Roman, New Horizons, Hayabusa2, DART and Hera look like nothing else in the AI industry - they publish the raw data, failures included.
The Roman telescope - named after Nancy Grace Roman, NASA's first chief astronomer - launches at the end of August on a SpaceX Falcon Heavy rocket, the company owned by Elon Musk, from the Cape Canaveral spaceport. Its mirror is actually a gift from a US military intelligence agency - a spy satellite, never launched, redirected from watching Earth to watching the sky. The goal is to count over 100,000 new planets beyond the solar system, more than every telescope in the history of astronomy combined, because its field of view is so wide it covers in one shot a patch of sky that would take the older Hubble telescope thousands of separate photos.
New Horizons is the probe that flew past Pluto - a dwarf planet at the edge of the solar system - about ten years ago. Today it's 9.5 billion kilometers from Earth, humanity's farthest working spacecraft after the two Voyager probes from the 70s. In June it woke up from its longest artificial sleep yet: 321 days, programmed in advance, because the signal travels almost nine hours one way and no one on Earth can react live. It isn't flying toward anything specific. It's just listening to the dust and particles in the Kuiper Belt, a ring of icy leftovers beyond Neptune, in a place no other working spacecraft has ever reached.
The Japanese agency JAXA, their version of NASA, sent Hayabusa2 to grab a sample from the asteroid Ryugu and bring it back to Earth - the probe finished the job in 2020. Instead of shutting it down, engineers left it fuel for a bonus tour. This July it flew past another asteroid, Torifune, as wide as a football stadium, at less than a kilometer, and found it isn't a solid rock but a pile of debris stuck together. The next stop isn't scheduled until 2031.
A few years ago, NASA deliberately crashed a probe the size of a vending machine into Dimorphos, the small moon of the asteroid Didymos, to test whether a pure mechanical impact could deflect its trajectory. It worked - Dimorphos's orbit changed. Now the European agency ESA is sending the Hera probe to inspect the crime scene up close, to measure exactly what's left of the crater and how much the debris weighs - data DART physically couldn't collect while it was crashing. Hera arrives at the end of the year, earlier than planned, because its engines held up longer than engineers expected.
Here's what holds up
None of the five missions hides a result it doesn't like. Hayabusa2 announced outright that Torifune doesn't match the model - a pile of debris, not a rock - and didn't rewrite the earlier forecast into a quietly deleted sentence. DART didn't hide that the first months showed a gap between the calculation and the impact's real effect; that's exactly why they're sending Hera to check on site where the model got it wrong. Roman promised to release its raw data publicly, not wait for a "final" result for a press conference.
Compare that to half the AI labs I write about here - they announce a "breakthrough" on a Friday afternoon and quietly delete the benchmark after the first criticism. Space agencies work with budgets voted by parliaments and schedules written a decade in advance; they don't have the luxury of sweeping a mistake under the rug, because the next mission depends on the exact numbers from this one. That's why I like them more than most things I write about: admitting a mistake here is built into the process, not an exception to it.