When Commands Arrive Twenty Minutes Late: How AI Steers Our Spacecraft

How do you drive a spacecraft millions of kilometres away if every command reaches it late? A NASA study (2022) puts the maximum one-way delay between Earth and Mars at about 22 minutes. And during solar conjunction, when the Sun stands between the two planets, communication is cut or disrupted for roughly two to three weeks on every mission. The practical answer is that you do not drive it; you give it some power to decide.

A rover that picks its path and its target

The Curiosity rover used to fire its ChemCam laser at a preset angle, and about half the time the beam hit soil rather than rock, according to NASA's Jet Propulsion Laboratory (JPL). Then the AEGIS system entered service in May 2016. It examines camera images for rocks, ranks them by colour, shape and size, and aims the instrument at the most suitable. Over a little more than a year it directed the laser 54 times and helped detect higher amounts of chlorine and silica in nearby rocks.

In driving, NASA says the autonomous navigation of Perseverance maps the ground ahead and chooses a route, with a top speed of about 120 metres an hour against about 20 for Curiosity. Human drivers stayed at their posts: the aim was to give the rover more independence where possible, not to remove people.

Telescopes that sift the sky

On Earth, telescopes face a flood of faint measurements. In 2017, Christopher Shallue and Andrew Vanderburg trained a neural network on signals from the Kepler mission. It ranked plausible planets above false positives in 98.8% of cases in its test set. Applied to known multi-planet systems, it allowed an eighth planet around Kepler-90 to be statistically validated, bringing its known planets to the same number as in our own solar system.

The 98.8% figure is a result on pre-labelled data, not a guarantee that the model will be right on every new observation. That is why the astronomers' review remains part of the discovery. NASA now lists more than 6,000 confirmed planets beyond our solar system, a sign of what human effort achieves when computation supports it.

A crowded sky

Ground networks track about 40,000 objects in orbit, roughly 11,000 of them working satellites, according to the European Space Agency's report (2025). The number of fragments larger than one centimetre is estimated at over 1.2 million. The report notes that events triggering collision-avoidance manoeuvres rise year after year. In a sky like that, the need to compute risks quickly grows, but the report does not specify the role of AI in it, so we should not credit it with what was not attributed to it.

What the machine does not see

The case of AEGIS shows something important: the machine's curiosity here is criteria written by people, a rock's colour, size and shape, which scientists can adjust. A system built on the familiar is likely to catch the familiar, and may treat the anomalous as noise. The history of science shows that some discoveries began as an anomaly that looked like an error, so it is wise for a team to review, from time to time, what the algorithm discarded.

Readers may also reflect on what they delegate. Calculation, sifting and long patience are work the machine does well; the question of why we look here and what it means is part of the researcher's trust. Part of being grateful for the blessing of knowledge is using its tools well: taking the means, and keeping in the loop a mind that wonders, asks and reviews.