
Adinfinitium Lander Pilot
Status: Final Β· License: Apache-2.0 Β· Domain: lunar south pole (NASA LRO LOLA 5 m)
A 14.8 M-parameter transformer that learns a residual on top of classic ZEM/ZEV guidance for lunar terminal descent. With the residual at zero it flies exactly like classic guidance, so it can only help on top of a known-good baseline.
Model details
| Task | Terminal-descent control (continuous thrust commands) |
| Parameters | 14,842,567 (counted from the released checkpoint) |
| Architecture | PilotFormer: conv stem cuts a 32Γ32 hazard-camera crop into 64 patch tokens + 24-number state token β 8-layer transformer β policy and residual heads. d 384, layers 8, heads 8, crop 32, stem 48 |
| Control law | action = clip(classic(obs) + 0.3 Γ clamp(r), β1, 1) |
| Training | PPO; the released checkpoint was saved at update 40 (170 episodes); best recorded validation success 96.9% |
| Checkpoint | lander_pilot.pt Β· 59,409,547 bytes Β· id d53a162df9 Β· sha256 2501e02bβ¦ed39c |
| Wiring | Separate package; its camera sees the slope-rule hazard map, not the hazard U-Net |
Benchmarks
Simplified 3-DOF terminal-descent simulation on NASA LRO LOLA 5 m terrain. Paired fixed-seed missions (pilot and classic fly identical missions). Touchdown limits: vertical β€ 2 m/s, horizontal β€ 0.5 m/s, slope β€ 12Β°. "Harsh" = 3Γ failure rates.
| Test | Classic guidance | Pilot | Difference [95% CI] | Exact McNemar p |
|---|---|---|---|---|
| 3 unseen maps (Site42, Shoemaker, DM2), normal, n=2,304 | 97.66% | 98.39% | +0.74 pts [+0.16, +1.32] | 0.019 (Holm-adjusted 0.074) |
| 3 unseen maps, harsh (3Γ failure rates), n=2,304 | 96.31% | 96.92% | +0.61 pts [β0.23, +1.44] | 0.18 |
| Site23 final exam (flown once), normal, n=1,024 | 97.95% | 97.95% | 0.00 | 1.00 |
| Site23 final exam, harsh, n=1,024 | 96.48% | 96.29% | β0.20 pts | 0.88 |
| Mission rehearsal at planned landing cell (Site23, 2027-03-17), normal, n=512 | 99.02% | 99.61% | +0.59 pts [β0.4, +1.6] | 0.45 |
| Mission rehearsal, harsh, n=512 | 99.22% | 99.02% | β0.20 pts [β1.3, +1.0] | 1.00 |
Also measured (unseen maps): 9% less fuel (65.0 vs 71.5 kg), same miss distance (4.07 vs 4.10 m).
Limitations
- Worse than classic after an engine-out (95.9% vs 98.4%).
- The unseen-map gain comes from Site42 only; Shoemaker and DM2 are ties, and on the Site23 final exam the pilot is not better than classic.
- Most crashes are sideways touchdowns at 0.51β0.78 m/s against the 0.5 m/s limit; the residual is capped at 0.3 of full thrust.
- Only the first benchmark row is nominally significant, and it does not survive Holm correction (p = 0.074).
- Simplified 3-DOF simulation; not validated against flight dynamics.
Usage
# project code: https://github.com/oasisk/adinfinitium-handoff (adinfinitium_lander/worldmodel/lander/pilot.py)
from worldmodel.lander.pilot import load_checkpoint
model, info = load_checkpoint("lander_pilot.pt", device="cpu") # also sets model.rule from info["residual_rule"]
Dream-RSI: gated offline self-improvement
Dream-RSI lets a model keep improving after the main training ends, offline and safely:
- Find failures: fly or replay many missions and record where the current model fails.
- Dream harder: generate and replay those hard cases (failure-boosted scenarios) and retrain a candidate, warm-started from the current model.
- Gate: the candidate is scored on a frozen validation exam (fixed seeds, paired with the current model and the physics backup).
- Lander: accept only if the candidate loses no mission that the current model or the physics backup lands (zero regressions) AND improves total success.
- Planner and rover: accept only if the frozen validation metric improves by a margin.
- Repeat from whichever model won. Rejected candidates are discarded, and the shipped model never gets worse.
Test maps and the final-exam site are never used in Dream-RSI: training uses training maps, and gating uses a separate validation map.
Measured results (all logged rounds)
| Loop | Rounds accepted | Effect on its gate metric |
|---|---|---|
| Planner (PlannerSetNet) | 1 of 5 | Validation Brier 0.01543 β 0.01353 (β12.3%) |
| Lander pilot (Site11 gate, 1,024 paired missions) | 0 of 7 | Pilot unchanged; every candidate lost 12β25 missions, so the gate held |
| Rover brain v3.1 (self-learning on stressed worlds) | 9 of 36 | Blocked route improved at 4/4 sites, late light 3/4, dusty panels 1/4, early light 0/4 |
| Rover brain v3 | 9 of 72 | Blocked route improved at 6/6 sites, dusty panels 2/6 |
| Rover learned driver v2 | 5 of 38 | Returned rate stayed 0.0% |
Across all loops, 18 of 42 first rounds on a new problem were accepted, against 6 of 120 later rounds. What this shows: Dream-RSI reliably repairs specific failures (blocked routes, late light, model faults) without ever shipping a regression. In unchanged conditions it rarely finds a safe gain (0 of 32 rover rounds accepted).
Projections after more Dream-RSI (Projected, not measured)
Median after 20 more generations, from the planner's own Dream-RSI log:
- Planner Brier: 0.0145 β 0.0094 [0.0043, 0.0145]
Scenario only, for models where Dream-RSI has not been run yet (it borrows the improvement rate seen in the planner and rover loops):
- Hazard fresh F1: 0.757 β 0.779
- Light error: 0.450Β° β 0.410Β°
- Terrain error: 2.01 β 1.84 m
The lander and rover v3.1 are projected to stay unchanged, because their gates have not accepted anything in unchanged conditions.
Limitations and next steps
- The zero-regression gate is strict: it also rejected lander candidates that were slightly better overall (best: 991 vs 990 landings).
- Training rounds were short (4β5 updates) because of the compute budget.
- Next: harder self-generated scenarios, longer rounds, and a fresh unseen-map exam after each accepted generation.
Intended use and scope
Research, education and benchmarking of lunar landing-site navigation methods. Not for flight, safety-critical or operational landing decisions. All results are simulation or held-out-lunar-site results, not real-mission performance. Models were frozen (SHA-256) before held-out evaluation; each test site was evaluated once.
About Adinfinitium
Adinfinitium is an open-source effort to make lunar landing and surface navigation easier to study. It began as our NASA Space Apps 2026 project, Lunar Landing Site Navigator, and trains on NASA LRO LOLA 5 m terrain of the lunar south pole. The model chain is: hazard network β light network β terrain model β rover brain β lander pilot. See the whole family at https://huggingface.co/Imok282.
Source code, full records and the hand-off package: https://github.com/oasisk/adinfinitium-handoff (checkpoint SHA-256 values: CHECKPOINTS.sha256).
Citation
@misc{adinfinitium2026lunar,
title = {Adinfinitium open models for lunar landing-site navigation},
author = {Adinfinitium},
year = {2026},
howpublished = {Hugging Face, https://huggingface.co/Imok282},
note = {NASA Space Apps 2026, Lunar Landing Site Navigator}
}

