SpikingNav: Spiking Neural Policies Boost Robust Embodied Navigation

Embodied navigation demands that an agent make sequential decisions based on egocentric observations within a physical environment. Conventional Artificial Neural Network (ANN) models have achieved impressive performance, yet they often rely on dense computation and can degrade when visual inputs are corrupted. Spiking neural networks (SNNs) offer event-driven computation and intrinsic temporal dynamics, which hold potential for compact and robust navigation on resource-limited hardware. However, whether spike-based sensing and policy dynamics can truly enhance robustness in visually rich embodied navigation has remained an open question. This paper introduces SpikingNav, a spiking framework designed for robust indoor embodied navigation, addressing this gap.

SpikingNav comprises two core components: a Spiking Sensing Encoder (SSE) and a Spiking Policy Network (SPN). The SSE employs a spike-based backbone to extract task-conditioned visual features from egocentric observations. The SPN maintains a recurrent policy state through membrane integration, thresholding, and spike-triggered reset, enabling temporal dynamics in decision-making. By integrating these components, SpikingNav exploits the inherent properties of SNNs—spike activation and dynamic state—to improve navigation performance and robustness.

The framework was evaluated on PointNav and ObjectNav tasks under both clean observations and various visual corruptions. Results demonstrate that SpikingNav achieves competitive performance in clean settings while showing significantly stronger robustness under corruptions, all with fewer parameters and lower per-step computation compared to a matched ANN baseline. For instance, on ObjectNav, SpikingNav increases success from 31.05% to 34.12%, and under visual corruptions, the average success rate rises from 8.45% to 13.71%, highlighting the benefits of spike-based sensing and policy dynamics.

To confirm practical deployability, the spike-based sensing method was validated on the Thruster-V2 neuromorphic chip. This physical hardware validation demonstrates that SpikingNav can be instantiated on a real neuromorphic substrate, paving the way for cyber-physical systems that require efficient and resilient navigation. The work thus provides a compelling case for SNNs in embodied agents, suggesting that spike-based approaches can offer robustness advantages without sacrificing performance.