We present a unified neural solver for multi-task VRPs using mixture of specialized experts based on the state-decomposable MDP.