Unmanned surface vehicle (USV) is a kind of important marine autonomous robots, which has been studied and applied to practice gradually. However, the autonomy of USV is still restricted by the performance of autonomous navigation technology. Especially, the problem of adaptive obstacle avoidance in complicated sea-state marine environments needs to be solved urgently. In the paper, an adaptive avoidance decision process model is proposed for USV to solve the problem of obstacle avoidance in complicated sea-state marine environments. By analyzing the disturbance factors from complicated sea-state marine environments, the model is constructed on the basis of Sarsa on-policy reinforcement learning algorithm. By setting the GLIE (greedy in the limit and infinite exploration) as the action exploration, the convergence of the adaptive avoidance decision process has been proved. The convergence shows that the action can converge to the optimal action strategy with the probability value of one. The proved result demonstrates that the performance of obstacle avoidance of USV in the complicated sea-state marine environment can be enhanced under the action of on-policy reinforcement learning algorithm.