Transformer-Based Atmospheric Rocket Landing Guidance Using Meta-Reinforcement Learning

Abstract

Rocket reusability is a key factor in enabling quicker and more cost-effective access to space. However, landing on Earth poses significant challenges due to the dynamic and highly uncertain environment. A robust guidance, navigation, and control system is essential to guide the vehicle to the landing site while fulfilling terminal constraints and minimizing fuel consumption. This research integrates meta-reinforcement learning with Gated Transformer XL neural networks to enhance the robustness of the powered guidance with respect to atmospheric and aerodynamic uncertainties, navigation and control errors, and dispersed initial conditions. By employing a 6-degree-of-freedom dynamics model and accurate vehicle and environmental simulations, the agent learns a higher-fidelity guidance policy compared to existing literature, demonstrating successful and robust performance in perturbed Monte Carlo campaigns. While all other terminal constraints are met using only the RL policy, a terminal patch can be included to reduce the violation of the vertical velocity constraint, achieving up to 99.7% of the simulations compliant with all terminal landing requirements.