In order to obtain the feasible solution in the realistic learning time, a layer architecture is often introduced. This paper proposes a behavior selection mechanism with activation/termination constraints. In our method, each behavior has three components: policy, activation constraints, and termination constraints. A policy is a function mapping the sensor information to motor commands. Activation constraints reduce the number of situations where corresponding policy is executable, and termination constraints contribute to extract meaningful behavior sequences, each of which can be regarded as one action regardless of its duration. We apply the genetic algorithm to obtain the switching function to select the appropriate behavior according to the situation. As an example, a simplified soccer game is given to show the validity of the proposed method. Simulation results and real robots experiments are shown, and a discussion is given.