on how the approach generalizes beyond Matterport3D to other datasets and real-world deployment settings.