Damage recovery for robot controllers and simulators evolved using bootstrapped neuro-simulation

Robots are becoming increasingly complex. This has made manually designing the software responsible for controlling these robots (controllers) challenging, leading to the creation of the field of evolutionary robotics (ER). The ER approach aims to automatically evolve robot controllers and morphologies by utilising concepts from biological evolution. ER techniques use evolutionary algorithms (EA) to evolve populations of controllers - a process that requires the evaluation of a large number of controllers. Performing these evaluations on a real-world robot is both infeasibly time-consuming and poses the risk of damage to the robot. Simulators present a solution to the issue by allowing the evaluation of controllers to take place on a virtual robot. Traditional methods of controller evolution in simulation encounter two major issues. Firstly, physics simulators are complex to create and are often very computationally expensive. Secondly, the reality gap is encountered when controllers are evolved in simulators that are unable to simulate the real world well enough due to implications or small inaccuracies in the simulation, which together cause controllers in the simulation to be unable to transfer effectively to reality. Bootstrapped Neuro-Simulation (BNS) is an ER algorithm that aims to address the issues inherent with the use of simulators. The algorithm concurrently creates a simulator and evolves a population of controllers. The process starts with an initially random population of controllers and an untrained simulator neural network (SNN), a type of robot simulator which utilises artificial neural networks (ANNs) to simulate a robot's behaviour. Controllers are then continually selected for evaluation in the real world, and the data from these real-world evaluations is used to train the controller-evaluation SNN. BNS is a relatively new algorithm that has not yet been explored in depth. An investigation was, therefore, conducted into BNS's ability to evolve closed-loop controllers. BNS was successful in evolving such controllers, and various adaptations to the algorithm were investigated for their ability to improve the evolution of closed-loop controllers. In addition, the factors which had the greatest impact on BNS's effectiveness were reported upon. Damage recovery is an area that has been the focus of a great deal of research. This is because the progression of the field of robotics means that robots no longer operate only in the safe environments that they once did. Robots are now put to use in areas as inaccessible as the surface of Mars, where repairs by a human are impossible. Various methods of damage recovery have previously been proposed and evaluated, but none focused on BNS as a method of damage recovery. In this research, it was hypothesised that BNS's constantly learning nature would allow it to recover from damage, as it would continue to use new information about the state of the real robot to evolve new controllers capable of functioning in the damaged robot. BNS was found to possess the hypothesised damage recovery ability. The algorithm's evaluation was carried out through the evolution of controllers for simple navigation and light-following tasks for a wheeled robot, as well as a locomotion task for a complex legged robot. Various adaptations to the algorithm were then evaluated through extensive parameter investigations in simulation, showing varying levels of effectiveness. These results were further confirmed through evaluation of the adaptations and effective parameter values in real-world evaluations on a real robot. Both a simple and more complex robot morphology were investigated