A rover on Mars spends most of its day being bored on your behalf. Light takes between four and twenty-four minutes to make the trip one way, and a command cycle usually means one uplink in the morning and one downlink at night. Everything in between is a machine executing yesterday’s opinion about today. If it rolls past something remarkable at eleven, nobody on Earth knows until evening, and by then the rover is forty metres away facing the wrong direction.
The fix that has matured over the past decade is to stop sending opinions and start sending appetites. Instead of a list of targets, you upload a model of what interesting looks like and let the vehicle go shopping. The technique comes out of the intrinsic-motivation corner of machine learning, where an agent is rewarded not for reaching a goal but for encountering something its model failed to predict. Researchers at the Halbrecht Institute have spent six years turning that into flight software, and they call the result a curiosity engine, which is either a very good name or a very good grant application.
Surprise as a scientific instrument
The mechanism is less exotic than the name. The rover keeps a compressed statistical picture of terrain it has already seen, built from texture, colour and a coarse mineralogical guess from its multispectral filters. Each new frame is scored by how badly that picture predicts it. A rock like ten thousand previous rocks scores near zero and is discarded onboard. One that does not scores high, gets a full-resolution capture, and jumps the downlink queue ahead of routine navigation imagery. Bandwidth is the scarce resource, and the model is really a triage nurse.
We are not asking the vehicle to do science. We are asking it to notice, which is the part of science that has to happen where the rock is.
That was Dr Ines Karvala, who leads the group’s field programme and has spent two summers in the Atacama driving a test platform across ground chosen to look like Gale Crater. Her most recent field trial ran a 41-day traverse with the engine in full authority over the imaging schedule. It returned 38 per cent fewer images than a conventionally scheduled control run and, in blind assessment by four geologists, produced roughly twice as many frames rated scientifically significant.
Novelty is a treacherous target, as Karvala is first to say. A model that chases surprise will fall in love with a lens flare, an unfamiliar shadow, or a smear of dust on its own optics. The Atacama run wasted an afternoon on a fragment of the test rig’s own packing foam, which was, to be fair, genuinely the most unusual object in the valley. Guarding against that means bolting hard priors onto an otherwise open-ended curiosity, and every prior you add narrows the thing you built the system to find.
Who gets to point the camera
The harder objection is institutional rather than technical. Science teams negotiate for target time in meetings that run for months, and handing an arbitration role to a model trained on last year’s terrain is not a small cultural ask. As two geologists put it at a workshop in Bremen, an autonomous prioritiser is also an autonomous suppressor, and nobody sees the frames it silently threw away. The current compromise is a thumbnail log, a few kilobytes per discarded frame, so a human can audit the machine’s taste afterwards.
None of this matters much at Mars distance, where a bad call costs a day. It matters enormously further out. An ice-shell lander on Europa would have a mission measured in weeks and a communications window measured in minutes. At that range you are not sending a robot to do what you decided. You are sending something you have taught to be interested, and then you are letting go of the wheel.
Leave a Reply