Six-degree-of-freedom Optimal Feedback Control of Pinpoint Landing using Deep Neural Networks

Machine learning regression techniques have shown success at feedback control to perform near-optimal pinpoint landings for low fidelity formulations (e.g. 3 degree-of-freedom). Trajectories from these low-fidelity landing formulations have been used in imitation learning techniques to train deep neural network policies to replicate these optimal landings in closed loop. This study details the development of a near-optimal, neural network feedback controller for a 6 degree-of-freedom pinpoint landing system. To model disturbances, the problem is cast as either a multi-phase optimal control problem or a triple single-phase optimal control problem to generate examples of optimal control through the presence of disturbances. By including these disturbed examples and leveraging imitation learning techniques, the loss of optimality is reduced for pinpoint landing scenario.