Neural network based feedback optimal control for pinpoint landers under disturbances

Imitation learning leverages example demonstrations to teach a policy to replicate a desired behavior. In this investigation, the example demonstrations consist of open-loop optimal trajectories calculated off-line. This paper introduces the use of imitation learning to demonstrate optimal feedback control of two different high-fidelity 6-degree-of-freedom (6DOF) lander models. A loss of optimality is shown when disturbances are applied to policies trained only with nominal trajectories. Methodologies such as multi-phase optimal control and triple-single-phase optimal control are applied to include disturbances in the optimal trajectory generation, and the policies are trained to mitigate loss of optimality when disturbances are applied. Monte Carlo simulations show that loss of optimality can be mitigated by including disturbances in the optimal trajectory generation and training data.