Robust Autonomous Lunar Landing Under Uncertain Physics

Abstract

The project studies whether training an agent across randomized environment physics can improve its performance on previously unseen conditions compared with training under fixed physics. It uses the LunarLander-v3 environment and Proximal Policy Optimization (PPO), alongside a custom PyTorch actor–critic implementation and controlled evaluation experiments. The research focuses on two questions: Does training across randomized physical conditions improve generalization to previously unseen environment conditions compared with fixed-condition training? How does robustness vary across different training configurations and evaluation scenarios? The repository contains experimental implementations, training and evaluation scripts, model artifacts where included, and supporting results intended to make the research process easier to inspect and reproduce. The experimental workflow includes comparisons between agents trained under fixed and randomized physics, evaluation across multiple physical scenarios and random seeds, and investigation of the limitations of different training approaches. A standalone actor–critic implementation is also included as an implementation experiment. This release is intended to support transparent experimentation, reproducibility, and further research into robust reinforcement learning for autonomous landing tasks. Repository: https://github.com/empowered-aerial/LunarWing