Author Topic: How Can AI Be Used for Space Applications?  (Read 71559 times)

Offline Twark_Main

  • Senior Member
  • *****
  • Posts: 3754
  • Technically, we ALL live in space
  • Liked: 1950
  • Likes Given: 1216
Re: How Can AI Be Used for Space Applications?
« Reply #360 on: 12/15/2023 01:52 am »
Previous prototype Optimus bots appeared to have a ITX motherboard with a big PC GPU strapped to the chest

I presume it's the same Autopilot hardware Tesla uses in their cars (like they said in the presentation), but it's funny how people see two cooling fans side-by-side and jump to "PC GPU."   8)



There are two major computing processes in AI:  Neural Net transfer function (NN) and Feedback for learning, said learning creating the weights for the neural net.

To aid searching, these are often termed "inference" and "training."

Does anyone know where the NN learning compute is done for robots?

If it's like Tesla cars (which I agree, it probably is), the inference hardware just detects "surprise" and uploads a 30 second clip to the mothership. Back at the mothership Tesla does training (using both Nvidia A100 GPUs and their own custom Dojo hardware) and it sends the fully-trained NN back down.

For safety reasons you really wouldn't want Tesla Bot (or cars) to re-train their own NNs. They run an extensive test harness of safety checks before pushing out a new NN, but you couldn't do that if each unit was doing online training in the field.
« Last Edit: 12/15/2023 02:04 am by Twark_Main »

Offline sanman

  • Senior Member
  • *****
  • Posts: 6009
  • Liked: 1324
  • Likes Given: 8
Re: How Can AI Be Used for Space Applications?
« Reply #361 on: 12/18/2023 07:48 am »
Tesla bot has a growing list of competitors:





The reverse knee on Digit looks like a useful variation. Our own anthropomorphic bias seems to prefer the standard android-like configuration, but there's no need to be confined to that.

This one looks even more practical, with its combination of limbs+wheels. It also looks more suited for the name Optimus - maybe Musk was premature in selecting which robot bandwagon to jump onto.




I'm imagining that some multi-limbed robot could move across the exterior hull of a space station, shifting from one hand-hold to another, tasked with inspection or repair.
« Last Edit: 12/18/2023 08:12 am by sanman »

Offline Twark_Main

  • Senior Member
  • *****
  • Posts: 3754
  • Technically, we ALL live in space
  • Liked: 1950
  • Likes Given: 1216
Re: How Can AI Be Used for Space Applications?
« Reply #362 on: 12/18/2023 11:32 am »
I'm imagining that some multi-limbed robot could move across the exterior hull of a space station, shifting from one hand-hold to another, tasked with inspection or repair.

"The best limbs is the fewest limbs."  -- Elron Moosk

Offline sanman

  • Senior Member
  • *****
  • Posts: 6009
  • Liked: 1324
  • Likes Given: 8
Re: How Can AI Be Used for Space Applications?
« Reply #363 on: 01/03/2024 05:39 pm »
Meet 'Coscientist,' Your AI Lab Partner

https://www.sciencedaily.com/releases/2023/12/231221012729.htm

Quote
Date: December 21, 2023
Source: National Science Foundation
Summary:
An artificial intelligence-driven system has autonomously learned about certain Nobel Prize-winning chemical reactions and designed a successful laboratory procedure to make them. The AI did so in just a few minutes and correctly on its first attempt. According to the authors, this is the first time that a non-organic intelligence planned, designed and executed this complex reaction that was invented by humans.

Sounds like we'll have a future of "Co-everything"
« Last Edit: 01/03/2024 05:41 pm by sanman »

Offline sanman

  • Senior Member
  • *****
  • Posts: 6009
  • Liked: 1324
  • Likes Given: 8
Re: How Can AI Be Used for Space Applications?
« Reply #364 on: 01/05/2024 11:32 am »
Google's new robot, Mobile Aloha (where do they come up with these random names?) might become your automated butler or housekeeper of the (near) future:




So analogous to a Turing test, being able to cook meals has got to represent some kind of useful threshold of capability for robotics. If robots can be trained to do that, then they might be able to do any complex job on an assembly line, or build you a mountain cabin on a mountaintop.

That robot moonbase idea is looking more and more within reach.

https://docs.google.com/document/d/1_3yhWjodSNNYlpxkRCPIlvIAaQ76Nqk2wsqhnEVM6Dc/edit
« Last Edit: 01/05/2024 11:54 am by sanman »

Offline Asteroza

  • Senior Member
  • *****
  • Posts: 2876
  • Liked: 1099
  • Likes Given: 33
Re: How Can AI Be Used for Space Applications?
« Reply #365 on: 01/07/2024 11:30 pm »
Google's new robot, Mobile Aloha (where do they come up with these random names?) might become your automated butler or housekeeper of the (near) future:




So analogous to a Turing test, being able to cook meals has got to represent some kind of useful threshold of capability for robotics. If robots can be trained to do that, then they might be able to do any complex job on an assembly line, or build you a mountain cabin on a mountaintop.

That robot moonbase idea is looking more and more within reach.

https://docs.google.com/document/d/1_3yhWjodSNNYlpxkRCPIlvIAaQ76Nqk2wsqhnEVM6Dc/edit

Webpage for the opensource hardware and software for mobile ALOHA (uses a commercial CoTS AGV body for body base/propulsion)
https://mobile-aloha.github.io/

I vaguely remember the ISS SPHERES program, and the more PR friendly JAXA Int-ball and ESA CIMON, but to actually get stuff done needs manipulators so you get into Robonaut 2 territory.

Offline sanman

  • Senior Member
  • *****
  • Posts: 6009
  • Liked: 1324
  • Likes Given: 8
Re: How Can AI Be Used for Space Applications?
« Reply #366 on: 01/25/2024 02:36 am »
Google DeepMind's AlphaGeometry is good at solving geometry problems:


Offline sanman

  • Senior Member
  • *****
  • Posts: 6009
  • Liked: 1324
  • Likes Given: 8
Re: How Can AI Be Used for Space Applications?
« Reply #367 on: 02/18/2024 07:22 am »
Webpage for the opensource hardware and software for mobile ALOHA (uses a commercial CoTS AGV body for body base/propulsion)
https://mobile-aloha.github.io/

I vaguely remember the ISS SPHERES program, and the more PR friendly JAXA Int-ball and ESA CIMON, but to actually get stuff done needs manipulators so you get into Robonaut 2 territory.

Here's something to consider:

https://en.wikipedia.org/wiki/Manifold_hypothesis

The Manifold Hypothesis basically says that real-world objects can be more efficiently/compactly represented in a latent space. This in turn can lower computational loads and memory footprints for calculations relating to them.

Can we take a high-dimensional representation of some real-world system like a launch vehicle or a rocket engine, and then condense it into a more compact representation in a latent space, in order to generate a "continuum" of various launch vehicle designs, or engine designs?
« Last Edit: 02/18/2024 07:23 am by sanman »

Online Solarsail

  • Member
  • Posts: 66
  • Liked: 29
  • Likes Given: 8
Re: How Can AI Be Used for Space Applications?
« Reply #368 on: 03/03/2024 04:16 pm »
Thinking of it, that observation that trade spaces around physical systems (engineering) fall into lower-dimensional planes of valid solutions within higher dimensional spaces, was what I was thinking of addressing with the mention of Prolog.  If you can figure out how to take a representation of a low-dimensional space within a high-dimensional physical description space... and perform operations on that, then you're getting somewhere.  If you've found a surface passing through n-space that amounts to an engineering solution to one set of concerns, and you add a new constraint to it, you've generated a subset of the previous solution space which still works with the new constraint.  If you can generalize that out to "cost" or some sort of difficulty of getting a particular solution to work, then you can get the algorithm to search for a global optima.

If memory serves...  Prolog tried to do something like that in the 80's.  The Prolog project wound up being a constraint solver with a fancy syntax, after all.  However, that constraint solver wound up being so limited in its ability to find solutions to arbitrary problems that it couldn't even encode (or reason about) arithmetic.  Many demonstrations of Prolog wound up similar to:

A man is a human.
Homer is a man.
(Prolog interpreter concludes) Homer is a human.

...Basic reasoning about set membership.

So; just how different are the 'generative neural nets' of this era's software?  Should I infer that current systems more or less encode the exact latent space that is described in their learning sets?  I've definitely read that human researchers cannot actually reason about the algorithms that GNNs implement.  Is anybody able to do operations upon these learned latent spaces?

Offline sanman

  • Senior Member
  • *****
  • Posts: 6009
  • Liked: 1324
  • Likes Given: 8
Re: How Can AI Be Used for Space Applications?
« Reply #369 on: 03/04/2024 08:45 pm »
Thinking of it, that observation that trade spaces around physical systems (engineering) fall into lower-dimensional planes of valid solutions within higher dimensional spaces, was what I was thinking of addressing with the mention of Prolog.  If you can figure out how to take a representation of a low-dimensional space within a high-dimensional physical description space... and perform operations on that, then you're getting somewhere.  If you've found a surface passing through n-space that amounts to an engineering solution to one set of concerns, and you add a new constraint to it, you've generated a subset of the previous solution space which still works with the new constraint.  If you can generalize that out to "cost" or some sort of difficulty of getting a particular solution to work, then you can get the algorithm to search for a global optima.

If memory serves...  Prolog tried to do something like that in the 80's.  The Prolog project wound up being a constraint solver with a fancy syntax, after all.  However, that constraint solver wound up being so limited in its ability to find solutions to arbitrary problems that it couldn't even encode (or reason about) arithmetic.  Many demonstrations of Prolog wound up similar to:

A man is a human.
Homer is a man.
(Prolog interpreter concludes) Homer is a human.

...Basic reasoning about set membership.

So; just how different are the 'generative neural nets' of this era's software?  Should I infer that current systems more or less encode the exact latent space that is described in their learning sets?  I've definitely read that human researchers cannot actually reason about the algorithms that GNNs implement.  Is anybody able to do operations upon these learned latent spaces?

Neural networks are capable of compressing data into a more efficient latent space representation. Once data is in that form, it is much more efficient to analyze and manipulate.

I was thinking that some 3D finite element analysis could then be carried out along all 3 axes bidirectionally with useful features being extracted, even as the system evolves across time.

Maybe we could have an FEA-Transformer.
« Last Edit: 03/08/2024 09:00 pm by sanman »

Offline sanman

  • Senior Member
  • *****
  • Posts: 6009
  • Liked: 1324
  • Likes Given: 8
Re: How Can AI Be Used for Space Applications?
« Reply #370 on: 03/14/2024 12:22 am »
These are the droids you are looking for



Offline sanman

  • Senior Member
  • *****
  • Posts: 6009
  • Liked: 1324
  • Likes Given: 8
Re: How Can AI Be Used for Space Applications?
« Reply #371 on: 04/17/2024 05:04 pm »
Atlas from Boston Dynamics has a new look:




Pretty flexible joint movement.

Could such tricks be useful in the future robotic exploration of the Moon & Mars?
« Last Edit: 04/17/2024 05:05 pm by sanman »

Online catdlr

  • Member
  • Senior Member
  • *****
  • Posts: 11942
  • Enthusiast since the Redstones
  • Marina del Rey, California, USA
  • Liked: 9566
  • Likes Given: 8192
Re: How Can AI Be Used for Space Applications?
« Reply #372 on: 04/17/2024 05:33 pm »
Atlas from Boston Dynamics has a new look:




Pretty flexible joint movement.

Could such tricks be useful in the future robotic exploration of the Moon & Mars?

BD put out a retirement video yesterday for the previous Atlas version and today the video you posted is the intro for the new Atlas  Ser# 001.  The new version is so much slimmer and its center of gravity issue with the first version is (hopefully) reduced.

« Last Edit: 04/17/2024 05:35 pm by catdlr »
It's Tony De La Rosa, ...I don't create this stuff, I just report it.

Offline sanman

  • Senior Member
  • *****
  • Posts: 6009
  • Liked: 1324
  • Likes Given: 8
Re: How Can AI Be Used for Space Applications?
« Reply #373 on: 05/06/2024 03:49 pm »
BD put out a retirement video yesterday for the previous Atlas version and today the video you posted is the intro for the new Atlas  Ser# 001.  The new version is so much slimmer and its center of gravity issue with the first version is (hopefully) reduced.

The new one has all-electric actuators and gets rid of the hydraulics. (Reminds me of Starship changing to e-TVC)

The joints can now move in new ways that weren't previously possible. Witness the new yoga flexibility.

But can it still do Parkour? Does it have the power for that? Or have there been tradeoffs with the new design?

Offline Star One

  • Senior Member
  • *****
  • Posts: 14076
  • UK
  • Liked: 4001
  • Likes Given: 220
How Can AI Be Used for Space Applications?
« Reply #374 on: 05/13/2024 08:36 pm »
A major update for Chat-GPT.

Introducing Chat-GPT 4o:

https://www.youtube.com/live/DQacCB9tDaw?si=yExFQ8OGLnqB4UA9

Two Chat-GPT 4os interacting and singing:



Math problems with Chat-GPT 4o:



Live demo of Chat-GPT 4o coding assistant and desktop app:



It seems like they’ve taken the training wheels off Chat-GPT. Google’s turn tomorrow with their big AI announcement. Can imagine AI might be useful on long term space exploration as a support assistant. Mind you radiation levels would probably preclude it really.

https://arstechnica.com/information-technology/2024/05/chatgpt-4o-lets-you-have-real-time-audio-video-conversations-with-emotional-chatbot/
« Last Edit: 05/13/2024 09:24 pm by Star One »

Offline InterestedEngineer

  • Senior Member
  • *****
  • Posts: 2414
  • Seattle
  • Liked: 1840
  • Likes Given: 3026
Re: How Can AI Be Used for Space Applications?
« Reply #375 on: 05/14/2024 04:52 am »

It seems like they’ve taken the training wheels off Chat-GPT. Google’s turn tomorrow with their big AI announcement. Can imagine AI might be useful on long term space exploration as a support assistant. Mind you radiation levels would probably preclude it really.


The model runs in a data center somewhere.   Using ChatGTP is interactive, and it does require iteration to get a useful answer.

No imagine you are 10 light minutes away.  Interaction is going to painfully slow.

Now, the compute power to actually run the model probably isn't so bad, so having a data center on a Mars base connected via Starlinks will probably work fine.   But I doubt you want that kind of power and mass budget for a traveling spaceship.

Offline Star One

  • Senior Member
  • *****
  • Posts: 14076
  • UK
  • Liked: 4001
  • Likes Given: 220
How Can AI Be Used for Space Applications?
« Reply #376 on: 05/14/2024 06:17 am »

It seems like they’ve taken the training wheels off Chat-GPT. Google’s turn tomorrow with their big AI announcement. Can imagine AI might be useful on long term space exploration as a support assistant. Mind you radiation levels would probably preclude it really.


The model runs in a data center somewhere.   Using ChatGTP is interactive, and it does require iteration to get a useful answer.

No imagine you are 10 light minutes away.  Interaction is going to painfully slow.

Now, the compute power to actually run the model probably isn't so bad, so having a data center on a Mars base connected via Starlinks will probably work fine.   But I doubt you want that kind of power and mass budget for a traveling spaceship.
Precisely. If you wanted a job free from AI then working on something like Mars exploration would seem to be ideal. All these fantasies about advanced robots in deep space beyond what we have now are just that fantasies because they never consider things like radiation limitations on what you can carry etc.
« Last Edit: 05/14/2024 06:20 am by Star One »

Offline Twark_Main

  • Senior Member
  • *****
  • Posts: 3754
  • Technically, we ALL live in space
  • Liked: 1950
  • Likes Given: 1216
Re: How Can AI Be Used for Space Applications?
« Reply #377 on: 05/14/2024 07:17 am »

It seems like they’ve taken the training wheels off Chat-GPT. Google’s turn tomorrow with their big AI announcement. Can imagine AI might be useful on long term space exploration as a support assistant. Mind you radiation levels would probably preclude it really.


The model runs in a data center somewhere.   Using ChatGTP is interactive, and it does require iteration to get a useful answer.

No imagine you are 10 light minutes away.  Interaction is going to painfully slow.

Now, the compute power to actually run the model probably isn't so bad, so having a data center on a Mars base connected via Starlinks will probably work fine.   But I doubt you want that kind of power and mass budget for a traveling spaceship.
Precisely. If you wanted a job free from AI then working on something like Mars exploration would seem to be ideal. All these fantasies about advanced robots in deep space beyond what we have now are just that fantasies because they never consider things like radiation limitations on what you can carry etc.

I wouldn't get my hopes up.  ;)

Onboard the ship is the one place you really want to invest in supporting a local AI. Imagine an AI trained on every mission control transcript, procedure, and NASA document ever, locally available without that pesky speed-of-light delay to the actual Houston, Texas.

This might be the ideal place for analog AI computing. It's low power, and energetic particles should "just" result in a higher noise floor, instead of the (numerically catastrophic) bit flips in digital systems.




« Last Edit: 05/14/2024 08:21 am by Twark_Main »

Offline Star One

  • Senior Member
  • *****
  • Posts: 14076
  • UK
  • Liked: 4001
  • Likes Given: 220
How Can AI Be Used for Space Applications?
« Reply #378 on: 05/14/2024 04:55 pm »
Doctor Mike Pound in relation to a new paper, linked to below, argues that generative AI is likely to son plateau, and that’s going to become increasingly difficult to get any kind of meaningful improvement in its performance.



https://arxiv.org/abs/2404.04125
« Last Edit: 05/14/2024 04:59 pm by Star One »

Offline Remes

  • Full Member
  • ****
  • Posts: 446
  • Germany
  • Liked: 263
  • Likes Given: 146
Re: How Can AI Be Used for Space Applications?
« Reply #379 on: 05/15/2024 07:50 pm »
This might be the ideal place for analog AI computing. It's low power, and energetic particles should "just" result in a higher noise floor, instead of the (numerically catastrophic) bit flips in digital systems.
No. In analog computing you encode one number in a voltage level. In digital computing you have 2 discrete voltage levels but several of them for each number (i.e. bits). A weak particle/ray might not flip your bit of your digital representation, but it might change your analog voltage signal by e.g. 30%.

Now a digital program and logic circuit is quite big, offers a lot of area for catching particles/rays (has memory, memory controlers, caches, instruction decoder, ...). On the other side: slight changes in the structure of your silicon based on particles doesn't matter that much in digital logic (a small threshold change doesn't matter). In analog any change is an offset for all calculations to come. In digital you have easier ways of calculating things redundantly (on different hardwares in parallel, running the same software cycle two times to detect spurious events, running short testvectors). In analog it is not that simple.

I was now thinking specifically about neural networks in digital vs analog and radiation tolerance. I guess for digital and analog the highest sensitivity is at the end of backbones or heads. Or something like attention layers, which have the ability to control which paths are weighed most.

It's never easy to determine really, what is better in which case. Lot of statistics and probabilites. And once in ten thousand years can be tomorrow.

Tags:
 

Advertisement NovaTech
Advertisement Northrop Grumman
Advertisement
Advertisement Margaritaville Beach Resort South Padre Island
Advertisement Brady Kenniston
Advertisement NextSpaceflight
Advertisement Nathan Barker Photography
1