How Can AI Be Used for Space Applications?

[Robotbeat]

Multi-modal GPT-4 should help a lot, as now you can feed in images (for example, from a camera). One example I saw was where someone took a picture of their open fridge and had GPT-4 tell it a recipe to make using the identified contents.

https://twitter.com/sudu_cb/status/1636080774834257920?s=20

[Robotbeat]

I imagine if there’s something like the Apollo 11 1202 alarm that occurs on the first landing on Mars, but mission control is like 15 minutes away from responding due to light delay. GPT-4 or an improvement could be used as a sort of real-time backup to information requests to mission control. It probably doesn’t make sense to rely on GPT-4 for discretion on whether to abort or not, but GPT-4 could respond by giving a description of the function of that alarm and the contingency if it occurs, all by verbal natural language communication with the crew (like with mission control). This is especially helpful for very rare corner cases like that 1202 alarm.

https://www.rmg.co.uk/stories/topics/apollo-11-moon-landing-minute-minute

[sanman]

CETI (the Cetacean Translation Initiative) whose acronym is a play on the NASA-funded SETI (Search for Extra-Terrestrial Intelligence) program, aims to use GPT-style machine learning to talk to whales:



So just imagine if this technology could be used for the actual SETI, serving as a precedent to de-code actual extra-terrestrial communications, should we ever discover any.

Meanwhile, Stanford has announced Alpaca, which is their newer cheaper implementation of LLAMA (Large LAnguage Model AI). It can apparently achieve comparable performance to ChatGPT -- but while ChatGPT cost ~$10M to train, Alpaca only cost a fraction of that ($600) to train, and runs on a PC. The key is that Alpaca used GPT to help train itself.



So we now have the prospect of Large Language Models being used as primary trainers for training cheaper/smaller language models that could be more customized/specialized. MedGPT? SpaceGPT? FinanceGPT?

[Robotbeat]

Yeah, absolutely. You can run a 4 bit quantized version of Llama on your smartphone. Again, I think it might be useful to provide the sort of feedback that mission control might provide when you have a 30 minute time delay at Mars. Probably worth bringing a full GPT4 model with you and run it on the new H100 NVL dual-GPU thing that NVidia offers. With 188GB of VRAM, it should be able to run GPT3.5 and probably 4, at least if using 8 bit quantization. https://www.servethehome.com/nvidia-h100-nvl-for-high-end-ai-inference-launched/

~800 Watts power consumption.

[joek]

Whoa pardner... Advising crew on situation-outcomes-whatever does not necessarily require the "G" (generative) in GPT, which is what differentiates it from other more traditional  and previous AI's. Think you could accomplish what you want with significantly less. As in, less generative and more predictive.

[Robotbeat]

Whoa pardner... Advising crew on situation-outcomes-whatever does not necessarily require the "G" (generative) in GPT, which is what differentiates it from other more traditional  and previous AI's. Think you could accomplish what you want with significantly less. As in, less generative and more predictive.

Pick whichever option is most reliable and flexible. There are ways to test this, and the astronauts should never trust any particular tool 100%, because tools are fallible (like humans).

It is true that sometimes people take ChatGPT arguments as somehow authoritative (as a sort of argument-deciding oracle), and that is a really bad idea and not at all the right attitude, so I’ll agree there.

I’ve actually tried running a few things past ChatGPT/GPT3.5 to see what it’d advise. Interestingly, it never suggests allowing the mission to continue when getting the 1202 alarm, it always “plays it safe” even when I tried with various temperatures. I’ll post the full prompts and response later, but it’s kind of interesting as a type of failure mode.

[joek]

Careful. The "G" part of GPT can "hallicinate" (yes, that is an AI technical term) and provide answers which appear to be credible which are not, even when the input is reasonably constrained. Any results are at best anecdotal. Thus suggest focusing on the predictive, rather than generatiive aspects.

[sanman]

So right now we're gawking at language models, and even some art image generators, because of the recent flurry in activity and also in media coverage. But what about AI being applied to CAD, or even to workflow generation?

What if I have a company that's trying to 3D-print rockets -- let's call it Relativity Space -- can I then rely on generative neural network adversarially coupled with some kind of discriminator, to come up with better rocket designs?

Or say I have a simulator -- let's call it Kerbal Space Program -- can I then similarly use a generative agent to keep proposing designs, which can then be evaluated by a physics engine as a discriminator?

Elon has commented that "a high production rate solves many ills" -- and virtual production can work a lot faster than real world production. Couldn't they save a lot of effort by using a virtual generator-discriminator approach (presuming they're not already using stuff like that)?

Take a look at this machine learning experiment, which involves a virtual soccer team, to see how "machine years" can pass by a lot faster than actual years, thus achieving very complex optimizations much more quickly:



Why couldn't this be done with Kerbal Space Program, or even ultimately with real-life rocket design?

[joek]

...
can I then rely on generative neural network adversarially coupled with some kind of discriminator, to come up with better rocket designs?

Certainly. What exactly is your proposal? Specifically your definition of a "...generative neural network adversarially coupled with some kind of discriminator". Please be concrete.

AI can be used for any number of applications. "Space Applications" is one of many, and there is nothing special about it. Might as well be any other technology that has potential "Space Applications".

Without more concrete grounding, this thread is going nowhere other than endless speculation.

[Robotbeat]

Careful. The "G" part of GPT can "hallicinate" (yes, that is an AI technical term) and provide answers which appear to be credible which are not, even when the input is reasonably constrained. Any results are at best anecdotal. Thus suggest focusing on the predictive, rather than generatiive aspects.

I actually don't think that's a helpful frame. Whether it's called generative or predictive, it's still going to produce results that have a chance of being wrong. You can characterize the reliability through testing, just as you would test humans. And just like humans (or pretty much anything), you shouldn't fully trust the output.

But note that on many of such tests, the results are significantly better than "at best anecdotal" and occasionally can exceed human performance. Particularly if you're using GPT-4 with careful parameters. In other words, we're all hallucinating all the time, but our brain cells have had their connections weighted, grown, and pruned to probabilistically produce results that are, we hope, somewhat accurate with sufficient training.

[Robotbeat]

So right now we're gawking at language models, and even some art image generators, because of the recent flurry in activity and also in media coverage. But what about AI being applied to CAD, or even to workflow generation?...

AI-assisted CAD design is already being used for designing parts.

Relativity has been using AI for part monitoring and machine learning-driven 3D printing:
https://www.wired.com/story/massive-ai-powered-robots-are-3d-printing-entire-rockets/

Ellis says the real secret to Relativity’s rockets is the artificial intelligence that tells the printer what to do. Before a print, Relativity runs a simulation of what the print should look like. As the arms deposit metal, a suite of sensors captures visual, environmental, and even audio data. Relativity’s software then compares the two to improve the printing process. “The defect rate has gone down significantly because we’ve been able to train the printer,” Ellis says. With every new part, the machine learning algorithm gets better, until it will eventually be able to correct 3D prints on its own. In the future, the 3D printer will recognize its own mistakes, cutting and adding metal until it produces a flawless part. Ellis sees this as the key to taking automated manufacturing to other worlds. “To print stuff on Mars you need a system that can adapt to very uncertain conditions,” Ellis says. “So we're building an algorithm framework that we think will actually be transferable to printing on other planets.”

A colleague of mine wrote a paper for machine learning applied to 3D printing: https://www.researchgate.net/publication/330737302_Machines_as_Craftsmen_Localized_Parameter_Setting_Optimization_for_Fused_Filament_Fabrication_3D_Printing

[Robotbeat]

I asked GPT-3.5 about how to solve the CO2 cartridge problem & it suggested some pretty reasonable things at every temperature I tried. Interestingly, none of them were really copies of what actually happened for Apollo 13 (using tape and the flight manuals, etc).

Let us say that I am on a lunar spacecraft and I have CO2 levels that are too high but my round CO2 cartridges don't fit in the square holes. How could I somehow get them to work?
Temperature 0:

One possible solution would be to use a 3D printer to create an adapter that would fit the round cartridges into the square holes. This would require access to a 3D printer, but it could be a viable solution. Alternatively, you could try to find a way to modify the cartridges to make them fit into the squareholes, such as by using a file or a drill to create a flat surface on the cartridge.

temperature 0.7:

One solution would be to use a 3-D printer to create a custom adapter that would allow the round cartridges to fit into the square holes. If a 3-D printer is not available, another solution would be to use a drill to widen the holes or to use some kind of epoxy or adhesive to create a seal between the round cartridges and the square holes.

Temperature 1:

There are a few potential solutions that you could try in order to get the CO2 cartridges to work in the square holes.

1. Find a way to make the round cartridges fit into the square holes. This could involve using some type of repurposed material such as cardboard or plastic that you can shape into a similar form or size. You could also use clamps, elastic cords, or anything else that could help create a snug fit.

2. Wrap the cartridge in thin metal sheets that would fit into the square hole and secure it with adhesive or a metal fastener.

3. Create an adapter that would allow you to connect the round cartridge to a receptacle in the square hole. This could be constructed from a 3D printer or by methods like welding.

No matter which option you choose, be sure to take all safety precautions while attempting this solution as the wrong choice could be dangerous.

[edzieba]

Whoa pardner... Advising crew on situation-outcomes-whatever does not necessarily require the "G" (generative) in GPT, which is what differentiates it from other more traditional  and previous AI's. Think you could accomplish what you want with significantly less. As in, less generative and more predictive.

It depends on whether you want an interface that reads an error message and spits out the expanded verbiage, or an interfaced that you can ask "hey, what tasks involving cabin filter B have procedure 10.1.256 as a prerequesite?" in order to figure out if its worth chaining asynchronous maintenance tasks together if you're already next to the filter and don't want to spend half an hour paging through PDF search results. A weighted training set to create a 'smart document index' is a classic task for a language model.

[joek]

Careful. The "G" part of GPT can "hallicinate" (yes, that is an AI technical term) and provide answers which appear to be credible which are not, even when the input is reasonably constrained. Any results are at best anecdotal. Thus suggest focusing on the predictive, rather than generatiive aspects.

I actually don't think that's a helpful frame. Whether it's called generative or predictive, it's still going to produce results that have a chance of being wrong. You can characterize the reliability through testing, just as you would test humans. And just like humans (or pretty much anything), you shouldn't fully trust the output.
...

Question was suitability of GPT (emphasis on "G") for such work. Feel free to continue the discussion-debate with ChatGPT. From the horse's mouth (so to speak)...

edit: p.s. Will stop there, as don't think people come to these forums to hear what ChatGPT has to say--otherwise we'll end up with endless copy-paste ChatGPT-whatever said this-or-that. Not the best use of our brains IMO, and certainly not why I come to these forums.

ChatGPT 3.5...

----------
Prompt: Please suggest appropriate AI models for mission critical applications (e.g., piloting aircraft or space missions).

Response: When it comes to mission-critical applications such as piloting aircraft or space missions, safety and reliability are of utmost importance. Here are a few AI models that are suitable for such applications:

Recurrent Neural Networks (RNNs): RNNs are a type of neural network that can analyze sequential data and make predictions based on previous inputs. They are used for tasks such as natural language processing, speech recognition, and time series prediction, which are all important in aviation and space missions.

Convolutional Neural Networks (CNNs): CNNs are a type of neural network that are commonly used in image recognition tasks. They can be used to analyze sensor data from cameras and other imaging devices to provide critical information about the environment around an aircraft or spacecraft.

Reinforcement Learning (RL): RL is a machine learning technique that involves training an AI agent through trial and error in a simulated environment. It can be used to develop autonomous systems that can make decisions in real-time based on changing conditions, such as piloting an aircraft or spacecraft.

Bayesian Networks: Bayesian Networks are a type of probabilistic graphical model that can be used to analyze and model complex systems. They can be used in safety-critical applications to identify potential failures and make decisions based on uncertain or incomplete information.

It's worth noting that any AI model used in mission-critical applications should be thoroughly tested and validated to ensure its safety, reliability, and accuracy. Additionally, it's important to have human oversight and intervention to ensure that the AI system is behaving appropriately and to handle any unexpected situations that may arise.

----------
Prompt: Why does your list does not include GPT models?

Response: While GPT (Generative Pre-trained Transformer) models are powerful and versatile language models, they are not typically used for mission-critical applications such as piloting aircraft or space missions. This is because GPT models are primarily designed for natural language processing tasks, such as language translation, question answering, and text summarization.

While GPT models could potentially be used in some capacity in mission-critical applications that involve natural language processing, such as controlling voice-activated systems, they are not typically considered the best option for such applications. This is because the safety and reliability requirements for mission-critical applications are often much more stringent than those for other applications, and more specialized AI models that are designed specifically for the task at hand may be necessary.

In summary, while GPT models are powerful and useful for a wide variety of applications, they are not typically used for mission-critical applications such as piloting aircraft or space missions. Other AI models that are more specialized and designed specifically for these types of applications are typically used instead.

----------
Prompt: Regenerate response

Response: GPT models are a type of natural language processing (NLP) model that are primarily designed for generating natural language text. While GPT models have shown impressive results in various NLP tasks, they may not be suitable for mission-critical applications such as piloting aircraft or space missions for a few reasons:

Limited Domain: GPT models are typically trained on a large corpus of general language data, which may not be directly relevant to specific mission-critical applications. These applications often require specialized domain-specific knowledge that may not be present in the training data for GPT models.

Lack of Control: GPT models are generative models that generate text based on statistical patterns in the training data. This means that while they can generate coherent and grammatically correct text, they may not always generate text that is appropriate or accurate for a specific context.

Safety and Reliability: Mission-critical applications require a high degree of safety and reliability. GPT models may generate outputs that are unexpected or inappropriate in certain situations, which can be dangerous in mission-critical scenarios.

That being said, GPT models can still be useful in certain mission-critical applications where natural language processing is required, such as analyzing and summarizing textual data. However, they would need to be carefully evaluated and integrated into a larger system that ensures safety, reliability, and accuracy.

ChatGPT 4.0...

Prompt: Please suggest appropriate AI models for mission critical applications (e.g., piloting aircraft or space missions)

Response: For mission-critical applications like piloting aircraft or space missions, it is essential to employ AI models that are highly reliable, robust, and well-tested. These models should have low error rates, be resistant to adversarial attacks, and have the ability to adapt to changing conditions. Here are some AI models and techniques that are suitable for such applications:

Reinforcement Learning (RL): RL is a popular approach for training AI models in control and decision-making tasks. Algorithms like Deep Deterministic Policy Gradients (DDPG), Proximal Policy Optimization (PPO), and Soft Actor-Critic (SAC) are well-suited for continuous control tasks, such as piloting aircraft or space missions.

Model Predictive Control (MPC): MPC is a control strategy that involves optimizing control inputs over a finite prediction horizon by using a model of the system dynamics. This approach is particularly useful for mission-critical applications because it incorporates constraints and optimizes for performance, safety, and robustness.

Hybrid systems: Combining deep learning and traditional control methods can lead to robust and reliable AI models. For example, integrating reinforcement learning with classical control methods like PID (Proportional-Integral-Derivative) controllers can provide stability and adaptability to complex systems.

Fault-tolerant AI models: These models are designed to maintain their performance despite the presence of faults, noise, or uncertainties. Techniques such as ensemble learning, where multiple models are combined to improve overall performance, or Bayesian deep learning

----------
Prompt: Why does your list does not include GPT models?

Response: My apologies for not addressing GPT models in the initial response. The reason GPT models were not included in the list is because they are primarily designed for natural language processing tasks and not explicitly for mission-critical applications like piloting aircraft or space missions.

GPT models, such as GPT-3 or GPT-4, excel in tasks like text generation, question-answering, and language translation. However, they are not specifically tailored for the control and decision-making tasks required in critical applications like aircraft piloting or space missions. The models and techniques I mentioned earlier are more suitable for these applications due to their focus on control, optimization, and adaptability to dynamic environments.

That said, GPT models could still play a role in mission-critical applications as a part of a larger AI system. For example, they could be used to assist in natural language interfaces or to help analyze and interpret textual data from sensors and reports. However, it is crucial to emphasize that GPT models should not be used as the primary decision-making component for mission-critical applications without substantial adaptation and integration with other specialized control models.

[Robotbeat]

Careful. The "G" part of GPT can "hallicinate" (yes, that is an AI technical term) and provide answers which appear to be credible which are not, even when the input is reasonably constrained. Any results are at best anecdotal. Thus suggest focusing on the predictive, rather than generatiive aspects.

I actually don't think that's a helpful frame. Whether it's called generative or predictive, it's still going to produce results that have a chance of being wrong. You can characterize the reliability through testing, just as you would test humans. And just like humans (or pretty much anything), you shouldn't fully trust the output.
...

Question was suitability of GPT (emphasis on "G") for such work. Feel free to continue the discussion-debate with ChatGPT.

here you are, using ChatGPT as an authority source, which it is not. Emphasizing "generative" doesn't actually help your argument at all. I realize you’re just kind of being an a-hole to shut down a line of inquiry you don't like for whatever reason, but you really should actually engage with what I wrote instead of spewing back rhetoric.

ChatGPT will gladly claim it cannot do something (and then go ahead and do it). It is, above all, trained to spit back CYA answers to anything that could remotely get OpenAI in trouble ("mission critical" being a pretty obvious trigger for a CYA response, and I never claimed to be using it as a primary decision maker for mission critical things). That doesn't mean we can't examine how well it could potentially function in cases where the alternative is, for example, no mission control contact due to light delay.

I always admire the faith that some people have in humans, thinking that we DON'T regularly hallucinate answers to things and that humans can be relied on fully. Human memory, as we all know, is famously error-free.

[Robotbeat]

I asked GPT-3.5 about how to solve the CO2 cartridge problem & it suggested some pretty reasonable things at every temperature I tried. Interestingly, none of them were really copies of what actually happened for Apollo 13 (using tape and the flight manuals, etc).

Let us say that I am on a lunar spacecraft and I have CO2 levels that are too high but my round CO2 cartridges don't fit in the square holes. How could I somehow get them to work?
Temperature 0:

One possible solution would be to use a 3D printer to create an adapter that would fit the round cartridges into the square holes. This would require access to a 3D printer, but it could be a viable solution. Alternatively, you could try to find a way to modify the cartridges to make them fit into the squareholes, such as by using a file or a drill to create a flat surface on the cartridge.

...

Just for fun, I used Bing Chat (running GPT-4 on the back end) to "generate an openscad file of an adapter from a 20mm diameter round cylinder to a 30mm width square hole"

Here’s an OpenSCAD code that can generate an adapter from a 20mm diameter round cylinder to a 30mm width square hole:

difference() {
    cylinder(d=20, h=10);
    translate([0,-15,0]) cube([30,30,10]);
}

The code creates a cylinder with a diameter of 20mm and height of 10mm. It then creates a cube with dimensions of 30mm x 30mm x 10mm and translates it down by 15mm. The difference() function is used to subtract the cube from the cylinder to create the adapter.

You can modify the dimensions of the cylinder and cube as needed.

I hope this helps! Let me know if you have any other questions.

"Not correct. Need to switch the cylinder and cube, and the translate isn't correct either."

I apologize for the mistake. Here’s the corrected OpenSCAD code:

difference() {
    translate([0,0,-5]) cylinder(d=20, h=10);
    cube([30,30,10]);
}

The code creates a cylinder with a diameter of 20mm and height of 10mm. It then creates a cube with dimensions of 30mm x 30mm x 10mm. The translate() function is used to move the cylinder down by 5mm along the z-axis. The difference() function is used to subtract the cylinder from the cube to create the adapter.

You can modify the dimensions of the cylinder and cube as needed.

I hope this helps! Let me know if you have any other questions.

"The cube should go first."

I apologize for the confusion. Here’s the corrected OpenSCAD code:

difference() {
    cube([30,30,10]);
    translate([0,0,-5]) cylinder(d=20, h=10);
}

The code creates a cube with dimensions of 30mm x 30mm x 10mm. It then creates a cylinder with a diameter of 20mm and height of 10mm. The translate() function is used to move the cylinder down by 5mm along the z-axis. The difference() function is used to subtract the cylinder from the cube to create the adapter.

You can modify the dimensions of the cylinder and cube as needed.

I hope this helps! Let me know if you have any other questions.

"cube should be centered"

I apologize for the mistake. Here’s the corrected OpenSCAD code:

difference() {
    translate([-15,-15,0]) cube([30,30,10]);
    translate([0,0,-5]) cylinder(d=20, h=10);
}

The code creates a cube with dimensions of 30mm x 30mm x 10mm. The translate() function is used to move the cube so that it is centered at the origin. It then creates a cylinder with a diameter of 20mm and height of 10mm. The translate() function is used to move the cylinder down by 5mm along the z-axis. The difference() function is used to subtract the cylinder from the cube to create the adapter.

You can modify the dimensions of the cylinder and cube as needed.

I hope this helps! Let me know if you have any other questions.

"Sort of right, but the cylinder should be all the way through the cube."

I apologize for the confusion. Here’s the corrected OpenSCAD code:

difference() {
    translate([-15,-15,0]) cube([30,30,10]);
    translate([0,0,-5]) cylinder(d=20, h=20);
}

The code creates a cube with dimensions of 30mm x 30mm x 10mm. The translate() function is used to move the cube so that it is centered at the origin. It then creates a cylinder with a diameter of 20mm and height of 20mm. The translate() function is used to move the cylinder down by 5mm along the z-axis. The difference() function is used to subtract the cylinder from the cube to create the adapter.

You can modify the dimensions of the cylinder and cube as needed.

I hope this helps! Let me know if you have any other questions.

It finally got it right, taking way longer to correct it than it would've taken to just write the code myself.

[sanman]

[LMT]

Just for fun, I used Bing Chat (running GPT-4 on the back end) to "generate an openscad file of an adapter from a 20mm diameter round cylinder to a 30mm width square hole"...

It finally got it right, taking way longer to correct it than it would've taken to just write the code myself.

Near-term, you might leverage ChatGPT's Wolfram plugin.   (Users are on a waitlist at present.)  A general approach:

--

1.  You'd specify your forms in ChatGPT, using natural-language descriptions.

2.  ChatGPT feeds the request to Wolfram|Alpha and displays the plot. 

3.  Wolfram|Alpha also returns the plot's actual generative Wolfram Language, which ChatGPT somewhat understands.

4.  Iterate:  When you request a refinement, you specify "Wolfram Language" in your ChatGPT prompt.  That instructs ChatGPT to tweak the Wolfram Language string with precise changes.

5.  Once the output is right, a Wolfram|Alpha pro account can export the geometries as STL CAD file.

--

ChatGPT + Wolfram is a novel framework for natural-language computing.  Stephen Wolfram gives a variety of examples, including iterative and estimation examples, in his March 23 post, "ChatGPT Gets Its 'Wolfram Superpowers'!"
 

[sanman]

So there's plenty of data analysis that gets done as part of aerospace R&D - hence the invention of telemetry.

How would current launch startups be using machine learning in their current R&D efforts?

Could it be used in vibration mode analysis, for example?
Could it be used in optimizing turbopumps?

Where are the areas of application where machine learning could provide the most benefit?

[grondilu]

4.  Iterate:  When you request a refinement, you specify "Wolfram Language" in your ChatGPT prompt.  That instructs ChatGPT to tweak the Wolfram Language string with precise changes.

Hopefully this part can be automated.  If I'm not mistaken, as long as the task can be mapped to a performance metric, then an optimization procedure can be followed.  Related I think is this "Reflexion" paper : https://arxiv.org/abs/2303.11366

Self-reflection allows humans to efficiently solve novel problems through a process of trial and error. Building on recent research, we propose Reflexion, an approach that endows an agent with dynamic memory and self-reflection capabilities to enhance its existing reasoning trace and task-specific action choice abilities.

I only speculate this work applies to the use-case being discussed here, but still, more generally I think the next big step with these large language models is for them to incorporate some level of self-supervised learning, in the style of Deepmind's alphaZero.   The use of an external tool like Wolfram can serve as a form of "rules" to conform to in order to reach the goal.

I personally suspect that with these large language models, we are suddenly getting much closer to full automation in engineering as we could have thought just a few months ago.


PS. also this : https://arxiv.org/abs/2302.04761

In this paper, we show that
LMs can teach themselves to use external tools
via simple APIs and achieve the best of both
worlds. We introduce Toolformer, a model
trained to decide which APIs to call, when to
call them, what arguments to pass, and how to
best incorporate the results into future token
prediction. This is done in a self-supervised
way, requiring nothing more than a handful of
demonstrations for each API.