Prof. Hyo-Sang Shin’s research team at KAIST developed a brain-inspired model that helps autonomous systems determine their positions. By combining neuroscience with probabilistic reasoning, the model maintains reliable estimates even when sensor information strongly conflicts with motion predictions.

As humans or animals move through space, their brains continuously estimate where they are by combining different sensory cues. One model that helps explain this process is the Continuous Attractor Neural Network (CANN), in which spatial information appears as a localized “bump” of neural activity. This can be imagined as a small hill moving across an internal map: the hill’s location indicates the estimated position, while its height reflects how certain the estimate is.

Just as biological systems continuously track their location, autonomous vehicles or robots must also accurately estimate their position. They must answer a basic question: where am I now? Under ordinary conditions, systems estimate the position quite well by combining motion predictions with sensor measurements. However, critical issues arise when sensor observations conflict with internal motion predictions. Under such conflicts, conventional Gaussian filters combine the prediction and observation into a single compromise belief, which often shifts the estimated location in the wrong direction. Particle filters attempt to resolve this by representing multiple hypotheses. Yet, their finite particles frequently become depleted when the prediction and observation diverge significantly.

To solve this problem, a research team led by Prof. Hyo-Sang Shin developed a 2D Bayesian CANN for robust spatial state estimation (see Figure 1 for a conceptual visualization). In the network, the location of the activity bump represents the estimated position, while its shape and strength reflect the confidence of the estimate. The team further showed analytically that the CANN dynamics can be matched to an information-form Kalman filter. This connection turns the brain-inspired network into a principled Bayesian estimator rather than merely a qualitative model of neural activity.

This new brain-inspired framework overcomes the limitations of conventional filters during severe sensory conflicts. When sensor information conflicts with motion predictions, the Bayesian CANN allows competing evidence to coexist across its continuous field rather than forcing an immediate compromise. Nonlinear attractor dynamics then suppress weak, dispersed activity and reinforce a coherent estimate. In random traction-loss simulations, it achieved the lowest position error, reducing slip-phase error by 61.7% versus a standard 2D Kalman filter and 24.2% versus an optimal-proposal particle filter.

From an engineering perspective, the significance of this work is that CANNs can be understood not only as brain-inspired models, but also as structured Bayesian estimators for robust state estimation. The framework offers a different way to represent and resolve uncertainty, helping autonomous systems maintain reliable state estimates under unpredictable conditions.

This study is currently under revision at Nature Communications. The preprint article, titled "A Two-Dimensional Bayesian Continuous Attractor Neural Network for Robust Spatial State Estimation," is available online (https://doi.org/10.21203/rs.3.rs-9897709/v1).