The point of humanoids beyond the hype is so they can integrate into existing infrastructure - like a commercial kitchen. So the question is can they cook an omelet and peel a potato. Much harder than kickboxing in a YouTube video.
Physical AI
- Thread starter JeffKeryk
- Start date
Why do you want to do that when there's already a version of "KitchenAid" machine in commercial Kitchen in Japan?
Was one example. In theory same design could do thousands of things just different software.
MolaKule
Staff member
Physical simulations have to be embodied in Physical AI entities.
NVIDIA: "To build physical AI, teams need powerful, physics-based simulations that provide a safe, controlled environment for training autonomous machines. This not only enhances the efficiency and accuracy of robots in performing complex tasks but also facilitates more natural interactions between humans and machines, improving accessibility and functionality in real-world applications."
An understanding of the laws of physics has to be implemented by human beings before it can be transferred to PAI.
Deleted.
Last edited:
Sure, and robotics will need to improve in so many ways, especially in uncertain, non-structured envirornments. But that is always true, humans included. In time, Physical AI is expected to exceed human capabilities in certain areas. Otherwise why do it? Personally I am convinced we will get there.
An understanding of the laws of physics has to be implemented by human beings before it can be transferred to PAI.
Regarding the laws of physics, AI will greatly further our understanding of the physical Universe via its capability to examine vast amounts of complex data and glean associations, etc. via complex algotithms.
I am not sure many humans can correctly add up 1M numbers in a short period of time. AI will examine 1B stars and reveal untold truths. No group of humans have that capability, at least in a relatively short period of time. Or perhaps ever. Then consider yottabytes of data...
Artificial intelligence is an essential tool. In fact, it is the only way to further disciplines like Astronomy. Human capabilities are limited.
You may be interested in my post on the Rubin Observatory development lecture by Prof. Kahn. Just a suggestion.
Last edited:
Based on what I have seen (presentation with NDA), they are using statistics instead of deterministic based calculation for that. This makes sense for their system as AI is not arithmetic based calculation but statistical training based.
Physical simulations have to be embodied in Physical AI entities.
NVIDIA: "To build physical AI, teams need powerful, physics-based simulations that provide a safe, controlled environment for training autonomous machines. This not only enhances the efficiency and accuracy of robots in performing complex tasks but also facilitates more natural interactions between humans and machines, improving accessibility and functionality in real-world applications."
An understanding of the laws of physics has to be implemented by human beings before it can be transferred to PAI.
The reason for this is for a simple system you can run a lot of known calculation to do simulation dot by dot. However for a large system you cannot do enough calculation by the trillions and you have to rely on an overall input and overall output based statistical distribution to do educated guess on what the outcome may be.
What "AI" is doing beyond the typical Monte Carlo / Random Walk based system is 1) breaking down system into linear components that you can multiply and add to each other (like a spring is just a metal chain with elastic joint), 2) simulate the behavior of the system based on previous training from a large sample (how dropping a metal spring looks from prior work). 3) Create "digital twins" of the system in AI, compare the educated guess against the actual working system. 4) train and refine the digital twins.
So, no they don't actually do the arithmetics but rather use statistic to do the math because it is simpler. You can think of it as having an artist watch how other architect design building, then draw the same design by mimicking, and let the architect review it and approve it as a functional design, all without learning how the math and physics work.
Or you can think of it as how a 3 year old learn how to ride a bike, without knowing how all the control loop and servo control work, and just trial and error and watching how people do it.
I think we are talking about discrete math vs AI both done by computers. We probably can agree that human cannot do these math by hands in scale.
Fundamentally underneath the hardware they are all multiplies and adds, just how we group things together and how we apply them turning them into "AI" vs "SIMD" vs "DSP". Marketing loves new names.
MolaKule
Staff member
The laws of physics were developed over hundreds of years of hypothesis, testing, and observation.
Once that data is gathered, a human will have to interpret the data to determine if any new enlightenment is present.
Finite Element Analysis (FEA) is an essential automation tool in which to analyze structures such as automobile and aircraft structures. I seriously doubt a single engineer could place a proper grid by hand to fully analyze structures. See page 34:
https://apps.dtic.mil/sti/tr/pdf/ADA173823.pdf and
https://simutechgroup.com/services/...fea/?msclkid=d133062e66c3141e727b6ec4c48f763e
AI is just another tool to automate processes.
I have studied the Rubin Observatory. It is an advanced optical and sensor earth-based observatory, nothing more. Why they didn't launch a space-based version, similar to JWST, is beyond understanding.
Last edited:
MolaKule
Staff member
That's my understanding of the AI algorithms as well. A good explanation. Why resimulate the physical phenomena when it has already been done; just gobble up the results and statistically analyze all of the available data at your disposal.
That could be a problem. What we found in the aerospace industry is unless one understands the details of Stress and Fatigue, his results may not be reliable and could not be validated.
Why "resimulate"? The answer I was told is a trade off between sample size and how you come up with the value of each sample. There's a believe that if you can simulate 1B times you can eventually reach the law of large number, and your values will fit a distribution that despite you having no official way to prove it, and could be a better result than 10k real number calculated with accurate real experiment, so you trust it. This is actually how a lot of people use statistics instead of deterministic way to figure out what is happening in real life. You can call it gut instinct, experience, trust, smell test, etc. You are not going to convince someone with a peer review with this but it is good enough if you have a human expert who you can count on to review it. You are not going to prove something scientifically this way, but you can probably form a tradition that others (robots) to follow because we have always done it this way. I guess you can call it a, culture?
So in a way you can say AI for self driving or robot doing your work is by observing what you do, and copy after watching you for hours, instead of reading an instruction and then apply tools work using the instruction line by line. You probably can trust it, but you as the owner is on the hook for check if it is done right.
MolaKule
Staff member
I think you misinterpreted this rhetorical question.
The point I was positing was this: There is no need to "resimulate" physics results internally within an AI module if the simulation results are already out there for the AI module to ingest.
AI is simply another knowledge tool and like an ax, it can be used to build useful things, or it can be used for nefarious purposes.
BTW, AI can be hacked such that AI results spew out gibberish. Use with caution.
True if you are doing the same thing. But typically what they use it for is to try new things half way between what we already have results for, or a different combination of things.
You can do everything you do in AI with a good software engineer and enough time, and enough computational power. AI in theory can save you time with some educated guess, that's basically what it is.
AI spew out gibberish just like an elementary school student pretending to do calculus and quantum mechanics after watching sci fi, their training is not quite there yet.
Similar threads
- Locked
