Enforcing constraints over learned policies via nonlinear MPC: application to the pendubot

G. Turrisi; B. Barros Carlos; M. Cefalo; V. Modugno; L. Lanari; G. Oriolo

doi:10.1016/j.ifacol.2020.12.2426

Enforcing constraints over learned policies via nonlinear MPC: application to the pendubot

G. Turrisi
, B. Barros Carlos
, M. Cefalo
, V. Modugno
, L. Lanari
, G. Oriolo

University of Rome La Sapienza

Research output: Contribution to journal › Conference article › peer-review

Abstract

In recent years Reinforcement Learning (RL) has achieved remarkable results. Nonetheless RL algorithms prove to be unsuccessful in robotics applications where constraints satisfaction is involved, e.g. for safety. In this work we propose a control algorithm that allows to enforce constraints over a learned control policy. Hence we combine Nonlinear Model Predictive Control (NMPC) with control-state trajectories generated from the learned policy at each time step. We prove the effectiveness of our method on the Pendubot, a challenging underactuated robot.

Original language	English
Pages (from-to)	9502-9507
Number of pages	6
Journal	IFAC-PapersOnLine
Volume	53
Early online date	21 Apr 2021
DOIs	https://doi.org/10.1016/j.ifacol.2020.12.2426
Publication status	Published - 21 Apr 2021
Externally published	Yes
Event	21st IFAC World Congress 2020 - Berlin, Germany Duration: 12 Jul 2020 → 17 Jul 2020

Bibliographical note

Journal issue date: 2020.

Keywords

Constraint satisfaction problem
Learning control
Model based control
Optimal control
Real time
Robotics

Access to Document

10.1016/j.ifacol.2020.12.2426Licence: Other

Cite this

@article{d44a791c85834837b2341a13f175b12c,

title = "Enforcing constraints over learned policies via nonlinear MPC: application to the pendubot",

abstract = "In recent years Reinforcement Learning (RL) has achieved remarkable results. Nonetheless RL algorithms prove to be unsuccessful in robotics applications where constraints satisfaction is involved, e.g. for safety. In this work we propose a control algorithm that allows to enforce constraints over a learned control policy. Hence we combine Nonlinear Model Predictive Control (NMPC) with control-state trajectories generated from the learned policy at each time step. We prove the effectiveness of our method on the Pendubot, a challenging underactuated robot.",

keywords = "Constraint satisfaction problem, Learning control, Model based control, Optimal control, Real time, Robotics",

author = "G. Turrisi and \{Barros Carlos\}, B. and M. Cefalo and V. Modugno and L. Lanari and G. Oriolo",

note = "Journal issue date: 2020.; 21st IFAC World Congress 2020 ; Conference date: 12-07-2020 Through 17-07-2020",

year = "2021",

month = apr,

day = "21",

doi = "10.1016/j.ifacol.2020.12.2426",

language = "English",

volume = "53",

pages = "9502--9507",

journal = "IFAC-PapersOnLine",

issn = "2405-8971",

publisher = "Elsevier B.V.",

}

TY - JOUR

T1 - Enforcing constraints over learned policies via nonlinear MPC

T2 - 21st IFAC World Congress 2020

AU - Turrisi, G.

AU - Barros Carlos, B.

AU - Cefalo, M.

AU - Modugno, V.

AU - Lanari, L.

AU - Oriolo, G.

N1 - Journal issue date: 2020.

PY - 2021/4/21

Y1 - 2021/4/21

N2 - In recent years Reinforcement Learning (RL) has achieved remarkable results. Nonetheless RL algorithms prove to be unsuccessful in robotics applications where constraints satisfaction is involved, e.g. for safety. In this work we propose a control algorithm that allows to enforce constraints over a learned control policy. Hence we combine Nonlinear Model Predictive Control (NMPC) with control-state trajectories generated from the learned policy at each time step. We prove the effectiveness of our method on the Pendubot, a challenging underactuated robot.

AB - In recent years Reinforcement Learning (RL) has achieved remarkable results. Nonetheless RL algorithms prove to be unsuccessful in robotics applications where constraints satisfaction is involved, e.g. for safety. In this work we propose a control algorithm that allows to enforce constraints over a learned control policy. Hence we combine Nonlinear Model Predictive Control (NMPC) with control-state trajectories generated from the learned policy at each time step. We prove the effectiveness of our method on the Pendubot, a challenging underactuated robot.

KW - Constraint satisfaction problem

KW - Learning control

KW - Model based control

KW - Optimal control

KW - Real time

KW - Robotics

U2 - 10.1016/j.ifacol.2020.12.2426

DO - 10.1016/j.ifacol.2020.12.2426

M3 - Conference article

AN - SCOPUS:85107824738

SN - 2405-8971

VL - 53

SP - 9502

EP - 9507

JO - IFAC-PapersOnLine

JF - IFAC-PapersOnLine

Y2 - 12 July 2020 through 17 July 2020

ER -

Enforcing constraints over learned policies via nonlinear MPC: application to the pendubot

Abstract

Bibliographical note

Keywords

Access to Document

Fingerprint

Cite this