Automation in outdoor applications (farming, surveillance, military activities, etc.) requires highly accurate control of mobile robots, at high speed, although they are moving on low-grip terrain. To meet such expectations, advanced control laws accounting for natural ground specificities (mainly sliding effects) must be derived. In previous work, adaptive and predictive control algorithms, based on an extended kinematic representation, have been proposed. Satisfactory experimental results have been reported (accurate to within ±10 cm, whatever the grip conditions), but at limited velocity (below 3 m·s-1). Nevertheless, simulations reveal that control accuracy is decreased when vehicle speed is increased (up to 10 m·s-1). In particular, oscillations are observed at curvature transition. This drawback is due to delays in sideslip angle estimation, unavoidable at high speed because only an extended kinematic representation was used. In this paper, a mixed backstepping kinematic and dynamic observer is designed to improve observation of these variables: the slow-varying data are still estimated from a kinematic representation, which is then injected into a dynamic observer to supply reactive and reliable sliding variable (namely sideslip angle) estimation, without increasing the noise level. The algorithm is evaluated via advanced simulations (coupling Adams and MatLab software) investigating high-speed capabilities. Actual experiments at lower speed (experimental platform maximum velocity) demonstrate the benefits of the proposed approach.