LunarHaul · lunar surface logistics

Results

Every figure and table here is read from the repository's figures/ and results/ folders when the site is built, with its caption, sample size and provenance flag. A figure that the latest run could not make says why. The numbers are whatever the last run produced; nothing on this page is typed in by hand.

Phase 3: learned landing, real terrain, plume and rover autonomy

Figures as last produced by experiments/phase3/make_figures_phase3.py; captions, sample sizes and flags from figures/phase3_captions.json. They are regenerated after the final runs, under the same names.

Figure lander mc
Figure 20. Terminal-descent Monte Carlo per lander class: the classical ZEM/ZEV law against the learned policies. Top: horizontal miss at touchdown in plan view (symlog axes, linear within ±5 m), passed episodes as filled points and failed ones as crosses; dashed circle = the 50 m pad radius. Bottom: pass rate (dot) with its Wilson 95 % CI (whisker) over the stacked outcome mix. Episodes: CLPS classical law n = 500, CLPS learned: bc v3 n = 500, CLPS learned: bcml n = 500, CLPS learned: baseline n = 500, CLPS learned: bc n = 500, CLPS learned: hover bonus n = 500, CLPS learned: hover bonus v1 n = 300, CLPS learned: no eta n = 500, MK1 classical law n = 500, LARGE classical law n = 500; sensor noise on. Flags: environment Modelled (3-DoF descent parameterised from flown landers, Spec); failure limits Spec; learned policies trained in-house (Modelled).

n: CLPS classical law n = 500, CLPS learned: bc v3 n = 500, CLPS learned: bcml n = 500, CLPS learned: baseline n = 500, CLPS learned: bc n = 500, CLPS learned: hover bonus n = 500, CLPS learned: hover bonus v1 n = 300, CLPS learned: no eta n = 500, MK1 classical law n = 500, LARGE classical law n = 500 · Intervals: Wilson 95 % interval on the pass rate · Flags: Modelled / Spec

PNG · PDF (repository)

Figure driving validation
Figure 21. Driving-table validation. Panel A: speed made good on the synthetic flat 2 km through the table's own leg integrator (deterministic made-good with whole stops; the stochastic mean over 400 integrator runs differs by DEPOT, max slope 5.8°), shown for context. Panel B: traction energy of the BASE→PSR and PSR→BASE legs at the 4 s latency bin, day and night, per autonomy level, against the Chen (2026) 0.4-4.4 kWh band for a 500 kg rover scaled linearly to the table's 1600 kg rover (Assumed). n = deterministic integrator means, no replication (the table has no seeds). Flags: JSC points Measured; Chen band Modelled; mass scaling Assumed; table legs Modelled on LOLA 5 m terrain (Measured).

n: deterministic integrator means (no replication) · Intervals: none (point checks) · Flags: Measured anchors, Modelled table, Assumed mass scaling

PNG · PDF (repository)

Figure dispersion standoff
Figure 22. Landing dispersion acts through off-pad landings. Rows 1-2: paired difference in the mean urgent-item delay against the classical miss level (dCL) as a function of the implied mean |miss| (log axis; flown-lander and requirement anchors on top), one line per pad radius, pad standoff 1 km, small multiples by fleet (columns) and dispatch mechanism (rows). Bottom row: share of landings beyond the pad radius. n = 30 seeds per cell; differences are paired on common seeds and the whiskers are percentile seed-bootstrap 95 % CIs (1000 resamples, fit_law.paired); off-pad whiskers are 1.96 SE over seeds. Flags: miss levels Measured (SLIM, Blue Ghost M1, IM-1, IM-2) and Spec (ADD Rev C 50 m, NASA/HLS 100 m); the classical level Modelled (classical (as flown in the env)); terrain Measured (LOLA 5 m); logistics Modelled.

n: n = 30 seeds per cell · Intervals: paired seed-bootstrap 95 % percentile CI; 1.96 SE for shares · Flags: Measured anchors, Spec requirements, Modelled logistics

PNG · PDF (repository)

Figure standoff
Figure 23. Pad standoff from BASE. Top: cell means of urgent mean delay (h), fleet km per year, driving energy (kWh / yr) against standoff for each fleet (colour) and dispatch mechanism (line style), classical (as flown in the env), 50 m pad radius; whiskers are 95 % CIs (1.96 SE over seeds). Bottom: paired differences against the 1 km standoff (dD1) on common seeds, whiskers = percentile seed-bootstrap 95 % CIs (1000 resamples); the 1 km point is the reference (0). n = 30 seeds per cell. Flags: terrain Measured (LOLA 5 m, Connecting ridge); driving legs and landings Modelled; standoff set Assumed (swept).

n: n = 30 seeds per cell · Intervals: 95 % CI = 1.96 SE over seeds (means); paired seed-bootstrap 95 % percentile CI (differences) · Flags: Measured terrain, Modelled logistics, Assumed sweep

PNG · PDF (repository)

Figure plume
Figure 24. Plume keep-out. Heatmaps of the paired difference in the mean urgent-item delay against the matching no-plume cell (dNP, hours, printed in each cell) over keep-out radius × stand-down time, pad standoff 1 km, small multiples by fleet and night share (rows) and dispatch mechanism (columns); cells whose percentile seed-bootstrap 95 % CI (1000 resamples) excludes 0 are outlined. Colour scale clipped at the 98th percentile of |dNP| (1.30 h). Bottom row: vehicle-hours lost to keep-out holds and evacuations per year against keep-out radius, one line per stand-down time (P2, 4 vehicles, night 15 %; whiskers 1.96 SE). n = 30 seeds per cell (P2, 4 vehicles: n = 30 seeds; P2, 6 vehicles: n = 30 seeds; P3, 9 vehicles: n = 30 seeds). Flags: keep-out radius and stand-down Assumed (swept, not predicted; NASA calls PSI predictions qualitative); impacts-vs-distance ordering Modelled (Wittal et al.); logistics Modelled.

n: n = 30 seeds per cell · Intervals: paired seed-bootstrap 95 % CI (outlined cells exclude 0); 1.96 SE for stand-down hours · Flags: Assumed sweep, Modelled logistics

PNG · PDF (repository)

Figure autonomy fleet
Figure 25. Vehicle autonomy level and fleet size (P2 demand, every vehicle at the same level): urgent mean delay (h) and operator console hours / yr against fleet size for teleoperation at 4, 6 and 8 s round-trip latency, supervised autonomy and full autonomy (ordered colours), in-house fleets vs the marketplace (columns). Bands are 95 % CIs (±1.96 SE over seeds). n = 30 seeds per cell. Flags: speed-vs-latency and stop cadence Modelled on JSC teleoperation (Measured); intervention rates Assumed by autonomy level; operators per fleet = max(2, n/2) (Assumed); latencies Spec (Moon direct 4 s) and Assumed (6, 8 s relayed).

n: n = 30 seeds per cell · Intervals: 95 % CI bands (±1.96 SE over seeds) · Flags: Modelled on Measured JSC curve; Assumed interventions

PNG · PDF (repository)

Figure siting
Figure 26. Pad siting. Top: total cost per year (operator cost of hauling + priced driving energy + amortised pad construction over 10 yr + expected plume damage = landings/yr × impacts per landing × line-of-sight shield × cost per impact + delay valuation (delivered items × mean delay × value of time)) against pad standoff for the reference cell (P2-default, marketplace, classical (as flown in the env), 50 m pad, delay valued at 100 $/item-h; cell means, n = 30 seeds per cell), one line per damage cost (colour) and pad-cost case (line style), small multiples by energy price; the filled marker is the minimum-cost standoff. Bottom: share of all cells and pad-cost cases whose minimum lies at each standoff, by energy price and damage cost (rows; 504 groups per row), small multiples by value of time (0, 100, 1000 $/item-h). Flags: hauling Modelled; energy price Assumed; pad cost Spec (Metzger & Autry 2022), independent of standoff; impacts/landing Modelled from Wittal et al. (OTPS 2022) anchors, shield 0.1 Assumed when PAD_A LOS to BASE is blocked; damage cost per impact Assumed; delay valuation Assumed (blended $/item-hour stand-in for the engine's per-priority vot_usd_per_h). The pad cost does not depend on standoff and shifts totals, not minima.

n: n = 30 seeds per cell · Intervals: none shown (cell means; the siting terms are deterministic given the means) · Flags: hauling Modelled; energy price Assumed; pad cost Spec (Metzger & Autry 2022), independent of standoff; impacts/landing Modelled from Wittal et al. (OTPS 2022) anchors, shield 0.1 Assumed when PAD_A LOS to BASE is blocked; damage cost per impact Assumed; delay valuation Assumed (blended $/item-hour stand-in for the engine's per-priority vot_usd_per_h)

PNG · PDF (repository)

Figure law
Figure 27. The dispersion / standoff / keep-out law: OLS of ln(mean urgent-item delay) on ln(mean |miss|), pad standoff, keep-out radius, stand-down time, the reference-cell utilisation, night share, off-pad share, pad radius, their interactions and mechanism dummies (in-house is the base), pooled over the E17 and E18 years (n = 62,640 simulated years from 30 seeds; pooled R² = 0.48). Coefficients are shown as the % change in mean urgent delay per the stated unit (100·(exp(b·unit) − 1)); whiskers are percentile seed-bootstrap 95 % CIs (1000 resamples, seeds drawn with replacement, every year of a drawn seed weighted by its multiplicity); filled markers exclude 0, open ones do not. Panel B repeats the fit within each mechanism (crosses: term constant in that subset). Flags: the fit is Modelled on simulated years; regressor levels Measured / Spec (miss anchors) and Assumed (standoff, keep-out, stand-down sweeps). Suites with fewer seeds enter fewer bootstrap resamples; a CI is only as good as the seeds behind the term.

n: n = 62,640 simulated years from 30 seeds · Intervals: percentile seed-bootstrap 95 % CI (fit_law.fit_ols) · Flags: Modelled fit; Measured/Spec/Assumed levels

PNG · PDF (repository)

Lander Monte Carlo (E15)

Experiment E15: terminal-descent Monte Carlo per lander class and controller (11 rows). Pass = touchdown inside the velocity, tilt and sensor limits; Δ pass rate against the classical law of the same class with its 95 % interval.
Class · controllerEpisodesPass rateΔ pass rate vs classicalMiss p50Miss p90Off the 50 m padPropellantFailures (count)
CLPS · classical law500 93.8%Modelledreference2.2mModelled4.7mModelled 0.4%Modelled89kgModelledvelocity 14, sensor 6, timeout 11
CLPS · classical law500 94.8%Modelledreference2.2mModelled4.7mModelled 0.4%Modelled85kgModelledvelocity 11, tilt 1, sensor 11, timeout 3
CLPS · PPO, reward v2500 63.2%Modelled-30.6ppModelled ± 4.74.0mModelled7.8mModelled 1.2%Modelled61kgModelledvelocity 52, tilt 22, sensor 110
CLPS · imitation + DAgger, filtered teacher500 58.2%Modelled-35.6ppModelled ± 4.82.2mModelled5.1mModelled 0.4%Modelled86kgModelledvelocity 20, tilt 49, sensor 133, timeout 7
CLPS · imitation + DAgger, observation v3500 89.0%Modelled-4.8ppModelled ± 3.52.2mModelled4.8mModelled 0.4%Modelled81kgModelledvelocity 16, tilt 6, sensor 33
CLPS · imitation + DAgger, memoryless teacher500 81.2%Modelled-12.6ppModelled ± 4.02.2mModelled4.6mModelled 0.6%Modelled88kgModelledvelocity 14, tilt 8, sensor 63, timeout 9
CLPS · PPO, hover bonus (env v2)500 51.2%Modelled-42.6ppModelled ± 4.94.7mModelled10.3mModelled 1.2%Modelled63kgModelledvelocity 59, tilt 45, sensor 140
CLPS · PPO, hover bonus (env v1)300 6.3%Modelled-87.5ppModelled ± 3.53.5mModelled8.4mModelled 1.0%Modelled145kgModelledvelocity 15, tilt 13, sensor 25, timeout 228
CLPS · PPO, no survival term500 0.0%Modelled-93.8ppModelled ± 2.19.3mModelled27.4mModelled 3.6%Modelled50kgModelledvelocity 500
LARGE · classical law500 97.0%Modelledreference2.1mModelled4.4mModelled 0.0%Modelled2,297kgModelledvelocity 1, sensor 9, timeout 5
MK1 · classical law500 93.0%Modelledreference2.1mModelled4.4mModelled 0.0%Modelled298kgModelledsensor 19, timeout 16

Flags: environment Modelled (3-DoF descent parameterised from flown landers), failure limits Spec, learned policies trained in-house (Modelled); sensor noise on in every row. table_E15.csv · per-episode outcome tables: data/policies/outcomes_*.json (repository).

Driving-table validation (E16)

Speed made good against round-trip time, flat ground (calibration): synthetic flat 2 km, max slope 0.0°, lighting none
AutonomyRound tripSimulatedMean of runsRatio to 0 sReferenceErrorInterventions λStops
teleop0 s3.26km/hModelled3.25km/hModelled1.000 3.24km/hMeasured0.66 %0.1010
teleop4 s2.57km/hModelled2.56km/hModelled0.788 2.56km/hMeasured0.38 %0.1020
teleop6 s2.03km/hModelled2.02km/hModelled0.623 2.03km/hMeasured0.07 %0.1034
teleop8 s1.76km/hModelled1.76km/hModelled0.541 1.76km/hMeasured0.20 %0.1041
supervised0 s2.71km/hModelled2.70km/hModelled1.000 ––0.204
supervised4 s2.71km/hModelled2.70km/hModelled1.000 ––0.204
supervised6 s2.71km/hModelled2.70km/hModelled1.000 ––0.204
supervised8 s2.71km/hModelled2.70km/hModelled1.000 ––0.204
autonomous0 s8.94km/hModelled8.89km/hModelled1.000 ––0.400
autonomous4 s8.94km/hModelled8.89km/hModelled1.000 ––0.400
autonomous6 s8.94km/hModelled8.89km/hModelled1.000 ––0.400
autonomous8 s8.94km/hModelled8.89km/hModelled1.000 ––0.400

flat ground as calibrated (Litaker/JSC teleop, Measured); same integrator as driving-v1; sim_kmh is the deterministic made-good with whole stops, sim_kmh_mean_runs the mean over 400 runs

Speed made good against round-trip time on a real leg: CONSTR->DEPOT, 0.66 km, max slope 5.8°, lighting 270
AutonomyRound tripSimulatedMean of runsRatio to 0 sReferenceErrorInterventions λStops
autonomous0 s7.93km/hModelled–1.000 ––0.10–
autonomous4 s8.00km/hModelled–1.009 ––0.10–
autonomous6 s7.81km/hModelled–0.984 ––0.00–
autonomous8 s8.04km/hModelled–1.014 ––0.05–
supervised0 s2.47km/hModelled–1.000 ––0.10–
supervised4 s2.41km/hModelled–0.974 ––0.05–
supervised6 s2.51km/hModelled–1.018 ––0.15–
supervised8 s2.47km/hModelled–1.001 ––0.10–
teleop0 s3.01km/hModelled–1.000 3.24km/hMeasured-7.12 %0.05–
teleop4 s2.39km/hModelled–0.793 2.56km/hMeasured-6.82 %0.10–
teleop6 s1.88km/hModelled–0.626 2.03km/hMeasured-7.19 %0.00–
teleop8 s1.68km/hModelled–0.558 1.76km/hMeasured-4.67 %0.00–

context: flattest real leg >= 500 m (max slope 5.8 deg, so below the flat-ground reference); table time excludes interventions

Energy of a round trip into the permanently shadowed region: BASE->PSR + PSR->BASE (6 rows)
CaseSimulatedReference bandIn band
autonomous, 4.0 s, day1.77kWhModelled1.3–14.1 kWhyes
autonomous, 4.0 s, night1.89kWhModelled1.3–14.1 kWhyes
supervised, 4.0 s, day1.83kWhModelled1.3–14.1 kWhyes
supervised, 4.0 s, night1.86kWhModelled1.3–14.1 kWhyes
teleop, 4.0 s, day1.84kWhModelled1.3–14.1 kWhyes
teleop, 4.0 s, night1.77kWhModelled1.3–14.1 kWhyes

Chen 2026 band 0.4-4.4 kWh at 500 kg scaled x3.2 (rolling + grade losses scale with m*g: Assumed linear); table energy is traction only

Flags: simulated values Modelled (rover model and driving tables, lunarsim.rover); teleoperation references Measured. table_E16.csv · raw runs: E16_driving_validation.csv (repository).

Version 2 results

Show the 19 figures of the version-2 paper (marketplace dispatch on an abstract site graph)

Captions from the version-2 paper. The engine and these results are unchanged in Phase 3 (lunarsim compare reproduces them).

Version 2 figure: sitemap
Figure 1. The nine-site south-pole base used in every scenario. Edge labels give path lengths after the tortuosity factor; colours give slope class. The dashed circle marks the 1 km plume-ejecta standoff that keeps pads away from the base.

PNG · PDF (repository)

Version 2 figure: architecture
Figure 2. Testbed architecture. The demand and environment submodels feed an event engine that calls a dispatch mechanism whenever cargo becomes ready, a vehicle becomes free, a night ends or a landing is near (or at fixed ticks for batched mechanisms). The engine prices every returned job: empty approach, loading, slope- and night-adjusted travel, teleoperation slowdown, operator availability, random faults and a hibernation penalty if a vehicle without night capability is caught in the dark.

PNG · PDF (repository)

Version 2 figure: demand
Figure 3. Demand. A: landed payload mass at the south pole by year, assembled from public mission manifests (low, base and high cases; Appendix A lists each mission). The 2026–2028 base case is five or six discrete sub-3-tonne items, which is the single most important fact for a marketplace design. B: item-mass histograms from ten simulated years per phase. Vertical lines mark the LTV's 800 kg full-performance and 1,600 kg reduced-performance cargo limits and the 10-tonne hauler we introduce.

PNG · PDF (repository)

Version 2 figure: main
Figure 4. Main comparison with each phase's default fleet. Bars are means over 30 seeds; whiskers are 95% confidence intervals. A: 90th-percentile delay of urgent items from release to delivery. B: share of urgent items delivered within 24 h. C: share of landed cargo mass that no vehicle or team in the fleet could ever move. D: number of times per year a vehicle without night capability is caught out by the lunar night. The grey bars are a 16-vehicle fully autonomous fleet with no other duties or faults, included as a bound.

PNG · PDF (repository)

Version 2 figure: timeline
Figure 5. One simulated Phase-2 year (seed 3, four-vehicle fleet) under pooled greedy dispatch (top) and the marketplace (bottom). Each dot is a delivered item at its release day and its delay. Grey bands are lunar nights at the base, orange lines are landings. The marketplace's deferrable lots cluster at the 24 h consolidation limit; its urgent items (red) are served at the same speed as under greedy dispatch, and no vehicle is caught by the night. The same replay, animated on the base map, is on the Replay tab of the browser version.

PNG · PDF (repository)

Version 2 figure: fleet size
Figure 6. Fleet-size sweep at Phase-2 demand (30 seeds; bands are 95% CIs). Fleets grow from two LTVs to eight mixed vehicles (Table 4). Left: urgent 90th-percentile delay. Middle: share of urgent items delivered within 24 h with the 90% target marked. Right: system cost per kilogram of all mass moved; the curves lie on top of each other because the fleet, not the dispatch rule, sets the cost.

PNG · PDF (repository)

Version 2 figure: thickness
Figure 7. Market thickness. Six vehicles (five LTVs and one hauler) are split among 1, 2, 3, 4 or 6 owners; cargo is drawn from the same owners. Dashed lines with squares are in-house contracting, solid lines with circles the marketplace; colours give the share of time vehicles spend on their owners' other duties. Bands are 95% CIs over 30 seeds. The marketplace lines are flat because pooling makes ownership irrelevant to dispatch.

PNG · PDF (repository)

Version 2 figure: heavy cargo
Figure 8. Share of landed cargo mass that no vehicle or team could move, by fleet composition and phase (30 seeds; 95% CIs). Team lift (blue) recovers about half of the mass stranded by an LTV-only fleet and most of the mass stranded with a single hauler.

PNG · PDF (repository)

Version 2 figure: latency
Figure 9. Autonomy level and control latency at Phase-2 demand with a five-vehicle fleet (20 seeds; bands are 95% CIs). A: urgent 90th-percentile delay with two Earth operators, on log axes; the dotted lines mark direct and relayed lunar latency and the near and far Mars cases. B: share of the year's mass delivered. C: operator console hours per year at the 5 s default latency for one, two and four operators.

PNG · PDF (repository)

Version 2 figure: standards
Figure 10. Interface standards at Phase-2 demand (30 seeds; 95% CIs). Three of five vehicles accept only the common standard. A: urgent same-day service level. B: urgent 90th-percentile delay.

PNG · PDF (repository)

Version 2 figure: environment
Figure 11. Environment sweeps at Phase-2 demand with a five-vehicle fleet (30 seeds; bands are 95% CIs). A: dark share of the lunar cycle at the base. B: mean time between faults as a multiple of the default. C: bulk regolith and feedstock flows, shown as fleet utilisation.

PNG · PDF (repository)

Version 2 figure: ablations
Figure 12. Ablations with tight fleets (four vehicles in Phase 2, six in Phase 3; 30 seeds; 95% CIs). A: urgent 90th-percentile delay with each component removed. B: vehicle trips per year. C: urgent delay as a function of the batching window; zero means dispatch on every event.

PNG · PDF (repository)

Version 2 figure: tuned
Figure 13. Hand-set versus tuned auction weights (30 fresh seeds per phase; 95% CIs). Weights were searched on Phase 2 only. The tuned weights shorten the urgent tail in Phases 2 and 3 and reduce operator hours, at the cost of longer waits for deferrable bulk lots.

PNG · PDF (repository)

Version 2 figure: economics
Figure 14. Economics. A: tornado of system cost per kilogram moved for the Phase-2 four-vehicle marketplace cell; vehicle unit cost and life are analytic swings on the cost model, the others are simulated (30 seeds). B: cost of landing a kilogram on three vehicles [46, 47, 4] against the simulated cost of moving it on the surface, per landed kilogram and per kilogram of all mass moved including regolith.

PNG · PDF (repository)

Version 2 figure: capacity ladder
Figure 15. Capacity ladder. Share of landed cargo mass that can be moved by a vehicle or team whose largest payload is the value on the horizontal axis, from 300 manifest-years per phase drawn from Table 2 and NASA's lander cadence. Dotted lines mark existing and proposed capabilities.

PNG · PDF (repository)

Version 2 figure: fragmentation law
Figure 16. The fragmentation law. A: in-house urgent 90th-percentile delay for 45 fleet configurations (fleet size by colour, other-duty share by marker; whiskers are 95% CIs) against effective vehicles per owner, with the fitted curve 25.1 + 20.5/e. B: pooling gain (in-house minus marketplace) against the number of owners, for the same configurations.

PNG · PDF (repository)

Version 2 figure: anticipation
Figure 17. Anticipatory staging before published landings, with and without, on five fleet-and-phase cells (30 seeds; 95% CIs). A: urgent 90th-percentile delay. B: urgent mean delay. C: empty kilometres driven to stage. Differences in A and B are within confidence intervals in every cell.

PNG · PDF (repository)

Version 2 figure: precision
Figure 18. Sequential replication of the Phase-2 default-fleet cells. 99% confidence-interval half-width as a share of the mean, against the number of replications, for the mean urgent delay (A) and its 90th percentile (B); the dashed lines are the targets.

PNG · PDF (repository)

Version 2 figure: reference design
Figure 19. Reference design of the lunar logistics marketplace: customers above, the neutral platform in the middle with its seven functions, vehicle operators and surface assets below. Arrows show the two interfaces the platform owns: orders, quotes, tracking and invoices towards customers; job offers, bids, telemetry and payments towards operators.

PNG · PDF (repository)