Data-Driven Neural Network-Based Learning For Regression Problems In Robotics

Abstract

Modeling is an important task in classic control system design. However, as the robotics systems are getting more complex, the modeling tasks using fundamental physical principles are becoming more difficult. One of the emerging approaches to avoid direct modeling is the data-driven techniques in which measurement data are collected, extracted and analyzed by some algorithms for the purpose of modeling and controlling the robots. In this paper, we present a data-driven technique that employs neural network (NN) to approximate robot kinematics without knowing the robot structure. The convergence of the algorithm is rigorously analyzed. Simulation results are presented to illustrate the performance of the proposed algorithm.