Research On Controlling Method Of Quadruped Robot Based On Worm Neural Networks

The nervous system architecture of the small nematode Caenorhabditis elegans shows that a simple way to apply animal intelligence in the robotics domain.We present in this thesis the control of the ant quadruped locomotion based on Pybullet framework.Although the augmented random search algorithm has achieved remarkable results and has been successfully applied to various forms of large-scale tasks,there are still some areas for improvement.In this paper,aiming at the control problem of quadruped ant robot in continuous time environment,we propose a control method combining random search algorithm with worm neural network,which makes the movement of ant have certain interpretability.The main contents of this thesis are as follows:First,we introduce on the worm C.elegans,and we have given the background corresponding to this animal,and we have mainly described the hypothesis concerning the locomotion of C.elegans.We also talk about bionics application in the quadruped robot field.Second,we talked about the theory and technology are used in reinforcement learning,and we used the worm brain to train the ant quadruped robot.Third,we have review quadrupeds among robots by giving an overview of the evolution.We have also described the mathematical model of structure and locomotion of ant quadruped robot in the PyBullet environment.Fourth,taking the ant quadruped robot as an example,the function and mapping process of tap extraction circuit are illustrated.The ant quadruped robot consists of 111 inputs and 8 outputs.In order to control the ant quadruped robot,we need to control its 8 outputs.First,we use a linear layer to map 111 input variables into two continuous variables,and then input them into two sensing neurons of tap-withdraw circuit(TW).Finally,we use a linear layer to map the neural potential of two motor neurons to eight control outputs.That is,the control of eight outputs of the ant quadruped robot is completed.Finally,in view of all the above,we have shown the interpretability of the worm's neural network by watching the results during the control of the PyBullet ant,the worm neural network has demonstrated the better performance as a simple nature-inspired artificial neural network architecture.The experiments show that neural circuit policies resulting from the combination of the Augmented Random Search algorithm and the worm network give the better performance than the Ant control with the sole use of the Augmented random search algorithm.