Within RAMSES, the Farinella CubeSat will approach and characterize near-Earth asteroid (99942) Apophis, requiring robust onboard centroid estimation for autonomous line-of-sight determination during proximity operations. The center of brightness is simple but degrades at high phase angles, where the visible surface is partly unilluminated, causing a systematic bias relative to the center of figure. Closed-form analytical corrections partly compensate but lose robustness for non-spherical shapes, extreme illumination, and unavailable or unreliable range information. Parametric fitting remains sensitive to a-priori shape uncertainty, while image-based convolutional networks exceed CubeSat resources and pose black-box verification and validation challenges. We propose NeuralCOB, a compact multilayer perceptron with 10,562 trainable parameters. Rather than the full image, it uses hand-crafted geometric and photometric features, namely phase and Sun-direction angles, bounding box, and brightness moments, mapped to an enhanced centroid estimate. It requires no externally supplied range. Across synthetic datasets spanning Sun–spacecraft–target geometries and Apophis shape realizations, NeuralCOB consistently outperforms the analytical centroid baseline, especially at high phase angles and for strongly non-spherical shapes where analytical assumptions fail. Explainability analyses show which inputs dominate the learned correction, providing evidence for flight qualification. NeuralCOB will fly as a navigation experiment and as backup solution onboard Farinella.