Fault-Tolerant Control for Unseen Faults using Randomized Methods

Real-time tolerance to novel faults is often needed to avoid catastrophic outcomes. However, no real-time solutions currently exist for novel faults: e.g., isolating the underlying fault and implementing an appropriate control may be too time-consuming or require a priori fault information. For such circumstances, we propose a novel approach that uses fast randomized control switching to stabilize a system, without any state estimation. We first show theoretically that, for a properly tuned controller, a randomized approach guarantees mean-square stability and is guaranteed to converge to an optimal solution. We then empirically demonstrate the efficacy of this approach on trajectory following in a quadcopter UAV.