Deep Reinforcement Learning-Based Accurate Control of Planetary Soft Landing

Area of Science:

• Aerospace Engineering
• Robotics
• Artificial Intelligence

Background:

• Planetary soft landing is critical for space exploration missions.
• Existing control algorithms face challenges in convergence and efficiency.
• Deep Reinforcement Learning (DRL) offers potential for autonomous control.

Purpose of the Study:

• To propose a novel soft landing control algorithm using DRL.
• To address the sparse reward problem in reinforcement learning for landing.
• To ensure fuel efficiency and adherence to attitude constraints during descent.

Main Methods:

• Formulated the soft landing problem for powered descent.
• Designed a reward function incorporating velocity tracking, fuel consumption, and constraint violation penalties.
• Trained policies using Deep Deterministic Policy Gradient (DDPG), Twin Delayed DDPG (TD3), and Soft Actor Critic (SAC) frameworks.

Main Results:

• All tested DRL frameworks (DDPG, TD3, SAC) demonstrated convergence.
• The trained DRL policy successfully achieved velocity tracking goals.
• Experimental deployment validated the algorithm's effectiveness in soft landing and velocity control.

Conclusions:

• The proposed DRL-based algorithm provides a robust solution for planetary soft landing.
• The designed reward function effectively guides the lander towards safe and fuel-efficient touchdowns.
• The approach shows significant promise for future autonomous space missions.