| Robotic vehicles are increasingly popular in the world, from the workplace to the home to the roadways. One of the underlying assumptions humans make when interacting with a robot is that of perfection: if a robot does something once, it will be able to replicate those movements exactly. While this is typically true, robots can behave unexpectedly when something breaks or faults. If the sources of faults can be detected directly, their effects can typically be minimized quickly. However, an unmeasurable or unanticipated fault can cause a system to become highly unstable.;The field of fault-tolerant control (FTC) has explored many ways to counteract these faults. Two typical approaches are to use a priori representations of the faults or to calculate the best response on-line. The a priori approach considered in this work, the multiple model (MM) approach, integrates the known dynamics of the system with the known failure modes to determine the expected model of the system under a set of given failures. By applying optimal control theory, this method can generate very efficient controllers, but only when the failures fall within the given set. If the faults fall outside of the given set, this method loses all guarantees of safe operation. On-line calculations can adapt to new faults much more easily, but they lose valuable information about how the system behaves.;Several approaches to merging these two methods have been suggested, but this work proposes a novel approach. By applying a genetic algorithm to the known model sets, new models can be generated, increasing the range of faults that can be accounted for without losing information on the system. This work will also demonstrate the effectiveness of this method by applying it to an autonomous vehicle and simulating actuator faults that are inside and outside the known fault set. This work will also show that a broad range of unknown faults can be accounted for within the vehicle without sensing the direct impact or source of the faults. |
