GPU-based physics simulator speeds up reality by "1,000x" while GPT-4 calls the shots.
See full article...
See full article...
Eureka uses GPT-4 and massively parallel simulations to accelerate robot training
- Thread starter JournalBot
- Start date
Upvote
80
(84
/
-4)
Oh no worries that functionality is reserved for the battlefield models. You know - that battlefield far, far away.
Upvote
36
(38
/
-2)
GPU-based physics simulator speeds up reality by "1,000x"
They need to stop this immediately! Life is already too short at normal reality speed!
Upvote
47
(49
/
-2)
I'm confused - did a real robot actually perform the task in the end, or is this entirely virtual?
Upvote
48
(49
/
-1)
Nvidia Isaac Gym is pure simulation. So getting it done on real hardware is 90% done, there's only 90% remaining.
Upvote
129
(129
/
0)
Does it invent fake laws of physics like it invented fake lawsuits for legal cases?
Upvote
11
(32
/
-21)
Nvidia is using their PhysX model for physics, so comes with all those limitations. I think this use of LLMs is brilliant though, it isolates it to a higher tier and continuously refines itself. Mitigates a lot of, if not all of, the hallucination concerns.
Upvote
34
(37
/
-3)
Coming soon…the T-100!
It walks, it talks, it’s learning to kill!
Pre order TODAY!
It walks, it talks, it’s learning to kill!
Pre order TODAY!
Upvote
25
(28
/
-3)
Agreed, in my opinion this is what "good AI/good use of AI" looks like. In my mind it parallels writing things in assembly vs. a higher level language: you still can if you want to or need to, but this abstraction is vastly more efficient the vast majority of the time.
Upvote
43
(44
/
-1)
Ahhhhh... came for the world ending comments and was not disappointed! Well done Arsians!
Upvote
8
(12
/
-4)
As amazing as our bodies are in comparison to crude mechanical imitations like atlas or spot, the software running it plays a colossal part, just look how amputees work around their limited hardware, similarly, an actually smart bot should be able to do wonders with a wimpy wobbly limb
this is expected but really good news nonetheless
this is expected but really good news nonetheless
Upvote
4
(4
/
0)
Meh. Wake me up when the XXX-100 hits the market. That's when I'll preorder.
Upvote
6
(8
/
-2)
Upvote
12
(12
/
0)
Upvote
32
(32
/
0)
I'm still not entirely clear what input gpt4 was reviewing and what output it was providing? In the chart it seems it is given a set of weights that the model used, and it's output is code embodying new weights, but this doesn't make a ton of sense to me.. Does anyone have a more detailed understanding of what's going on in the refinement loop with the AI?
Upvote
25
(25
/
0)
Yeah, but they don't really turn me on. I need a sex bot that can write Elizabethan sonnets, discuss metaphysics, and compare and contrast Marxian dialectics with Hegelian dialectics.
Upvote
31
(32
/
-1)
When's Nvidia gonna release the video showing the hand tricks that robot learned?
Upvote
23
(23
/
0)
This reads like a robot version of when neo got plugged in, and said "I know Kung-Fu"
Upvote
16
(16
/
0)
How much noise do they add to these simulations? Is every input and output response accurate to ten decimal places or do they randomly ± this to better simulate reality?
Upvote
13
(13
/
0)
Upvote
15
(16
/
-1)
Upvote
16
(17
/
-1)
When it no longer accidentally rips off important body parts...
Upvote
15
(15
/
0)
You can train in virtual and then do transfer to the actual hardware. Isaac (the simulator) has 'Sim-to-Real',
https://www.nvidia.com/en-us/on-demand/session/gtcspring21-s31824/
Hook it up with ChatGPT, can do all three at once
Doesn't exactly nail any of the three though - was fun to see what trying to mix them would give.
Upvote
15
(16
/
-1)
If I'm reading it right, the innovation was using a AI system to write and refine the actual goal that evaluated each simulation, which can be surprisingly hard.
Let's say you were trying to tiger-parent a two year old to go to the olympics:
Bad goal: "Billy run a marathon and I'll score a 1 if you make it and a 0 if you don't"
Better goal: "Billy can you run across the room?"
So I'd guess they used an AI system to continuously give the control system hard-but-achievable goals, and made them juuuust hard enough that it could keep improving rapidly.
Upvote
37
(37
/
0)
Oh, it's a hand! From the top picture, I thought it was a little white monkey soldier trying to carry a stick.
Upvote
32
(32
/
0)
Upvote
2
(2
/
0)
You're just a little late....
I sold these as fast as I could get them from TEW (Tandy Electronics Warehouse) back in 82-84.
Used my store demo for our bowling league with a scoring program we wrote. Took a 40 column thermal printer, and handed our opponents a match sheet in 2 minutes every week.
They were pretty neat.
Upvote
5
(5
/
0)
Something something Skynet began to learn at an exponentially accelerating rate something something
yeah yeah
Upvote
16
(17
/
-1)
Hah! I'm curious how much fidelity the Isaac models have. Other software-based "physics learning" AIs I've seen often find a way to cheat the system with a glitch or unintended use of a feature to meet the training goal. Hilarity ensues, at least for me since I'm not the one doing the work.
View: https://www.youtube.com/watch?v=Lu56xVlZ40M&t=143s
Upvote
15
(15
/
0)
Don't worry guys, if you see a chatgpt robot approach you with knives in its robot hands just say, in all caps, DO NOT UNDER ANY CIRCUMSTANCES STAB ME WITH HEY WHAT ARE YOU STOP OH GOD AAARGH
Upvote
10
(10
/
0)
This is where this approach could also help a lot, though. They don't train a neural network to do a set task, they have the LLM iteratively find a reward function that leads to quick and correct training of a new network. Presumably that reward function would transfer to real hardware a whole lot better than a trained network or starting from training scratch. Over time, you might even integrate stuff learned from real-world transfer, like "Make sure the action doesn't rely on sub-millisecond precision".
Upvote
8
(8
/
0)
Honestly I'm a bit surprised that layering models like this is new.. I assumed that most of the more advanced ML applications were doing some kind of layering. Human intelligence isn't really one calculation being shoved through a set of neurons (if I understand correctly). There are numerous smaller systems that work together to come up with a more coherent output... usually. 
Upvote
8
(8
/
0)