Optimization of a Simulation Environment for Generic Action Model Acquisition Based on Monocular Camera Image Input

Abstract

This study introduces a training system of action models for mobile robot navigation based on reinforcement learning with monocular camera input. The authors have developed a synthesis system that includes automatically building simulation environments and training of models. The purpose of this paper is to obtain a general-purpose action model by quantitatively verifying the effects of the proposed system. The simulation result shows that the trained model with the proposed system provides stable navigation performance.