13
$\begingroup$

Technology has apparently advanced to the point some states in the Untied States are debating AI managed cars. The thought comes to mind that space is relatively less crowded than most streets on Earth. An AI system could probably be tested out there with less risk than down here. Of course the same argument may be used to test AI systems in the oceans rather than on roads .. Apparently on most modern jetliners too in-flight controls are managed by a computer bank. But that's a different subject...

AI having been proven to be mature enough to handle atleast traffic patterns in a city, the questions that come to mind are -

  • What systems/programs in a mission may be considered for delegation to AI?
  • Is there any move to use AI onboard systems in the space program?
$\endgroup$
  • $\begingroup$ Ah. Well, we're discussing spaceex aren't we? Surely there may be scenarios where we would think thrice before sending a primate in ... where an autonomous system may simply be better suited. Say we develop systems to explore the Io, or Europa. Places where the system should be able to adapt, and time-lag (just two but there may be more) could be critical - AI could probably handle that better autonomously $\endgroup$ – Everyone Aug 16 '13 at 17:07
  • $\begingroup$ Yes... but it seems more sci-fi-y to me. Downvote removed $\endgroup$ – Undo Aug 16 '13 at 17:09
  • $\begingroup$ Forgive me if this appears to be argumentative - it isn't meant to be. But with AI already this mature IMO it's time AI started to be considered legit - if it is not already. $\endgroup$ – Everyone Aug 16 '13 at 17:13
  • 7
    $\begingroup$ Depending on your definition of AI, lots of satellites already have some degree of autonomy, particularly when it comes to fault detection, isolation, and recovery (FDIR) and "routine" things like momentum management. $\endgroup$ – user29 Aug 16 '13 at 17:16
  • 2
    $\begingroup$ Too often people think AI is "a machine that thinks like a human" and forget AI is your laundry machine adjusting speed and rotating the drum in such a way as to make the laundry distributed evenly for rapid spinning. $\endgroup$ – SF. Aug 16 '13 at 23:45
8
$\begingroup$

There are already many systems that fall under the umbrella of Artificial Intelligence in use in space (and on earth). Here are two examples:

However, if you are wanting to know when HAL 9000 will appear Soyuz spacecraft, the answer is probably not within our lifetimes...

$\endgroup$
  • 1
    $\begingroup$ There's more than a single robot on board the ISS at the moment. I was also hoping for a bit more detailed answer than listing a few AI currently in use in space missions, there's more examples listed just in the comments to the question. $\endgroup$ – TildalWave Aug 16 '13 at 17:46
  • $\begingroup$ Sorry, I realized afterwards that comments had been posted while I was writing my answer $\endgroup$ – mayaPapaya Aug 16 '13 at 17:53
  • 1
    $\begingroup$ HAL is probably too far in the future for this question (+: But I did wonder about the approximate complexity equivalent of the Google Car, or better on board missions. Chris' example of momentum control, and your mention of the Rovers autonomous mobility are good examples - albeit one may argue these do not involve 'learning' and nets. $\endgroup$ – Everyone Aug 16 '13 at 18:13
5
$\begingroup$

It depends entirely on what you qualify as 'artificial intelligence'. I'm not aware of any self-learning systems in space to the point of worrying about SkyNet, but there are plenty of ones that are built to be more autonomous.

For instance, on STEREO, if the spacecraft detects something interesting, it'll take additional images, and if there's sufficient time in the next DSN passes, it'll be downlinked.

In the earth-observing realm, there's the Autonomous Sciencecraft Experiment which uses analysis software to detect interesting events (eg, erupting volcano) and focus instruments there. Their website says the spacecraft does 'machine learning' so it's possible that it might qualify as AI. The software from ASE (Sciencecraft) was also used on the Mars Rovers.

Unfortunately, NASA stopped funding the AISR program years ago, which was for developing low TRL technology that could be used to do better science. There's still an archive of past PI presentations, and you can try digging through archive.org for more info.

(disclaimer: I work for the Solar Data Analysis Center which manages the STEREO Science Center, and I might have been on AISR grant review boards)

$\endgroup$
4
$\begingroup$

The Buran spacecraft ( Snowstorm or Blizzard ), a Soviet orbital vehicle analogous in function and design to the US Space Shuttle, but it made unmanned spaceflight, but again, only once in 1988.

It was the first Unmanned Space Vehicle ( USV ) with humans on board, all of
blast-off, flight, landing made by machine.

However, it was destroyed in a 2002 hangar collapse.

enter image description here

$\endgroup$
  • $\begingroup$ Not sure if entry and landing autopilot qualifies as Artificial Intelligence. $\endgroup$ – Deer Hunter Aug 19 '13 at 10:24
  • $\begingroup$ @DeerHunter: If autofocus in a compact camera does, then autonomous landing on autopilot definitely does. $\endgroup$ – SF. Aug 19 '13 at 11:06
  • 1
    $\begingroup$ That picture makes me cry. $\endgroup$ – Mark Adler Mar 9 '14 at 5:31
3
$\begingroup$

It depends on your definition of artificial intelligence.

Is a Kalman filter AI? By most definitions, no. Yet a Kalman filter's job is to generate an ever improving estimate of state. It "learns" by comparing what it think it knows (imperfectly) versus the imperfect results from sensors. Some Kalman filters are extremely sophisticated. They handle sensors that report at different rates, they estimate biases in the sensors as well as where the vehicle is, and they help detect sensor failures. The mathematical hearts of a Kalman filter and a Bayesian belief network (which is AI) are one and the same. The key reason a Kalman filter isn't deemed to be AI is that it wasn't developed by AI researchers.

Is a star tracker AI? A modern star tracker looks for multiple bright stars at once and matches the detected stars against a database of stars. It is doing simple pattern matching. There's not much room in a star tracker's microprocessor for any advanced AI techniques. There's not much room for what most would call AI, period.

Is the Autonomous Landing Hazard Avoidance Technology (ALHAT) being developed by NASA AI? This is a lot closer to what many would call AI. However, is edge detection AI? Is detecting circles and ellipses (aka craters) AI? Does it matter whether the algorithm was developed by an AI practitioner or a signal processing expert? Does matching the detected craters against a database AI? Does it matter that the point of all this is to feed the information to a Kalman filter?

That's the long to midrange part of ALHAT. As the vehicle get's close to the landing site, the technology has to switch from "where am I" mode to "where can I land" mode. It must autonomously find and guide the vehicle to a good landing site. A good landing site is relatively flat and free of obstructions. Does it matter if the technique used to detect a good landing site uses what most would call AI to categorize obstructions as rocks, versus purely mathematical techniques such as wavelet or Fourier transformations?

It shouldn't matter whether the techniques came out of the signal processing community or the AI community, but it does. (Better: AI communities; there's a noted lack of communications amongst different branches of AI, and even a lot of bickering of what AI is.) That I could rewrite a Kalman filter as a Bayesian network and suddenly it's AI? That doesn't make sense.

Another issue that gets in the way of "true" AI from being used in space operations is the languages preferred by AI researchers. LISP won't fly, nor will Prolog (Buran excluded, and it didn't fly). The last thing one wants to happen is to have the computer go into garbage collection mode just as the vehicle is about to land. The algorithms have to be translated to the stodgy languages used by flight software developers, and then toned down further because dynamic memory allocation is taboo in flight software.

One last thing that gets in the way of employing AI in space exploration: A state of the art flight computer is a state of the practice desktop computer from the previous millennium. We won't be able to say Siri, open the pod bay doors on a flight computer anytime soon. Your smart phone has a much faster CPU and a whole lot more memory than does the flight computer on a spacecraft.

$\endgroup$
  • $\begingroup$ "We won't be able to say Siri, open the pod bay doors on a flight computer anytime soon." Of course not, you need to say "open the pod bay doors please". $\endgroup$ – a CVn Jun 5 '14 at 21:01
  • $\begingroup$ "Your smart phone has a much faster CPU and a whole lot more memory than does the flight computer on a spacecraft." That's certainly true, but I feel it's worthwhile to point out it's partly because the failure mode is so very much more catastrophic. If your smartphone breaks that's an annoyance. If an airliner autopilot fails it's a major nuisance, but often recoverable. If the spacecraft flight computer fails, the hardware and crew can very easily be lost. And it's subjected to a very much more harsh environment. Thus, no taking chances. $\endgroup$ – a CVn Jun 5 '14 at 21:02
  • $\begingroup$ I wasn't denigrating flight computers. It's just how it is. The primary reason is that flight computers need to be robust against radiation. The traditional approach is rad hardened processors, and those are a good ten, fifteen years behind what's on your desktop or in your cellphone. SpaceX took a non-traditional route. That means their avionics has a lot more processing power, but it also means the SpaceX Dragon is more susceptible to single event upsets and the like. It was a well-thought through tradeoff by SpaceX. $\endgroup$ – David Hammen Jun 5 '14 at 22:35
  • $\begingroup$ "The primary reason is that flight computers need to be robust against radiation." Agreed. $\endgroup$ – a CVn Jun 6 '14 at 13:13

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.