Introducing Semaphore

We’re building Semaphore, a digital twin of the Moon for autonomous systems, their operators, and analysts. It is a unified operational picture of the Moon, from orbit to below the surface. It joins observations of the real Moon with models of it, holds the current state as measurements arrive, and projects what comes next.

We started with public data from NASA, including 17 years of maps from its Lunar Reconnaissance Orbiter (LRO). Next, we launch a screening service that lets lunar operators plan together and assess risk. We also build training and test data from the twin, offer it to others, and train our own models on it. Starting in 2028, we plan to launch hardware and software systems built by Semaphore and begin extending our compute to cislunar space. (NASA Science)

Why the Moon needs a digital twin now

No sensor network routinely tracks spacecraft at lunar distance. NASA’s Jet Propulsion Laboratory (JPL) screens lunar orbits with the paths that missions report about themselves. In 2021, JPL predicted that LRO and India’s Chandrayaan-2 would pass 110 m apart, and Chandrayaan-2 moved. In 2024, some reported paths still arrived by email, a practice JPL calls “not sustainable”. (NASA JPL)

The traffic is about to grow. NASA has awarded 17 commercial deliveries to the Moon through 2028. It plans its first crewed landing for early 2028, on Artemis IV. The European Space Agency (ESA) counts over 400 planned lunar missions in the next two decades. (NASA on deliveries; NASA on Artemis IV; ESA)

Those missions crowd into a few places. The Moon’s uneven gravity pulls most low orbits down, and orbits that hold without fuel are rare. On the surface, NASA’s nine candidate regions for a crewed landing are all near the south pole. China’s Chang’e 7 targets the same area. There, the Sun stays low, and a rover has only a few sunlit corridors to drive in. (NASA Ames on orbits; NASA; Scientific American; NASA Ames on rovers)

Neighbors at the Moon affect each other. A lander’s exhaust throws dust at more than 2 km per second, fast enough to reach hardware nearby and in orbit. A rover’s next move depends on its operator’s plan. Other missions know where it goes only if the operator shares that plan. (NASA Kennedy; NASA Ames)

The ground changes too. Images from LRO found a third more new craters than models predicted. A crater 222 m across formed in spring 2024 and went unnoticed until fall 2025. A lander that tips over stays where it falls. In 2025, one carrying a NASA drill came down on its side. Today’s lunar maps were made for science, not for landing. (Nature; Science Advances; NASA; NASA Goddard)

When a model misses the terrain, missions fail. In 2023, ispace’s Hakuto-R lander crashed after its landing simulations “did not adequately incorporate the lunar environment” along its route. (ispace)

Governments see the gap. Seventy-five nations have signed the Artemis Accords. Each commits to notify and coordinate before it operates in another’s safety zone. No shared system carries those notices today. A United Nations lunar action team calls this a “procedural vacuum”. NASA lists extending U.S. space tracking to cislunar space, the region between Earth and the Moon, among its technology needs. (NASA on the signatories; NASA on the accords; United Nations; NASA on shortfalls)

How we’re building it

First, we seed the platform with public data. LRO’s measurements are public in NASA’s Planetary Data System (PDS). We turn them into a model of the Moon that people and software can query through software interfaces. It holds terrain, sunlight, and the published paths of spacecraft. Operators and scientists add what they publish, and they choose who sees each item. Semaphore adds its own datasets, from derived layers such as sunlight and shadow to the training and test data in the next step. (PDS)

Second, we turn the digital twin into training and test data. Landers and rovers read the Moon through their sensors, and the Moon is hard to read. A lander finds its position by matching its camera to a reference map that NASA requires to match the light at landing. Real images under that light are rare, so NASA renders each map from elevation models and validates it for every site. In 2025, a lander with noisy altimeters came down on its side, because in the low light the craters looked unlike its reference images. The twin holds the Moon’s terrain and the Sun over time, so we build labeled training and test data from them across the whole Moon, as NASA builds its labeled crater data. People review the labels, and each dataset states its sources and its limits. Others train and test their systems on it. We train our own models on it too, for terrain and for the environment, starting with sunlight, shadow, and radiation, and offer them through software interfaces as they become available. Every measurement published to the twin improves the next dataset and the next model. (NASA Goddard; NASA Johnson on rendered maps; SpaceNews; NASA Johnson on crater labels)

Third, starting in 2028, we launch our own hardware and software systems. Public archives update slowly, and no network watches traffic at the Moon. Our systems fill those gaps with independent observation.

Fourth, we move computing closer to the Moon. Semaphore runs in the cloud today, but distance has a cost. On lunar missions, a signal can take 3 to 14 seconds each way, depending on the communication links. NASA’s Deep Space Network is already oversubscribed, with demand at times 40 percent above capacity. As computing moves closer to the Moon, our models move with it. (NASA; NASA Inspector General)

Let’s collaborate

We believe in simple systems that adapt quickly, so we shape Semaphore around what its users tell us. If you run lunar missions, study the Moon, build machines that sense it, or follow it as a hobby, let’s work together.