Abstract: The curved surface climbing gait is the key of wall-climbing robots for detecting the wind turbine blades. A mechanical model of an inchworm-like climbing robot with 5 DOF is designed, including three T type joints and two I type joints. The adaptability of the vacuum sucker and mechanical model to the arc surface is analyzed based on geometric method. The fliping gaits on curved surface are analyzed by the circular trajectory planning method. The function relationship between the joint angle and curvature radius is established based on the stable adsorption state. Based on the supervised learning method，the adaptive frequency Hopf oscillator and Kuramoto couple, a central pattern generator (CPG) module corresponding to a joint and a CPG network of robot are designed. The steady-state values of the parameters which are learned from the flat flip gait can be the initial values of the parameters of CPG network, and then the flip gait on curved surface is planned by adjusting the angle amplitudes online. The coordinated simulation of the curved surface flip gait based on Matlab and Adams is carried out，which validates the stability of the proposed CPG network. The physical prototype robot is developed, and the flip gait experiments on curved surface are carried out. The results show that the angle amplitude gained from the stable adsorbing status can be used to transfer the plane gait to curved surface gait, and the online adjustment CPG planning is valid.

Key words: control science and technology, wall-climbing, blades curved surface, Hopf oscillator, gait planning, central pattern generator