Self-Generation of Gait Pattern in Quadruped Walking Robot Using Vector Quantizing Search Method (VQSM)

Abstract

In this paper, we describe the self-generation of gait pattern in quadruped walking robot. The gait of robot is generated by the neural network CPG : Central Pattern Generator, and the robot learns the gait pattern by finding the optimal weights in CPG. In addition to the traditional optimization methods such as GA : Genetic Algorithm, and LSM : Likelihood Search Method, we propose the new optimization method VQSM : Vector Quantizing Search Method, which finds gradient from the finite quantized vectors to be applicable to the discontinuous search space, and which inherits intensification and diversification of search from LSM to find the global optimum in the search space. The simulation result of self-generation of gait pattern shows that VQSM has an advantage of optimization over LSM and an advantage of computational amount over GA. It is also shown that, like the gait of animals, according to walking speed, the optimal gait becomes Bound from Trot.