Decentralised iterative learning control for high performance collaborative tracking and formation objectives

Abstract

Multi-agent systems have become highly popular in recent years due to their wide range of applications spanning industrial, military, domestic and research domains. Two common goals are for the agents to attain a physical separation (a formation) or for their states/outputs to combine to equal a specified target (a collaboration). Many multi-agent systems repeat the task over and over again, and a common requirement is to achieve the desired objective with the highest accuracy possible. Iterative learning control (ILC) is a well-known approach that enables systems that repeat the same task to improve performance by using data collected over previous attempts. ILC has been successfully applied to collaborative and formation control problems, but unfortunately existing designs have important limitations. Firstly, algorithms have addressed each problem separately, not enabling collaborative and formation objectives to be amalgamated. Secondly, existing approaches for formation control tackle restricted classes of system and/or communication structures, and are not model based. They therefore can only deliver slow, asymptotic convergence. A model based controller has been developed to solve the collaboration problem, however, it can only be used with simple (relative degree zero) dynamics and is not robust to model uncertainty. Thirdly, there is little analysis of how robust existing schemes are to model uncertainty, nor clear mechanisms that would enable the designer to trade robustness against performance. This thesis develops a powerful new ILC framework to solve these limitations. It is applicable to the general class of linear agent dynamics, and addresses common forms of communication architecture. The framework simultaneously combines both tracking and formation control, and equips the designer with the ability to choose from a whole class of updates rather than a single specified algorithm. A key benefit is that the update structure is decentralised, thereby simplifying design and broadening the range of potential application areas. Convergence properties and robust performance properties are derived, and are used to develop a comprehensive design procedure to transparently balance practical trade-offs. The control structures are then experimentally applied to rehabilitation engineering, where each agent corresponds to an electrically stimulated muscle with the global goal of assisting movement. Experimental and simulation results confirm accurate assistance of human movement and illustrate the utility of decentralised control to substantially reduce hardware and communication overheads.

More information

Identifiers

ORCID for Shangcheng Chen: orcid.org/0000-0001-8032-2779

ORCID for Christopher Freeman: orcid.org/0000-0003-0305-9246

Catalogue record

Export record