Multi-Objective Biped Robot Gait Control using Modified Continuous Ant Colony Optimized Recurrent Neural Networks

• Main Authors: Yen-Ting Yeh, 葉晏廷
• Other Authors: Chia-Feng Juang
• Format: Others
• Language: en_US
• Published: 2014
• Online Access: http://ndltd.ncl.edu.tw/handle/02630485508891758117

碩士 === 國立中興大學 === 電機工程學系所 === 102 === This thesis proposes optimization of a fully connected recurrent neural network (FCRNN) using modified continuous ant colony optimization (MCACO) for gait control of a forward-moving biped robot, the NAO. There are five degrees in each leg of the robot. The FCRNN is optimized to control the hip roll and pitch, the knee pitch, and the ankle pitch of a leg. The other angles are obtained by using the symmetry property of the walking posture. Two CACO-based learning approaches, supervised learning and evolutionary learning, are proposed. For supervised learning, the desired angle trajectories of a leg are collected in advance. Two CACO-based algorithms are used to optimize an FCRNN to generate trajectories that follow the desired ones based on error minimization. Performances of the two algorithms are analyzed with different numbers of nodes in the FCRNN. For the evolutionary learning approach, no supervised training data are collected in advance. The performance of an FCRNN is evaluated based on multi-objective functions. Five objective functions with imposed constraints are defined to evaluate the walking speed, trajectory straightness, oscillation, walking posture, and stability of the robot. For this multi-objective optimization problem, a multi-objective MCACO is proposed to find the Pareto optimal solutions. Simulations are conducted using the webots robot simulator. The software-designed FCRNN controllers are then applied to control the gait of a real NAO robot. To show the advantage of the MCACO optimization performance, comparisons with other swarm intelligence and genetic algorithms are conducted.