---
title: Adinfinitium
emoji: 🌙
colorFrom: gray
colorTo: gray
sdk: static
pinned: false
---

# Adinfinitium
**Open models for lunar landing-site navigation**, built from NASA LRO LOLA 5 m terrain of the lunar south pole.
Adinfinitium began as our NASA Space Apps 2026 project, *Lunar Landing Site Navigator*. Models were frozen before testing,
held-out sites were evaluated once, and failures are reported alongside successes.
| Model | What it does | Params | Status |
|---|---|---|---|
| [lunar-hazard-unet](https://huggingface.co/Imok282/lunar-hazard-unet) | per-pixel slope hazard map | 60.4 M | final |
| [lunar-light-skyline](https://huggingface.co/Imok282/lunar-light-skyline) | skyline / illumination prediction | 38.3 M | final |
| [lunar-terrain-diffusion](https://huggingface.co/Imok282/lunar-terrain-diffusion) | terrain gap-filling | 49.9 M | experimental |
| [lunar-rover-brain](https://huggingface.co/Imok282/lunar-rover-brain) | scripted rover planner v3 / v3.1 | 0 learned | final |
| [lunar-lander-pilot](https://huggingface.co/Imok282/lunar-lander-pilot) | residual-on-classic landing guidance | 14.8 M | final |
| [lunar-rover-driver-v1](https://huggingface.co/Imok282/lunar-rover-driver-v1) | PPO rover policy | 11.4 M | **deprecated** (research record) |
| [lunar-rover-driver-v2](https://huggingface.co/Imok282/lunar-rover-driver-v2) | imitation rover policy | 11.4 M | **deprecated** (research record) |
## Headline results (held-out, simulation)
| Model | Metric | Ours | Baseline |
|---|---|---|---|
| Hazard U-Net | F1, held-out Site23 + Site42 | **0.951** | 0.817 (slope rule) |
| Hazard U-Net | F1, six fresh sites | **0.757** | 0.292 |
| Light transformer | skyline MAE, Site23 + Site42 | **0.450°** | 0.570° |
| Lander pilot | landing success, 3 unseen maps | **98.39%** | 97.66% (classic) |
| Rover brain v3.1 | returned, four untouched sites | **33.7%** | 17.9% (naive) |
| Terrain diffusion | RMSE, held-out | 2.01 m | **0.97 m** (kriging wins) |
### Landing-cell planner (not yet released)
PlannerSetNet (646,401 params) and a 14-parameter logistic planner pick the landing cell and hour (held-out Site23, 7 dates × 96 candidates):
| Method | Chosen-candidate success | Regret |
|---|---|---|
| Logistic planner | **15.2%** | 27.2 pts |
| Navigator ranking (shipped) | 12.8% | 29.6 pts |
| PlannerSetNet | 12.2% | 30.2 pts |
None of the differences is significant.
Each model card documents **Dream-RSI**, our gated offline self-improvement loop that never ships a regression, with measured results and clearly labelled projections.
Research and educational use only, not for flight or safety-critical decisions.
Source and hand-off package: https://github.com/oasisk/adinfinitium-handoff