This paper presents a technique for the estimation of the sideslip angle and longitudinal speed of a high-performance four-wheel drive vehicle by means of Artificial Neural Networks. The proposed architecture relies on the combination of a pattern recognition neural classifier with two stages of cascaded regression neural networks. The classifier allows identifying the road condition and the regression stages perform the two estimation tasks. The networks are trained by means of datasets recorded on an instrumented vehicle. The strategy is tested in different road adherence conditions, namely dry, wet and icy, and in a variety of driving maneuvers. The effectiveness of the technique is demonstrated experimentally by deploying the algorithm on the vehicle and comparing the sideslip angle estimation with the measurement computed by an optical sensor and the longitudinal speed estimation with both the estimation already present on the vehicle and obtained as the mean value of the wheels’ velocity and with the optical sensor measurement.