Towards Autonomous Lunar Exploration: A Learning-based Lunar Rover Mobility System

Abstract

Autonomous navigation of planetary rovers on deformable, heterogeneous terrains presents challenges such as high-slip hazards, mobility limitations, and unpredictable terrain dynamics. This dissertation addresses three core problems: (1) inefficient wheel-legged locomotion systems that reduce actuators while maintaining adaptability, (2) the lack of predictive models for accurate slip estimation on heterogeneous terrains without explicit terrain classification, and (3) the inefficiency of current navigation methods in detecting, avoiding, or recovering from slippage-induced entrapment. To address these issues, this research develops a unified framework to improve rover mobility, terrain assessment, and adaptive navigation. First, a novel wheel-legged system driven by a single motor with bevel and worm gears and electromagnetic clutches is introduced. This system enables self-locking for enhanced resilience and fault tolerance in case of clutch or motor failure. The concept was implemented in the Khalifa University Space Rover (KUSR), which demonstrated superior obstacle and slope traversal, achieving high success rates in both high-fidelity simulations and real-world tests. Second, a vision-based slip prediction model, SlipNet, is developed, which dynamically estimates slip across segmented terrain without prior classification. SlipNet uses DeepLab v3+ for segmentation and combines probabilistic reasoning with a Vision Transformer (ViT) to predict wheel slip on diverse terrains. Simulations show that SlipNet outperforms state-of-the-art models like TerrainNet in accuracy across deformable soils. Third, the DRL-SAPER framework (Deep Reinforcement Learning for Slip Avoidance in Planetary Exploration Rovers) is developed for non-geometric hazard avoidance. DRL-SAPER uses a weighted slip cost map in its reward function, explicitly penalizing traversal through high-slip zones. It outperforms traditional planners in navigation success rate, slip reduction, and path efficiency. Finally, a hierarchical RL with teacher-student policies is proposed to guide rovers out of unavoidable high-slip zones using push-pull locomotion, enabling entrapment recovery. High-fidelity simulations and real-world validations confirm the robustness of these solutions. The key contributions of this dissertation are: (1) the design and implementation of a single-motor wheel-legged mobility system with self-locking and fault tolerance, validated through simulations and real tests; (2) the development of SlipNet, a learning-based slip prediction model that generalizes across heterogeneous, deformable terrains without prior classification; (3) the formulation of DRL-SAPER, a reinforcement learning framework with slip-aware reward for safe terrain navigation; and (4) a hierarchical RL strategy enabling recovery from entrapment through push-pull locomotion. Together, these contributions enhance the mobility, perception, and autonomy of planetary rovers in challenging terrains.

Date of Award	20 May 2025
Original language	American English
Supervisor	Yahya Zweiri (Supervisor)