A Neural Network Control System for the Segway Robotic Mobility Platform

Abstract

An Artificial Neural Network (ANN) is a network of simple processing elements that emulate neurons in the brain. The behavior of such a network is characterized by the synaptic connections between the input data and the processing elements. Here, an ANN was generated and used as part of a control system for a Segway Robotic Mobility Platform (RMP) being trained in obstacle avoidance behavior. The single sensor input to the control system is a SICK laser, a range-finding sensor; the control output is Pulse Width Modulation commands to the RMP's motors. The Segway RMP, neural network maps input sensor data directly to appropriate motor output commands for obstacle avoidance.

Obstacle avoidance training was accomplished in a simulated LabView world using supervised reinforcement learning and practices from evolutionary robotics. Synaptic connection strengths were stored in an array called the artificial "chromosome". The chromosome was randomly modified, and the response of the network was compared to a pre-defined desired output. The goal of the genetic algorithm training was to minimize the error between the desired and actual outputs, yet to ensure that local minima were avoided. Once the ANN was trained in simulation, it was transferred to an actual RMP for obstacle avoidance testing in the real world

. The benefits of training ANN's for obstacle avoidance tasks in simulation are demonstrated here. In the simulated world, training and testing can be done in virtual environments: offering greater control over environment complexity, testing the robustness of the controllers generated, and filtering the training data set. All of the foregoing reduces the cost of training and lead to the development of an optimized ANN controller for RMP obstacle avoidance. The ANN provided input pattern generalization for smooth motion, improved computational speeds, and added to the body of knowledge for RMP controller development.