Learning-Based End-to-End Path Planning for Lunar Rovers with Safety Constraints

Area of Science:

• Robotics
• Artificial Intelligence
• Planetary Science

Background:

• Autonomous exploration of the Moon requires robust path planning for lunar rovers.
• Existing algorithms may not adequately address safety constraints or real-world lunar terrain complexities.

Purpose of the Study:

• To propose a learning-based, end-to-end path planning algorithm for lunar rovers incorporating safety constraints.
• To enhance the safety and efficiency of autonomous lunar exploration missions.

Main Methods:

• Developed a Gazebo simulation environment with real lunar terrain data and a lunar rover simulator.
• Designed a deep reinforcement learning (DRL) algorithm, including state/action spaces, network architecture, and a reward function considering slip behavior.
• Employed proximal policy optimization (PPO) for training and curriculum learning to improve generalization across diverse lunar terrains and scales.

Main Results:

• The proposed DRL algorithm successfully achieved end-to-end path planning for lunar rovers in simulations.
• Paths generated by the algorithm demonstrated a higher safety guarantee compared to classical path planning methods.
• The use of curriculum learning improved the algorithm's generalization capabilities for varied lunar environments.

Conclusions:

• The learning-based end-to-end path planning algorithm offers a promising solution for safe and efficient lunar rover exploration.
• The integration of safety constraints and DRL, coupled with curriculum learning, significantly enhances navigation capabilities on extraterrestrial surfaces.