Neural-based trajectory shaping approach for terminal planetary pinpoint guidance

In this paper, we present an approach to pinpoint landing based on what we consider to be the next evolution of path shaping methodologies based on potential functions. Here, we employ Extreme Learning Machine (ELM) theories to devise a Single Layer Forward Network (SLFN) that learns the relationship between current spacecraft position and the optimal velocity field required to shape the path to the surface in a fuel efficient fashion. ELM techniques enable fast and accurate training as well as better generalization. The network is trained using open-loop, fuel-efficient trajectories that are numerically generated using pseudo-spectral methods. After test and validation, the SLFN becomes a critical element in the linear guidance algorithm loop. More specifically, a Linear Quadratic Regulator (LQR) is employed to track the optimal velocity field which is naturally defined to be attractive to the landing target. The guidance approach is tested on a simulation environment to evaluate the performance of proposed algorithm. Monte Carlo simulations show that the algorithm achieve a low guidance residual error which is less than one meter in position and less than -0.9 m/sec in impact velocity.