Works like this deserve a lot more attention. In embodied AI, the most common argument against physics-based simulation is that deformables are hard/expensive to simulate. Offset Geometric Contact just made it 300x faster. It computes vertex-specific displacement bounds to guarantee penetration-free simulation. To scalably simulate the world, start with physics instead of pixels.

Oct 12, 2025 · 4:41 PM UTC

Replying to @sunfanyun
I love this paper, but I doubt truly scaleable world sim will start w/ physics instead of pixels. Simply because that’s not how human NN’s operate. But open to being wrong here.
4
9
Evolution shaped humans, and that's an important part of the physics prior we carry
1
1
14
Replying to @sunfanyun
I thought the problem was solved with protein folding.
Replying to @sunfanyun
I bought the first 3DFX card and i can't believe that tying a knot inside a game is now finally solved, only 28 years later.
8
Replying to @sunfanyun
Thank you for post!
1
5
Replying to @sunfanyun
Except this is not physics, it’s a surrogate model deaigned to look real and support fast computations.
3
Replying to @sunfanyun
Finally ffs. As a kid I assumed the ps3 would have physics simulation. Software laws modeled off of Newtonian physics.
3
Replying to @sunfanyun
I could do this about 19 years ago on the CPUs/GPUs of that time… I would be surprised if it wasn’t possible with the hardware we have now
2
Replying to @sunfanyun
stimulating the physics is the only way to make biological breakthroughs
2
Replying to @sunfanyun
This is huge for embodied AI! ⚡ Making deformable simulations 300x faster removes a massive bottleneck—physics-first approaches are finally catching up to pixel-first shortcuts. 🚀
1
Replying to @sunfanyun
I agree, when I talk about Kung Fu Cybernetics, I’m talking at this level of electromagnetic and innervation layers. Something only a Cyberneticist would recognize, and only a martial arts expert would know where to look thats also a computer scientist with sigint understanding
Replying to @sunfanyun
or no pixels at all.
Replying to @sunfanyun
I'm sure some dev out there thinks it's worth 187518754917854 dollars lol
Replying to @sunfanyun
이건 진짜 월드빌드 판 뒤집는 기술이네
Replying to @sunfanyun
這技術聽起來很厲害,對模擬真實環境應該有很大幫助!
Replying to @sunfanyun
Another also very strong argument against sim is the system identification (real2sim) problem for deformables in the general sense is a nightmare.
Replying to @sunfanyun
Hey… you might be on the missing part of restabalization (putting back together) the broken down data code for quantum teleportation. -you are on the brink of being the idea that makes teleporting of humans & matter REAL… God bless you & Lee going!! We need this ❤️❤️
Replying to @sunfanyun
Perhaps this can help NCAP crash test cars without crash testing them.
Replying to @sunfanyun
That's beautiful
Replying to @sunfanyun
Game changing!
Replying to @sunfanyun
ten years folding clothes, i toiled through the night one day a bright figure asked me for a light he said boy listen up your folding work is nice but this world that you’ve been living in well something isnt quite right he showed me on a computer screen and along side thousands more like me he said I have been “training” and that I leave tomorrow for the real world. I asked him if I could bring my cat, the one that gave me company while I folded he said, “cat?!” -“ghastly, what on hells earth is going on, that’s not supposed to be here. You said he keeps you company?? That will never do. no no that was not part of the goal” he scanned the number on my arm and said “delete”