Good inertial navigation units are really good. I had a professor who liked to tell the story of having one on the top floor of the engineering building and if you left it on all day you could watch your altitude go up and down as the building warmed up with the sun and cooled down at night.
Some level of drift is acceptable depending on the mission. Aircraft and missiles only fly for a few hours at most, and may be able to get a satellite or stellar fix at least intermittently.
dead reckoning has to account for currents etc, which is something you could use AI to improve as per the paper. It's actually a _good_ application of AI?
Subs use very expensive and very impressive inertial guidance systems, which can go up-to a week without correction. But that doesn't scale down to the modern world of small autonomous systems.
Huh, I don't understand why. :-D Someone please explain. If I tugged the sub, wouldn't the IMU detect that? Why is a current doing the tugging any different.
now I'm guessing, but perhaps it's like if you walk up the aisle of a moving train, how fast are you travelling? And how fast does it feel like you're travelling?
Sure, I would get confused by that too, but I thought the point of the IMU is that it has its own frame of reference. Elsewhere in these comments was someone mentioning an IMU which could detect a building growing in size when warmer. The IMU doesn't "know" there's water outside or how the water is moving. It knows how much force it's been pushed around by.
Aha! IMU <> "dead reckoning"? In the wikipedia article they are listed separately. Dead reckoning is meaning tracking things like the revolutions of a propeller and a compass to compute position, whereas an IMU is three accelerometers, three gyros and three magnetometers? https://en.wikipedia.org/wiki/Inertial_measurement_unit
Is it just me or does this feel like another "we must stuff AI into everything, just because"? Inertial navigation has been around for most of a century, i.e. before GPS even existed, and is basically self-contained.
Learning that is difficult. It is a skill that needs to be worked on by learning maths, then control, dynamics. Instead, learn to use an AI framework and you can solve the problem with orders of magnitude more computing.
Not advocating one way or another, just that that's probably the thinking. Also, it's in vogue.
Its far easier to stuff 8 H100s in a submarine than to add another headcount.
Every person in a submarine consumes valuable space/food/oxygen. Cooling and electricity on the other hand are easy to solve, when you have a nuclear reactor and a dead cold ocean right outside.
So if possible, AI will be used to perform these duties.
And there will always be low tech, self contained back up systems in place. And I was not aware that current military subs had problems with navigation that require AI to solve, or issues with life support.
I don't think you understand, inertial navigation is not computing (as in predicting) anything. It is a bunch of sensors to measure where you are.
Using AI to predict where you might be based on crappy cheap sensor data and ocean simulations is never going to be precise enough. AI can't pull signal out of noise and uncertainty.
> Using AI to predict where you might be based on crappy cheap sensor data and ocean simulations is never going to be precise enough. AI can't pull signal out of noise and uncertainty.
Yeah, but people with monies see this as "AI can help replace expensive milspec hardware with cheap COTS garbage plus some software magic pixie dust, for approximately the same effect [during peace time, in harbor]".
AI is a contemporary marketing term. Statistical methods that go under that umbrella term has been used for a few hundred years. AI has now become hip, so anyone using any methods that could conceivably be put under that marketing term is now being done.
AI is definitely a big buzzword right now, but in this particular case inertial navigation has a pretty big limitation: cost. Cheaper sensors with better software would be a legitimately compelling alternative to traditional solutions.
My sense is that the authors' focus was on stabilizing the vehicle in varying conditions (ocean currents), and the training time needed to achieve that, rather than wayfinding. Not sure why the beginning of the article is so oriented towards navigation.
Although underwater vehicles are very much not in my wheelhouse this seems actually more interesting; I wonder what other applications there might be for the learning method they propose.
The sub knows where it is at all times. It knows this because it knows where it isn't. By subtracting where it is from where it isn't, or where it isn't from where it is (whichever is greater), it obtains a difference, or deviation. The guidance subsystem uses deviations to generate corrective commands to drive the sub from a position where it is to a position where it isn't, and arriving at a position where it wasn't, it now is. Consequently, the position where it is, is now the position that it wasn't, and it follows that the position that it was, is now the position that it isn't.
46 comments
[ 4.9 ms ] story [ 110 ms ] threadMore interesting would be to understand how much its position drifts over a period of days or weeks.
Subs use very expensive and very impressive inertial guidance systems, which can go up-to a week without correction. But that doesn't scale down to the modern world of small autonomous systems.
Not arguing that small autonomous systems could use something smaller and cheaper.
Not advocating one way or another, just that that's probably the thinking. Also, it's in vogue.
Every person in a submarine consumes valuable space/food/oxygen. Cooling and electricity on the other hand are easy to solve, when you have a nuclear reactor and a dead cold ocean right outside.
So if possible, AI will be used to perform these duties.
Using AI to predict where you might be based on crappy cheap sensor data and ocean simulations is never going to be precise enough. AI can't pull signal out of noise and uncertainty.
Yeah, but people with monies see this as "AI can help replace expensive milspec hardware with cheap COTS garbage plus some software magic pixie dust, for approximately the same effect [during peace time, in harbor]".
if it checks up with the map of how our brains learn to keep us upright, then we learn other side effects of how brains grow and learn
if it doesn't... then it's even more interesting because there's another confounding variable.
My sense is that the authors' focus was on stabilizing the vehicle in varying conditions (ocean currents), and the training time needed to achieve that, rather than wayfinding. Not sure why the beginning of the article is so oriented towards navigation.
Although underwater vehicles are very much not in my wheelhouse this seems actually more interesting; I wonder what other applications there might be for the learning method they propose.
> Stochastic parrots and other animals in linear algebra zoo ain't going to end the world
> Proceeds to put black-box SOTA ML models at the core of Mutual Assured Destruction dynamics
What could possibly go wrong.
Instead of gps, submarines can use maps of the gravitational constant for navigation
(Couldn’t resist).
* https://en.wikipedia.org/wiki/Underwater_acoustic_communicat...
In 2017, NATO formalized the JANUS standard (STANG 4748):
* https://www.cmre.nato.int/rockstories-blog-display/398-janus...
* https://nso.nato.int/nso/nsdd/main/standards/ap-details/1770...
If it were even possible.