To make the robot avoid obstacles in time while performing complex tasks, an interactive learning approach for robot obstacle avoidance based on DP-KMP is proposed. Firstly, the whole framework of this approach is constructed, which adopts the segmentation-generalization strategy, to implement the learning of demonstrated trajectories with rapid segmentation and the planning of sub-trajectories for obstacle avoidance. In the learning phase, a trajectory segmentation strategy based on the DP algorithm is proposed to improve the efficiency of segmentation, and Gaussian mixture model strategy is used to extract the reference database from each sub-trajectory. In the trajectory planning phase, the KMP model is used to implement the trajectory reproduction and generalization, while the reference database update strategy based on humanrobot interaction feedback is introduced, to enhance the success rate of human-robot interaction for obstacle avoidance. Aiming at the issue that this update strategy may be ineffective to cause failure in planning the trajectory for obstacle avoidance, two available conditions are proposed for inspecting the sub-trajectories generated by segmentation. Finally, the effectiveness of mentioned available conditions is verified by simulation, respectively. The real experimental results show that it takes only 0.084 s and 0.107 s, respectively, to segment the demonstrated trajectories of the two experiments using the proposed approach, and KUKA cobot successfully avoids all static and suddenly changing obstacles through multiple interactions with user during the execution of the different lift-place tasks.