Reinforcement-Learning-Enhanced Model Predictive Control with Application to Autonomous Planetary Landing

This paper presents a reinforcement-learning (RL)-enhanced model predictive control (MPC) framework, referred to as RLE-MPC, for robust spacecraft guidance and control. The RL agent learns the parameters of a quadratic cost function, which is used to formulate MPC’s recursively solved optimization problem. By training in a perturbed environment, RL enhances robustness to uncertainties and learns the long-term effects of control actions, mitigating the limitations of nominal MPC under unmodeled dynamics. Conversely, MPC explicitly specifies and enforces the cost function and constraints, thus preserving within the control policy the constraint-awareness and optimality guarantees that are generally absent in standard RL methods. Numerical results for a soft pinpoint lunar landing scenario under uncertainty evaluate the performance and robustness of RLE-MPC in comparison to standalone MPC and RL, both in the perturbed environment used during training and in a more challenging one to assess generalization.