The pursuit of robotic autonomy is a long-standing ambition motivated by the need for robots to operate independently in dynamic and complex environments, such as space exploration and deep-sea missions. This thesis addresses the simulation-to-reality problem, where the transfer of robotic behaviors optimized in simulation is hindered by the reality gap—a dis- crepancy between simulated and real-world dynamics.
To bridge this gap, this work presents a holistic approach considering both the optimization of behaviors and the simulation. We develop a framework for comparing different simulation-to-reality approaches based on a formal definition of the problem. For behavior optimization, we employ a multi-objective evolutionary optimizer
(NSGA-II) for generating robotic behaviors with respect to the goal of solving a task and exploring the environment. A heuristic for selecting behaviors from the Pareto-optimal solutions provided by NSGA-II is proposed, referred to as ’constructive disagreement’, that prioritizes task-solving whenever a candidate for solving a given task in reality can be found. Furthermore, the thesis explores the application of the Niching Migratory Multi-Swarm Optimizer (NMMSO) for simulation optimization, which facilitates the generation and maintenance of a di- verse set of simulation parameterizations.
Experimental evaluations in a simulated environment demonstrate the potential of a multi-objective behavior optimization when task-solving and exploration are conflicting goals, compared to a single-objective approach with CMA-ES. Moreover, we conclude that a multi-modal simulation optimization could be beneficial for finding a diverse set of solutions when real-world samples are scarce and the ambiguity of suitable simulation parameters is high. We hope that the results presented in this thesis can be used as a foundation for future work on multi-objective behavior optimization and multi-modal simulation optimization in the context of simulation-to-reality problems.