Associative memory used for trajectory generation and inverse kinematics problem

Abstract

Proposes a neural network system to perform trajectory generation and inverse kinematics. Such a system is composed of two neural network blocks based on associative memory principles. The first block is formed by the model called temporal multidirectional associative memory (TMAM). This block is responsible for producing a desired spatial trajectory given part of it. The second block includes a radial basis function (RBF) model that provides a set of joint angles associated with the trajectory. The system has a fast training stage, is able to interpolate and extrapolate points to a trained trajectory is able to deal with multiple trajectories, and is able to produce viable joint angles even if the spatial position slightly violates the robot constraints. So far, the RBF model was tested only for single trajectories.