Map-less Navigation in Novel Environments using Deep Reinforcement Learning: A Curiosity-driven approach for Mobile Robotics

Rasmussen, Anders Frederik Nørby

4. term

Global Systems Design (cand.tech.), Master

2019

2019-06-03

51 pages

Dette projekt undersøger mulighederne for at udvikle autonome robot agenter der kan navigere i usete miljøer vha. \textit{deep reinforcement learning}. For at træne agenterne udvikles der et simuleret miljø med Unity, som anvender principerne af \textit{domain randomization}. For at anspore agenterne til at undersøge usete områder, implementeres der et \textit{Intrinsic Curiosity Modul} (ICM). En \textit{Deep Recurrent Q-network} (DRQN) agent og en DRQN agent med et ICM trænes i en indskrænket version af miljøet for at teste systemets potentiale. Under træningen udnytter agenterne tricket \textit{Learning from Easy Missions} samt det foreslåede trick \textit{Naive Visual Hindsight Experience Replay}.

This project investigates the possibilities of creating an autonomous robot deep reinforcement learning agent for maples navigation in novel environments. To train the agent a simulated environment is created with Unity, that applies the principles of domain randomization. An \textit{Intrinsic Curiosity Module} (ICM) is implemented, to encourage the agent to explore unseen areas of the environment. A \textit{Deep Recurrent Q-network} (DRQN) agent and a DRQN agent with an ICM are trained on a limited version of the environment to explore the system's potential. During the training the agents use the trick of \textit{Learning from Easy Missions} along with the proposed trick of \textit{Naive Visual Hindsight Experience Replay}.

Download
View record in AAU Student Projects