ANGS: A Synergistic Neuro-Mechanical Framework for Autonomous Legged Robots in Uncharted Extraterrestrial Landscapes

Planetary exploration, such as on the Moon and Mars, demands mobile robots capable of traversing rugged, low-gravity environments. Quadruped robots, particularly quadrupeds, offer superior adaptability for such missions. This paper presents Adaptive Neuro-Mechanical Gait Synthesis (ANGS), a novel hierarchical control framework integrating Deep Reinforcement Learning (DRL), Terrain Mechanics Predictor (TMP), and Model Predictive Control (MPC) to enable robust locomotion in challenging extraterrestrial terrains. ANGS overcomes limitations of traditional methods reliant on pre-defined gait libraries or purely reactive control by dynamically adapting gait parameters (stride length, foot trajectory, body posture) to real-time terrain mechanics (e.g., soil stiffness, slip dynamics). TMP employs deep neural networks to predict terrain properties from LiDAR and proprioceptive data, while DRL optimizes high-level motion policies, and MPC refines these policies for stability and energy efficiency. Simulation and real-world experiment results show that our proposed method enables the robot to move stably and efficiently on various surfaces, including low-gravity surfaces. The proposed approach overcomes the limitations of traditional movement strategies in unstructured environments. Combining terrain-aware adaptation and real-time trajectory optimization. TMP uses deep neural networks to predict terrain properties (e.g., soil stiffness, slip dynamics) from visual and tactile inputs, to dynamically adapt gait parameters (step length, foot trajectory, and body posture) to unknown terrain in real time. This framework advances the autonomy of legged robots for future unmanned missions in extreme extraterrestrial landscapes.