Autonomous Hazard Detection and Avoidance for ESA's Argonaut Moon Lander : Boulder, Slope, and Shadow Detection with Multi Objective Landing Site Optimisation on Synthetic LiDAR and Camera Data

Sammanfattning

Safe, autonomous lunar landing requires real-time Hazard Detection and Avoidance (HDA). This thesis develops an integrated HDA and landing-site-selection simulation framework covering the final descent phase for the European Space Agency’s Argonaut Moon lander. Three hazard classes are addressed: terrain slopes exceeding 15°, boulders larger than 30 cm in diameter, and shadowed regions, each extended by a safety buffer. To begin addressing the fragmentation of the Synthetic Data Generation (SDG) toolchain landscape for HDA, this thesis includes a wide survey of candidate SDG tools and the first known comparison of lunar images rendered across separate SDG tools under matched sensor and terrain conditions. Within the present pipeline, synthetic sensor data are produced by rendering procedurally generated terrain through the Planet and Asteroid Natural scene Generation Utility (PANGU), populated with boulders and craters drawn from lunar-typical size-frequency distributions. Boulder detection uses a U-Net trained on synthetic imagery generated within the same pipeline. The HDA system operates in two sequential phases, HDA1 and HDA2. The first occurs at 700 m altitude, where slopes and shadows are identified from LiDAR-derived Digital Elevation Models and camera imagery, with up to 100 m of divert budget available. During the second phase, at 200 m altitude, boulders become resolvable and are detected alongside slopes and shadows, with up to 30 m of divert budget remaining. Per-modality hazards are extended with a safety buffer and fused into a local hazard map, on which a two-stage Multi-Objective Optimisation (MOO) algorithm ranks candidate landing sites and selects a touchdown target at each descent stage, weighing site safety against manoeuvre cost within the available divert budget. The MOO is integrated end-to-end with the HDA chain, closing the loop from raw sensor input to a concrete touchdown target. Evaluation on a 600-frame synthetic descent corpus produces a hazard-free touchdown target on every one of the top-50 pre-screened candidates, with all HDA2 diverts completing within the 30 m budget (maximum 22.6 m), and yields two important findings. First, the as-written 10 m ESA exclusion buffer is unflyable on this dataset: applied uniformly across all three hazard modalities, 12 of 50 candidates exhaust the HDA2 divert budget without reaching a safe landing site, against zero failures under a relaxed 5 m buffer. Second, the MOO selects touchdown targets in substantially safer immediate surroundings than a naive nearest-safe-pixel baseline: across the diverted candidates, the mean safe fraction within 5 m of touchdown is 87% at the MOO target versus 27% at the naive target, a 60 percentage-point gain.