Towards a Unified Framework of Efficient Algorithms for Numerical Optimal Robot Control

Abstract

In the recent past, the wide availability of digital technologies has strongly disrupted many well-established manufacturing techniques and initiated a rapid transformation process in the corresponding industry sectors.However, there are some domains which to date seem to have less profited from digitization.One such domain is building construction and civil engineering.The emerging research field digital fabrication promises a revolution in the construction industry through digitization and robotization of design and manufacturing processes and exhibits a great potential for novel construction technologies and architectural approaches.In this thesis, we introduce the concept of an In situ Fabricator, a versatile and flexible mobile robot dedicated to on-site fabrication.We present a prototype system, show several example applications and motivate the necessity of high-performance optimal control and estimation algorithms for achieving desired construction goals.Such methods enable non-expert users to operate complex robotic systems and are an important building block for advanced autonomous capabilities.For this reason, the main objective of this work is the development of efficient optimal control algorithms and software for high-dimensional robotic systems.We introduce a family of multiple shooting algorithms which exploit the sparsity of the optimal control problem and offer interesting properties for motion planning and nonlinear model predictive control.We derive equalityconstrained versions of these algorithms and demonstrate their potential for motion planning and real-time control of non-holonomic vehicles.To facilitate our implementations and speed up our algorithms, in particular to minimize the time required for computing gradient information, we introduce automatic differentiation and code generation for rigid body systems.Our customized solvers combined with derivative code generation and state-of-the-art software engineering form a framework which allows us to perform nonlinear model predictive control for longer time horizons or at higher update rates than other approaches.