TY - JOUR
T1 - Reinforcement Learning for Position Control Problem of a Mobile Robot
AU - FARIAS CASTRO, GONZALO ALBERTO
AU - Garcia, Gonzalo
AU - Montenegro, Guelis
AU - Fabregas, Ernesto
AU - Dormido-Canto, Sebastian
AU - Dormido, Sebastian
N1 - Publisher Copyright:
© 2013 IEEE.
Copyright:
Copyright 2020 Elsevier B.V., All rights reserved.
PY - 2020
Y1 - 2020
N2 - Due to the increase in complexity in autonomous vehicles, most of the existing control systems are proving to be inadequate. Reinforcement Learning is gaining traction as it is posed to overcome these difficulties in a natural way. This approach allows an agent that interacts with the environment to get rewards for appropriate actions, learning to improve its performance continuously. The article describes the design and development of an algorithm to control the position of a wheeled mobile robot using Reinforcement Learning. One main advantage of this approach concerning traditional control algorithms is that the learning process is carried out automatically with a recursive procedure forward in time. Moreover, given the fidelity of the model for the particular implementation described in this work, the whole learning process can be carried out in simulation. This fact avoids damages to the actual robot during the learning stage. It shows that the position control of the robot (or similar specific tasks) can be done without the need to know the dynamic model of the system explicitly. Its main drawback is that the learning stage can take a long time to finish and that it depends on the complexity of the task and the availability of adequate hardware resources. This work provides a comparison between the proposed approach and traditional existing control laws in simulation and real environments. The article also discusses the main effects of using different controlled variables in the performance of the developed control law.
AB - Due to the increase in complexity in autonomous vehicles, most of the existing control systems are proving to be inadequate. Reinforcement Learning is gaining traction as it is posed to overcome these difficulties in a natural way. This approach allows an agent that interacts with the environment to get rewards for appropriate actions, learning to improve its performance continuously. The article describes the design and development of an algorithm to control the position of a wheeled mobile robot using Reinforcement Learning. One main advantage of this approach concerning traditional control algorithms is that the learning process is carried out automatically with a recursive procedure forward in time. Moreover, given the fidelity of the model for the particular implementation described in this work, the whole learning process can be carried out in simulation. This fact avoids damages to the actual robot during the learning stage. It shows that the position control of the robot (or similar specific tasks) can be done without the need to know the dynamic model of the system explicitly. Its main drawback is that the learning stage can take a long time to finish and that it depends on the complexity of the task and the availability of adequate hardware resources. This work provides a comparison between the proposed approach and traditional existing control laws in simulation and real environments. The article also discusses the main effects of using different controlled variables in the performance of the developed control law.
KW - Mobile robot
KW - position control
KW - reinforcement learning
UR - http://www.scopus.com/inward/record.url?scp=85090599506&partnerID=8YFLogxK
U2 - 10.1109/ACCESS.2020.3018026
DO - 10.1109/ACCESS.2020.3018026
M3 - Article
AN - SCOPUS:85090599506
VL - 8
SP - 152941
EP - 152951
JO - IEEE Access
JF - IEEE Access
SN - 2169-3536
M1 - 9171241
ER -