Robotic motion control using machine learning techniques

摘要

This paper proposes a new technique for path planning of mobile robotic locomotion using a multi-layered Auto Resonance Network (ARN). Architecture of these networks is different from the Convolutional Neural Networks and other related structures used in Deep Learning methods for image recognition and game playing. The proposed network can search through space to find multiple paths around obstacles. They can also be used for solving the Movers' Problem in a work area cluttered with obstacles. When the network is used for joint control, required joint angles and torque can be interpolated without any need for computation of non-linear inverse kinematic expressions generally used for such problems. The proposed structure combines features of Auto Resonance Network and Self Organizing Maps. Cells in lower layers map input to output using a ARN like structure. These nodes are perturbed to generate a local SOM like structure. Higher layers can identify and optimize the paths that can be used to solve motion problems. These ANNs have been implemented using R simulation language. Results of the implementation for three segment joint with six Degrees of Freedom (DoF) are presented in this paper.