Loading the terrain…
Cargo logistics at a lunar south-pole base, on real terrain
A discrete-event simulator of how landed cargo gets from the pads to where it is needed, with learned landing and driving models on NASA terrain of the Connecting Ridge.
Terrain near the lunar south pole from NASA's Lunar Orbiter Laser Altimeter, lit by a sun that stays within 1.5° of the horizon and circles once per lunar day; long shadows cross the ridges and low ground the sun cannot reach stays black. Shadows are cast only by terrain inside this tile.
Three results
From version 3: one fit over every simulated year of the pad-standoff, landing-dispersion and plume keep-out experiments (62,640 years, 30 seeds), read at the typical fleet load (4.6 % mean utilisation). Each number is the change in the mean delay of urgent cargo, with its 95% interval and its source.
- Added delay of urgent cargo per kilometre of pad standoff 35.0% per kmModelled (Modelled: produced by the LunarHaul simulation. Source: results/table_E20.csv, pooled fit of ln(urgent mean delay), term standoff: 100(e^b − 1) with a 95% bootstrap interval over seeds; n = 62,640 simulated years, 30 seeds (experiment E20))95% interval 32.6 to 37.6 %, at the typical fleet load, every fleet and dispatch rule. The relation is concave: a log-delay slope of about 0.52 per km below 1 km and 0.10–0.19 per km beyond 2 km.
- A plume keep-out radius and a stand-down add no measurable delay at these loads -0.5% per kmModelled (Modelled: produced by the LunarHaul simulation. Source: results/table_E20.csv, pooled fit of ln(urgent mean delay), term keepout: 100(e^b − 1) with a 95% bootstrap interval over seeds; n = 62,640 simulated years, 30 seeds (experiment E20))95% interval −1.5 to 0.8 %; stand-down -0.8% per hourModelled (Modelled: produced by the LunarHaul simulation. Source: results/table_E20.csv, pooled fit of ln(urgent mean delay), term standdown: 100(e^b − 1) with a 95% bootstrap interval over seeds; n = 62,640 simulated years, 30 seeds (experiment E20)), interval −1.1 to −0.3 %. The fleets, at 2–8 % utilisation, absorb the evacuations; a cost appears only in the lightest-loaded night cells at 2–3 km radii.
- Added delay per e-fold of the mean landing miss 8.2%Modelled (Modelled: produced by the LunarHaul simulation. Source: results/table_E20.csv, pooled fit of ln(urgent mean delay), term log_miss: 100(e^b − 1) with a 95% bootstrap interval over seeds; n = 62,640 simulated years, 30 seeds (experiment E20))95% interval 6.2 to 10.5 %. It acts through landings off the pad: a 100 m pad absorbs today's 50 m accuracy, a 25 m pad does not.
Read at build time from results/table_E20.csv. Every number carries its flag; hover a flag for its source (screen readers read it out).
Version 2's comparison of dispatch rules still holds: pooling the fleets of several owners cuts the delay of urgent cargo.
| Rule | Mean (h) | ±95% CI (h) |
|---|---|---|
| In-house fleets | 16.2 | 4.5 |
| Pooled greedy | 7.61 | 2.3 |
| Batched assignment | 7.46 | 2.1 |
| Marketplace (ours) | 5.95 | 1.7 |
| Unconstrained fleet (bound) | 2.25 | 0.19 |
results/table_main.csv.What it is
A simulator of surface cargo movement at a south-pole base on real terrain: lander manifests and landing outcomes, rover fleets owned by several organisations, the lunar night, Earth–Moon control latency, and four ways of dispatching the vehicles, from each owner using its own rovers to a pooled marketplace. The landing and driving outcomes it samples come from controllers flown and driven in their own physics models, and the same physics runs in Python and in your browser.
What is new
Distance is what costs. At the typical fleet load, each kilometre between pad and base adds 35.0 % to the delay of urgent cargo, most of it in the first kilometre. A plume keep-out radius and a stand-down after each landing add no measurable delay at these loads: the fleets absorb the evacuations. Landing dispersion adds 8.2 % per e-fold of the mean miss, only through landings that miss the pad: a 100 m pad absorbs today's 50 m accuracy, a 25 m pad does not.
A learned lander comes close when it sees what the guidance law sees. Imitation of the classical law, given the altimeter's recent history, lands 89.0 % of 500 dispersed landings in the touchdown box, against 93.8 % for the law itself; reinforcement learning on the same rewards trades touchdown accuracy for reward.
Rover autonomy stands in for fleet size. Two autonomous rovers deliver urgent cargo about as fast as six teleoperated ones.
Try it
Replay a simulated year on the terrain in 3D, fly the lander and drive the rover yourself, read every figure with its data, and run the checks that compare the browser with Python. Nothing needs an account.
Lander lab · Rover lab · Results · Verification · Simulator
Credits
Terrain: LOLA 5 m digital elevation model, PGDA product 90 (Barker et al. 2021), NASA Goddard Space Flight Center; averaged here to 16–20 m posts for the web. Polar lighting geometry after NASA Scientific Visualization Studio item 4893. Vehicles and landers in the 3D views are parametric, representative models. Typeface: Source Sans 3 (Adobe, SIL Open Font License). Icons: Lucide (ISC). 3D: three.js (MIT).