Safe Rocket Landing Using Lyapunov-Based Reinforcement Learning

Reusable rockets are significant for reducing space program costs, improving human safety, and expanding exploration opportunities. This study introduces a safe and robust control framework, termed Lyapunov Proximal Policy Optimization (LPO), designed for the vertical landing of first-stage rocket boosters at a predetermined target location. By integrating reinforcement learning (RL) techniques for achieving rapid and precise control under unknown system dynamics, the proposed method extends conventional approaches through the incorporation of Lyapunov stability theory to ensure safety during the training process. Experimental evaluations are performed using two RL agents, one incorporating Lyapunov stability constraints and one without them. The controllers are tested within a physics-based simulation environment, where the LPO agent demonstrates superior landing precision and stability compared to the baseline Proximal Policy Optimization (PPO) algorithm. The results indicate that the proposed method provides a promising foundation for extending Lyapunov-based reinforcement learning to broader stability, tracking, and control applications.