With the development of intelligent robot navigation technology, much more attention has been paid on the research of path planning. An improved gray wolf optimization(GWO) algorithm was proposed for local optimum and slow convergence seed in robot path planning. Firstly, a two-dimensional spatial model was used to imitate the path planning process of robots. In the global search process, a dynamic perturbation coefficient and nonlinear decreasing control parameters(which are improved from linear decreasing) were proposed in the formula of position updating for the aim of enhancing global search performance. Then, considering the diversity of gray wolf population and local mining ability of the algorithm, a reverse learning selection strategy was introduced, which led to the improvement of convergence accuracy of the algorithm. The testing data from eight common functions has shown the effectiveness of the improved algorithm. Finally, a comparative experiment has been made among the improved gray wolf optimization algorithm, the original gray wolf optimization algorithm and particle swarm optimization algorithm, and the data has shown that the average path distance is shortened by 11.99%, 7.79% and 5.78%, and iteration times is reduced by 75.63%, 59.78% and 43.67%, respectively in simple, general and complex environments, which indicates the effectiveness in optimal distance planning and obstacle avoidance of this improved algorithm.