This Isn’t AI - It’s Even Wilder: Squishy Physics That Learn to Move!

This is a new squishy simulation  technique and it’s a bit like   teaching a pile of Jell-O how to win  an Olympic gold medal in gymnastics. Yes! So what if a desk lamp could  learn gymnastics? You know, jump,   twist, stick the landing - all with real  physics! That is really hard. Why? Well,   you see, most characters you meet in video  games have a bone and joint structure,   which determines how they can move around. But  jellyfish, worms, or a stress ball don’t have   skeletons. They move by squishing, stretching,  and contracting their own soft bodies. To animate that, you need to figure  out a series of muscle contractions   and relaxations that make physical  sense. Simulating that is brutally hard:   thousands of interacting parts, collisions,  friction, and no neat closed-form equations. So much so that old gradient descent-based  methods, with the blue and orange lines fail   to launch this ball into the hoop. Now hold on  to your papers Fellow Scholars, because this   new method can do it every single time. And  the results are wild: starfish now can crawl   realistically, gummy caterpillars wriggle forward  with incredible realism, lamps are doing backflips   on trampolines, even chess pieces can hop  across the board. Or if you are in a library,   you can also have a little personal butler. I  absolutely love this one! Amazing! This level   of control was simply not possible with older  first-order methods like gradient descent. But that’s not even the best part.   I’ll show that to you in a moment. So, how is all this insanity even possible?   And how long does it take to compute? Dear   Fellow Scholars, this is Two Minute Papers  with Dr. Károly Zsolnai-Fehér. Dr. Carroll. Imagine you are hiking in fog. Gradient descent  is feeling the local slope with your boots and   taking cautious steps. But Newton’s method  is different - it not only feels the slope,   it also senses how the ground curves, so  instead of just stepping, it can leap to   the right spot. Much quicker! The trick here  is: how do we measure that curvature quickly   and accurately in a challenging soft-body world  with contacts and friction? That is super tough. And in this paper, scientists fuse two superpowers  together: automatic differentiation for precise   slopes, and a clever complex-numbers probe that  sneaks a microscopic step into an imaginary   direction to read curvature cleanly. Crazy  mad science, I love it. They call it mixed   second-order differentiation - which gives  you the information needed for true Newton   updates. Now the optimizer has a map and  a compass, not just a feeling underfoot. And now, here is the best part.   My favorite experiment. Oh yes,   it is not hard to see that soon, even  computer games based on this kind of   physics will be possible. Man, I  cannot wait to see and play that! But based on my calculations, this still  takes quite a while to compute - 10 to 25   minutes for one second of movement  are typical. So not real-time yet,   but absolutely within reach for movies  and, two more papers down the line,   probably video games. And don’t forget,  this makes the impossible now possible.    Now we only need to make it faster for games.   And man, I cannot wait to see and play that!    And here is the other best part: the paper is  not from Disney, so I can even talk about it! And with this amazing paper, we are  witnessing the birth of a new kind of   animation - where soft, squishy worlds are  no longer puppeteered but are truly moving   realistically. This isn’t just eye candy -  it’s a glimpse of creatures that move with   the richness of real life. Absolutely  brilliant. What a time to be alive! And once again, a paper that almost nobody is  talking about. You can likely only hear about   it here on Two Minute Papers. So please don’t  forget to like and leave a kind comment so Youtube   starts recommending this kind of content, because  otherwise, unfortunately, it does not do that.