Goal-seeking Navigation based on Multi-Agent Reinforcement Learning Approach

Abstract

Navigation for the robot has numerous applications in industries such as agriculture, couriers, autonomous vehicle, and many more. Navigation seems a simple problem for humans but complex for machines. Robot navigation problems can be made up of mapping, localization, path planning, and motion control. But there are limitations when it comes to the real world. The nature of dynamics in the real world makes it hard to adapt to certain situations since a lot of conventional controllers cannot adapt. This is where machine learning can be used since it can adapt according to the situations it has learned. The closed-loop control systems have a feedback loop, and Reinforcement Learning is most suitable to develop where this paper explores how to develop a controller using deep reinforcement learning. The algorithm used is the Deep Deterministic Policy Gradient algorithm with multi-agents. The simulated robot inside the environment has a simple dynamics system for simplicity's sake for this proposed project, and the algorithm is designed to work with the environment developed by a third-party using OpenAI Gym. The robot is controlled using the neural network or controller network with the same as the multi-input multi-output system. The neural network was designed and also constructed using (Full abbreviation) LSTM with a fully connected neural network. The algorithm and evaluation of the model will be part of future works to achieve the objectives of the proposed project.