AI is likely to become important to future space travel for a fairly unglamorous reason: Earth is too far away to make every decision in time.

A spacecraft near Mars may wait minutes for a radio command. At the outer planets, the delay stretches into hours. A vehicle approaching an asteroid, descending towards unfamiliar ground or responding to a failing component may need to act before a human team can see the problem, discuss it and send an answer back.

That does not mean future spacecraft will be handed to a talking machine captain. On the evidence available today, the practical role is narrower: recognising a local situation, choosing among approved actions and keeping a mission working within limits set by people.

Spacecraft autonomy began before the current AI boom

NASA tested an onboard artificial-intelligence system on Deep Space 1 in 1999. The Remote Agent experiment received high-level goals, produced plans, executed them and responded to four simulated faults inserted by the mission team. It also encountered a software bug during its first run, a useful reminder that autonomy creates new failure modes as well as new capability.

Deep Space 1 separately tested autonomous optical navigation. The spacecraft photographed asteroids against background stars, calculated its position and commanded its ion propulsion without waiting for engineers to perform each navigation calculation on Earth.

These systems were not generative AI. They were purpose-built planning, diagnosis and navigation software, designed around a constrained mission. That distinction remains useful today.

I wrote earlier about how AI already flying in space tends to solve one specific physical problem. Future capability will probably grow by joining many such bounded systems, not by installing a single general intelligence and asking it to improvise.

DART showed why decisions sometimes have to stay onboard

NASA’s DART spacecraft had to strike Dimorphos, an asteroid moonlet roughly 150 metres across, while closing at about six kilometres per second. During the final phase, human intervention would have been too slow and the target did not separate clearly from the larger asteroid until late in the approach.

DART’s SMART Nav system used images from its camera to identify the bodies, select Dimorphos, calculate corrections and command manoeuvres. The system guided DART through the last four hours without human control, ending in the intended collision on 26 September 2022.

That is autonomy with a sharply defined objective.

The spacecraft was not deciding whether the mission was wise or inventing a new scientific goal. People had designed, tested and authorised the decision structure. The machine handled the part that had to occur faster than the radio link allowed.

Newer AI can assist with planning before it flies

In December 2025, NASA’s Perseverance team completed the first two drives on another world whose waypoints had been generated by AI. A vision-language model examined orbital imagery and elevation data, identified terrain features and proposed a route.

The AI was on Earth, not steering the rover live. JPL engineers ran its commands through a digital replica of Perseverance and checked more than 500,000 telemetry variables before transmission. The rover then drove 210 metres on 8 December and 246 metres on 10 December, according to JPL’s account of the demonstration.

This suggests a near-term division of labour. AI can examine large image sets, propose waypoints and reduce routine planning work. Engineers can test those proposals against mission rules. Onboard navigation can then handle nearby rocks and other hazards as the vehicle moves.

Each layer has a different job, and each can be checked differently.

Hera will test more autonomy around an asteroid

ESA’s Hera spacecraft is due to reach the Didymos system in November 2026 to examine the aftermath of DART’s impact. ESA describes Hera as having autonomy comparable to a self-driving car, although the comparison refers to sensor fusion and local navigation rather than road-driving software transplanted into space.

Hera will combine camera images, laser-altimeter readings, inertial sensors and star-tracker data to build a working picture of the two asteroids. Its experimental modes are intended to recognise surface features and make navigation decisions near bodies whose weak gravity and irregular shapes are difficult to model perfectly from Earth.

If those tests work, the same broad approach could help later spacecraft rendezvous with small objects, inspect damaged satellites or land in places where detailed maps are incomplete. It would be evidence for particular navigation methods, not proof that a spacecraft can manage an entire mission alone.

Future travellers will need systems that know their authority

For robotic probes, autonomy can mean more observations and less time waiting. For crews, it could help monitor equipment, flag anomalies, manage inventories or compare possible responses while Earth is unavailable. Yet a system making suggestions for astronauts is not the same as one controlling life support or propulsion, where an error could be irreversible.

The harder engineering question is therefore not simply how capable the AI becomes. It is how its authority is bounded, how uncertainty is displayed, what happens when sensors disagree and how people regain control.

I recently looked at the much larger physical barrier in the possibility of sending a probe to Alpha Centauri. AI cannot provide the propulsion, shielding or energy such a mission lacks. It could, however, help a small vehicle operate for decades without expecting Earth to supervise every choice.

That is the credible future role: not replacing mission control, but carrying a carefully tested part of mission control aboard the spacecraft when distance makes immediate help impossible.