Abstract: Q-learning algorithm has a fundamental flaw,that is,prone to error intervals overlap,and thus overestimation of the correct Q-value.These are likely to lead to low convergence speed and continuous decline in the performance of Q-learning,an improved Q-learning algorithm was proposed,that was rough sets Q-learning algorithm.The algorithm can be able to minimize the overestimation caused by Q-values and improve performance of learning through effectively deal with incomplete information and uncertain knowledge.Navigation experiments based on these two algorithms were conducted,the results showed that rough sets Q-learning algorithm had higher efficiency of learning and stronger ability of obstacle avoidance than Q-learning algorithm.