Study on Self Learning Systems for Autonomous Soft Landing of Spacecrafts on Lunar Surfaces

With the rising interest in space exploration there is a rapid development of spacecraft technology. One of the biggest issues faced in space exploration is the soft landing of the exploration devices, due to its complex and stochastic nature. This study aims to propose a solution which solves the problem with the help of a self-learning, single-agent system powered by Reinforcement Learning. The Reinforcement Learning agent is made to act out in a simulation comprising of a realistic 2D model of the lunar surface and the lunar lander. A comprehensive comparison is conducted between the Proximal Policy Optimization, Deep Q Network, Advantage actor-critic algorithms to verify the best performing algorithm. The problem is approached as a combination of supervised and unsupervised learning techniques. The unsupervised learning component is used to alter the lander's path based on its own experience whereas the supervised learning component directs the lander to follow a planned course toward the lunar surface. This paves way for the automation of soft moon landing while minimizing human intervention, greatly increasing the probability of success of the mission. After thorough experimentation it is found that Proximal Policy Optimization performs the best with the highest mean rewards and lowest standard deviation rewards respectively.