A Hybrid AI-assisted Hazard Detection and Avoidance Pipeline for Autonomous Lunar Landing

Mohamed El Awag, Riccardo Teodori, Ludovica Cavalieri, Antonio Genova

Keywords: Space Robotics and Automation, AI-Enabled Robotics, Motion and Path Planning

TL;DR: This work presents a AI-assisted framework that combines vision-based hazard detection, learned optimal guidance, and robust control to achieve safe and real-time autonomous lunar landing.

Abstract:

Safe and precise autonomous landing is a critical capability for future lunar exploration missions targeting scientifically valuable yet hazardous regions, such as the lunar South Pole. These environments require real-time perception and decision-making to identify hazards and replan trajectories under strict computational and time constraints. This work presents a hybrid AI-assisted Hazard Detection and Avoidance (HDA) pipeline for autonomous lunar landing, integrating learning-based perception and guidance with model-based control.
The proposed framework combines Convolutional Neural Networks for vision-based terrain analysis, a Multi-Layer Perceptron (MLP) for near-optimal trajectory replanning, and a Model Predictive Control (MPC) scheme for robust trajectory tracking. In the Hazard Detection phase, dense regression networks estimate terrain slope and roughness from monocular images, while uncertainty is quantified through Monte Carlo Dropout and conformal prediction to ensure reliable safety assessment. These outputs are used to generate probabilistic landing safety maps and select optimal landing sites.
For Hazard Avoidance, an MLP trained on a dataset of fuel-optimal trajectories approximates the solution of a mass-optimal control problem, enabling real-time trajectory generation. The resulting trajectory is tracked by an MPC controller that enforces system constraints and compensates for model uncertainties, ensuring soft landing conditions.
Preliminary results demonstrate good terrain characterization with robust uncertainty quantification, and near-optimal trajectory generation with real-time performance. The proposed pipeline represents a promising approach toward robust and computationally efficient autonomous landing in challenging planetary environments.

Submission Number: 23