Quasi-Dynamic Walking Optimization of Humanoid Robot Using Genetic Algorithm

Abstract

Humanoid robot has been developed in design methods and functionality in recent years. In its application, a humanoid robot is required to interact with human, tools, or the environment. For a bipedal humanoid robot, efficient, precise, and stable walking is required, where the humanoid robot is expected to walk in a bigger step in a predetermined direction without falling. In this paper, the Genetic Algorithm (GA) is implemented to optimize the quasi-dynamic walking of a humanoid robot. The walking is optimized in terms of distance and precision while keep considering stability. For this purpose, a 10-DoF humanoid robot is designed and constructed to resemble a half-body of a human, from waist to feet. The humanoid robot is built of metal brackets where 10 servo motors are integrated for a coordinated movement. The walking gait of the humanoid robot for one complete walking cycle of one right step and one left step is divided into 8 walking phases. In each walking phase, the input to the 10 servo motors can be set whether with the same value as the previous phase or with a new value. The GA takes all possible new input values to the servo motors as the genes of an individual. At the population initialization, the first individual that can move the humanoid robot with adequate stability is found by using the forward kinematics method. Five individuals are derived from the first individual through mutation with the rate of 40-60 %. Thus, the GA starts with an initial population of these 6 individuals. A novel fitness function is introduced with positive weight on straightforward displacement and negative weight on deviation. This also emphasizes the merit of this research in the quantification of a robot's walking performance. The GA cycles include the uniform crossover with 25 % gene exchange probability and the mutation with the rate of 10 %. The GA is conducted for 4 cycles, where every individual is tested on the humanoid robot 10 times. In each test, the humanoid robot performs 3 complete walking cycles and the fitness score is assessed. The GA is successful to increase the fitness score of the population’s best individual from 12.02 to 25.42. The walking distance is increased by 26.12 % from 25.33 cm to 31.94 cm, while the deviation angle is reduced by 57.65 % from 25.39° to 10.75°. Further application of the proposed method is to obtain the best individual for the robot to walk in a certain direction, which will possible by adjusting the fitness function. This is to be done with the support of sensor feedback and reverse kinematics in the robot’s modeling.