To solve the problems of slow convergence speed and poor stability of reinforcement learning method used on mobile robot for path planning tasks, an improved reinforcement leaning algorithm is proposed. In this paper, the algorithm allow the agent to move with 8 optional self-adapting directions. Also, we apply route expansion method to initialize the final route. Finally, we set dynamic exploration factor to accelerate the convergence of the final route. Simulation experiments undertaken on grid maps created by canvas indicates that the improved reinforcement learning can largely increase the speed of the convergence for the final path comparing to the classic Q-learning algorithm and A * algorithm, which has highly application value in the future.