Locomotion behavior modeling is a key and difficult problem in the research field of robotic fish. In this paper, a novel biomimetic learning approach is applied to the movement patterns generation of a multi-joint robotic fish. Firstly, the real fish behaviors are recorded by video system. By analyzing the recorded data, three basic swimming patterns, “cruise”, “cruise in turning” and “C sharp turn”, are extracted. Then the general internal model (GIM) is employed to learn the swimming patterns of carangiform fishes. Based on the approximation ability and the temporal-spatial scalabilities of GIM, robotic fish owns the capability of learning, modifying and regenerating the similar swimming patterns of the real fish. Finally, autonomous obstacle avoidance behavior of the robotic fish is realized by combining infrared sensors and obstacle avoidance algorithm. The experiment result verifies the validity of the proposed biomimetic learning approach.