| Summary: | Evolutionary Robotics (ER) is the field of study concerned with the application of evolutionary computation to the design of robotic systems. Two main issues have prevented ER from being applied to real-world tasks, namely scaling to complex tasks and the transfer of control to real-robot systems. Finding solutions to complex tasks is challenging for evolutionary approaches due to the bootstrap problem and deception. When the task goal is too difficult, the evolutionary process will drift in regions of the search space with equally low levels of performance and therefore fail to bootstrap. Furthermore, the search space tends to get rugged (deceptive) as task complexity increases, which can lead to premature convergence. Another prominent issue in ER is the reality gap. Behavioral control is typically evolved in simulation and then only transferred to the real robotic hardware when a good solution has been found. Since simulation is an abstraction of the real world, the accuracy of the robot model and its interactions with the environment is limited. As a result, control evolved in a simulator tends to display a lower performance in reality than in simulation. In this thesis, we present a hierarchical control synthesis approach that enables the use of ER techniques for complex tasks in real robotic hardware by mitigating the bootstrap problem, deception, and the reality gap. We recursively decompose a task into sub-tasks, and synthesize control for each sub-task. The individual behaviors are then composed hierarchically. The possibility of incrementally transferring control as the controller is composed allows transferability issues to be addressed locally in the controller hierarchy. Our approach features hybridity, allowing different control synthesis techniques to be combined. We demonstrate our approach in a series of tasks that go beyond the complexity of tasks where ER has been successfully applied. We further show that hierarchical control can be applied in single-robot systems and in multirobot systems. Given our long-term goal of enabling the application of ER techniques to real-world tasks, we systematically validate our approach in real robotic hardware. For one of the demonstrations in this thesis, we have designed and built a swarm robotic platform, and we show the first successful transfer of evolved and hierarchical control to a swarm of robots outside of controlled laboratory conditions.
|
|---|