Underactuated MIMO Airship Control Based on Online Data-Driven Reinforcement Learning

In this work, a novel online model-free controller for an underactuated dirigible is developed based on reinforcement learning and optimal control theory. A reinforcement learning structure is used while overcoming the dependence of the value function on future values by introducing a neural network that is adapted using input-output data. The suboptimal critic neural network is structured such that optimality is guaranteed over the interval from which the data is valid. The system performance is validated using a highly realistic physics engine, Gazebo, with the robot operating system (ROS) interface and the results are compared to the performance of a model-based controller specifically designed to control the airship model. It is emphasized that the proposed formulation does not leverage any knowledge of vehicle dynamics and thus is considered a vehicle agnostic control strategy.