Plans for robot crews building on the Moon. NVIDIA Nemotron proposes the work; OR-Tools CP-SAT proves it, or refuses with the numbers.
Open the landing pad runThe landing pad run’s verified schedule: 12 tasks for 7 robots, done at hour 72 of a 336-hour deadline. The bronze row is the task the solver refused twice. Choose any task to read it.
Each was recorded live on Nebius Token Factory on 24 October 2026 and replays in the browser. Pick a run to see what the checks refused, how Nemotron repaired it, and where it ended.
A 20 m landing pad and a 40 m blast berm near Shackleton crater: 7 robots, 14 days, no recharging.
energy:printer, 95 Wh over.A 30 m heavy-lander pad and a 70 m spoil berm in 60 hours. The brief forbids a smaller pad.
horizon and zone:Z1. Five pad tasks need 75 h in one zone.A foundation pit, a 70 m plume berm and a borrow strip in 3 days, before a crew lander arrives.
horizon, 10 precedence edges, robot:SV1, zone:Z2.Three exchanges from the landing pad run, with its real numbers. The model reads and writes; the solver checks and decides.
First, write the brief in plain English.
Requirements · Nemotron 3.5 Lightning · 7.3 s
| Structure | Size | Work |
|---|---|---|
PAD1 pad | Ø 20 m, 0.4 m deep | excavate, compact, sinter |
BERM1 berm | 40 × 4 m, 1.5 m high | haul, compact, sinter |
Then Nemotron proposes the tasks, and the checks answer with numbers.
T9 | sinter PAD1 · printer | 314 m² | 2,702 Wh |
T10 | sinter BERM1 · printer | 160 m² | 1,376 Wh |
Refusedenergy:printer
Printer tasks need at least 4,095 Wh: 4,078 Wh of work and 18 Wh of travel. PR1 carries 4,000 Wh.
Last, repair from the proof. Try it yourself.
Refusedenergy:printer
PR1 needs 4,095 Wh of 4,000
2,702 pad + 1,376 crest + 18 travel
Nemotron’s first graph. Over by 95 Wh, so the checks send it back with this core.
CP-SAT’s receipt for the 145 m² plan
| Precedence | 18 edges hold |
| Robots | 7, no overlaps |
| Safety zones | 2, one robot each |
| Batteries, with travel | PR1 3,966 of 4,000 Wh |
| Deadline | 72 of 336 h |
| Status | optimal, 13 ms, re-checked |
Nemotron gets three repairs. In the last evaluation, 55 % of plans verified on the first graph and 86 % after repairs. When none verifies, CP-SAT tries single changes and returns only the ones it can prove: a later deadline, a bigger battery, one more robot.
Every planning call goes to Nebius Token Factory with a strict JSON schema; Cosmos runs beside Isaac Sim on one GPU, as it would on the rover.
Across 66 live runs, 86 % ended with a plan the solver proved. Shown the solver’s proof, Nemotron repaired 70 % of refused plans; told only “refused”, 32 %.
Earth, 1.3 light-seconds away
Earthrise over Compton crater · LRO, 2015
Nobody steers an excavator across that delay, and nobody on Earth catches a bad plan in time. The plan has to be right before it runs.
Shackleton crater, south pole
LROC and ShadowCam mosaic · NASA, 2023
The demo’s site, LS-1, is a plateau 1.8 km east of this rim: 80 by 80 m of loose regolith, a depot on its south edge, nothing to recharge from.
Regolith
Bootprint · Apollo 11, 1969
Loose to 40 cm over firmer ground. The pad is dug 0.4 m to firm regolith, compacted, then sintered 30 cm deep.
Batteries are the budget
Lunar rover practice drive · Apollo 16, 1972
With nothing to recharge from, each robot’s battery covers its whole job, driving included. The solver counts every leg it drives.
When a robot fails
RASSOR excavator test · NASA Kennedy, 2025
CP-SAT replans the remaining work in a median 9 ms. With three spares, the landing pad plan finishes on time in 86 % of 200 random-failure runs, up from 37 %.
The run opens on its brief. Step through the refusal, both repairs, the verified schedule and a rover that stalls in Isaac Sim, then break any robot yourself before the 200 Monte Carlo runs.
Open a run
Replays the run recorded on Token Factory on 24 October 2026. No sign-in, no backend.
Open the landing pad run