Performance Analysis of Adaptive Dynamic Tube MPC

TL;DR

This paper experimentally compares traditional, dynamic, and adaptive Tube MPC methods, demonstrating that adaptive DTMPC offers superior control efficiency and accuracy for uncertain nonlinear systems like a pendulum.

Contribution

It provides the first comprehensive experimental analysis of DTMPC and ADTMPC, showing their advantages over standard TMPC in real-world uncertain systems.

Findings

DTMPC reduces control effort by up to 30% compared to TMPC.

ADTMPC further reduces control effort by up to 35%.

ADTMPC improves trajectory tracking accuracy by up to 34%.

Abstract

Model predictive control (MPC) is an effective method for control of constrained systems but is susceptible to the external disturbances and modeling error often encountered in real-world applications. To address these issues, techniques such as Tube MPC (TMPC) utilize an ancillary offline-generated robust controller to ensure that the system remains within an invariant set, referred to as a tube, around an online-generated trajectory. However, TMPC is unable to modify its tube and ancillary controller in response to changing state-dependent uncertainty, often resulting in overly-conservative solutions. Dynamic Tube MPC (DTMPC) addresses these problems by simultaneously optimizing the desired trajectory and tube geometry online. Building upon this framework, Adaptive DTMPC (ADTMPC) produces better model approximations by reducing model uncertainty, resulting in more accurate control policies. This work presents an experimental analysis and performance evaluation of TMPC, DTMPC, and ADTMPC for an uncertain nonlinear system. In particular, DTMPC is shown to outperform TMPC because it is able to dynamically adjust to changing environments, limiting aggressive control and conservative behavior to only the cases when the constraints and uncertainty require it. Applied to a pendulum testbed, this enables DTMPC to use up to 30% less control effort while achieving up to 80% higher speeds. This performance is further improved by ADTMPC, which reduces the feedback control effort by up to another 35%, while delivering up to 34% better trajectory tracking. This analysis establishes that the DTMPC and ADTMPC frameworks yield significantly more effective robust control policies for systems with changing uncertainty, goals, and operating conditions.