Deep reinforcement learning and randomized blending for control under novel disturbances

Enabling autonomous vehicles to maneuver in novel scenarios is a key unsolved problem. A well-known approach, Weighted Multiple Model Adaptive Control (WMMAC), uses a set of pre-tuned controllers and combines their control actions using a weight vector. Although WMMAC offers an improvement to traditional switched control in terms of smooth control oscillations, it depends on accurate fault isolation and cannot deal with unknown disturbances. A recent approach avoids state estimation by randomly assigning the controller weighting vector; however, this approach uses a uniform distribution for control-weight sampling, which is sub-optimal compared to state-estimation methods. In this article, we propose a framework that uses deep reinforcement learning (DRL) to learn weighted control distributions that optimize the performance of the randomized approach for both known and unknown disturbances. We show that RL-based randomized blending dominates pure randomized blending, a switched FDI-based architecture and pre-tuned controllers on a quadcopter trajectory optimisation task in which we penalise deviations in both position and attitude.