Research On Evaluation Model Of CPG Self-growing Network And Strategies Of Network Correction

In the field of robot control, the CPG neural network which is designed out based on the goals of control and applied for specific occasions has many advantages such as less control variables, strong rhythm and preserving the complexity of the output pattern, therefore CPG is widely used in locomotion control of robots. However, this kind of artificial CPG neural network emphasize function more than the rationality on biological side, once new control objects are proposed, the network has to be designed again at basic levels, which resulting in poor network function adaptability and inheritance.With the development of bionics and bio-neurological, scholars have put forward that combine the growth and development process of CPG networks with the actual control together, determine the parameters of the control model into the growing model and achieve the control objectives on the premise of corresponding to biological realities. However, the current research in this field is still in initial stage, and most of them are still in morphologic and statistical aspects, thus, it lacks the significance of practical application.This paper based on CPG self-growing model, according to the analysis of the characteristics of animal movement, the evaluation model of self growing network is established and the optimization algorithm which meet the needs of quadruped robot control and kalman filtering mechanism is proposed.Firstly, analyze the relationship of the walking speed with energy consumption and gait, determined when the total energy consumption is minimized, the coordinated relationship between the optimal walking pattern and legs under different walking velocity of quadruped mechanisms.Furthermore, we put out the derivative graph of different walking mode conversion between the gait. Based on it, we established the self growth evaluation model of network output signal, and ensured the continuity of the gait transition.Secondly, on the basis of the relationship between the growth parameters of the network and the output of the control signal, found the signal transmission path which satisfied the above evaluation model via deep optimized search algorithm; based on it, We built a self-predict correction model combined with the characteristics of the growth and connection of the biological neurons and the actual control characteristics of the four legged mechanisms, to guide the formation of the self growing neural network with the output of trot at different speeds; then, a phase adjustment mechanism based on the perception of the body is proposed by using the robot’s perception of the state of travel.Next, based on the idea of Kalman filter, the combined,optimized and updating model was established which was combined with self-predict correction and environmental feedback correction. We conducted simulation experiments for practical examples which verified the algorithm proposed in this paper has favorable approximation ability and fast correction ability.Finally, we set up the simulation platform of the quadruped robot and conducted joint simulation experiment of Matlab and Adams, regard the self growing network output signal as the control signal of the hip joint.The results of experiments verified the self growth network correction algorithm is effectiveness for walking and learning of robots, the following character for speed command and quick adjustment ability for the changes of environments. Based on it, the hip joint control curve resulted from simulation was applied in specific Rhex robot prototype, and the experimental results verified the validity of self growth network output signal can be applied to actual robot control.