Repairing the damaged road network is one of the most important and basic parts of disaster emergency response. It mainly deals with how to effectively dispatch the repair crew to quickly restore the traffic capacity of the damaged road network and provide an effective guarantee for the smooth implementation of the subsequent emergency rescue. However, when the road network is severely damaged, the existing algorithms often fail to find a feasible solution. Therefore, this paper first simplifies models of damaged road network and decision-making on the basis of the existing work. Then, an improved algorithm for repair crew scheduling on severely damaged road network is developed according to Q-learning and action set reduction. Particularly, in the proposed algorithm, the repair crew can only choose the next action from the set of current damaged road sections which are unrepaired but reachable, ensuring the continuity of Q-learning. Finally, simulation results show that the proposed algorithm can ensure that all demand nodes are reachable, has higher stability and reliability, and can obtain better scheduling schemes at lower time and repair cost, even if the road network has been seriously damaged with a great number of damaged nodes and a big damage rate.