The usual human behavior learning methods are in the robot task space, i.e., three-dimension Cartesian space. After learning, the desired trajectories have to be transformed into the joint space by the inverse kinematics of the robot. However, for most robots, the analytical solutions of the inverse kinematics cannot be obtained. In this paper, we learn human behavior directly in the joint space. There are some problems to learn the demonstrations in the joint space, such as the demonstration trajectories depending on different velocities and tremors in Cartesian space produced by natural human behavior. We use dynamic time warping and neural networks to solve these problems. More importantly, we avoid calculating the inverse kinematics. Experiment results have shown this method to be effective.