Reinforcement learning-based powered descent and landing for planetary exploration

Abstract

This study introduces a reinforcement learning-based powered descent and landing technique for autonomous planetary landing. The main problem is decomposed into landing site selection and landing guidance subproblems to reduce complexity and ensure real-time capability. The landing site selection problem uses image-processing techniques to determine a safe location for landing based on altitude sensor data. The factors related to landing site selection are created as individual maps. An optimization problem formulation to maximize the weighted sum of these factors is developed and validated through a case study. A reinforcement learning technique is then applied to find an efficient and safe landing trajectory for the selected site. An episode reward (instead of a step reward) adopted in this study can ensure learning autonomy. A gradual reward system based on curriculum learning addresses the sparse reward. The proposed technique is compared with the solution obtained by offline optimization to demonstrate its effectiveness in stable autonomous landing. The model fidelity was enhanced to account for the uncertainty in real-world landing situations. The reinforcement learning-based landing framework proposed in this study can provide a real-time autonomous planetary capability.