Neural Networks with Physical Meaning: Representation of Kinematic Equations from Robot Arms Using a Neural Network Topology
Raphael C.O. Jesus, Eduardo Oliveira Freire, Lucas Molina, Elyson Á. N. Carvalho · 2018
This paper proposes a new way of representing the kinematic equations of a manipulator robot in a neural network topology, having link lengths and joint offsets as neural weights. The proposed method of constructing a network from a set of equations is generic for models obtained through the Denavit-Hartemberg convention and can be done directly, by only knowing the number of joints and their planes of movement. As an application for the proposed method, a simple neural network learning algorithm is implemented in order to recover the unknown structural parameters from a robotic arm directly from the weight matrices.