LARRUZ explains a computer-graphics technique that speeds up elastic simulations by 110 times in some cases. The method keeps virtual objects stable and prevents them from passing through obstacles even with larger time steps, turning minutes of computation into seconds per frame.
Dear fellow scholars, this is 2 Minute Papers with Dr. Today we are going to do this and this and this on a budget. Today, through the power of computer graphics research works, we can simulate all these amazing elastic interactions if we are very patient. That is because they take forever to compute. But if we wish to run these simulations quicker, what we can do is increase something that we call the time step size. Usually this means that the simulation takes less time, but is also less accurate. Let's see this phenomenon through a previous method from just a year ago. Here we set the timesteps relatively small and drop an elastic barbarian ship onto these rods. This is a challenging scene because the ship is made out of half a million tiny elements and we have to simulate their interactions with the scene. How does this perform? Uh oh, this isn't good. Did you see the issues? Issue number one is that the simulation is unstable. Look, things remain in motion when they shouldn't. And two, this is also troubling penetrations. Now let's increase the time step size. What do we expect to happen? Well now we advance the simulation in bigger chunks, so we should expect to miss even more interactions in between these bigger steps and whoa. Sure enough, even more instability, even more penetration. So what is the solution? Well, let's have a look at this new method and see if it can deal with this difficult scene. Now hold on to your papers and wow, I am loving this. Issue number one, things coming to rest is solved. And issue number two, no penetrations. That is amazing. Now what is so interesting here? Well, what you see here should not be possible at all because this new technique computes a reduced simulation instead. This is a simulation on a budget. And not only that, but let's increase the time step size a little bit. This means that we can advance the time in bigger chunks when computing the simulation at the cost of potentially missing important interactions between these steps. In short, expect a bad simulation now like with the previous one. And wow, this is amazing. It still looks fine. But we don't know that for sure because we haven't seen the reference simulation yet. So you know what's coming. Oh yes. Let's compare it to the reference simulation that takes forever to compute. This looks great. And let's see, this looks great too. They don't look the same, but if I were asked which one the reference is and which is the cheaper reduced simulation, I am not sure if I would be able to tell. Are you able to tell? Well be careful with your answer because I have swapped the two. In reality, this is the reference and this is the reduced simulation. Were you able to tell? Let me know in the comments below. And that is exactly the point. All this means that we got away with only computing the simulation in bigger steps. So why is that good? Well of course because we get through it quicker. OK, but how much quicker? 110 times quicker. What? The two are close to equivalent, but this is more than a hundred times quicker. Sign me up right away. Note that this is still not real time, but we are firmly in the seconds per frame domain so we don't need an all nighter for such a simulation, just a coffee break. Now note that this particular scene is really suited for the new technique. Other scenes aren't typically a hundred times faster, but worst case scenario is when we throw around a bunch of furballs. But even that is at least 10 to 15 times faster. What does that mean? Well, well an all nighter simulation can be done. Maybe not during a coffee break, but during a quick little nap. Yes, we can rest like this tiny dinosaur for a while and by the time we wake up the simulation is done and we can count on it being close to the real deal. So good. Just make sure to keep the friction high while resting here or otherwise this happens. So from now on we get better simulations up to a hundred times faster. What a time to be alive. This video has been supported by Weights and Biases. They have an amazing podcast by the name Gradient Descent where they interview machine learning experts who discuss how they use learning based algorithms to solve real world problems. They've discussed biology, teaching robots, machine learning in outer space, and a whole lot more. Perfect for a fellow scholar with an open mind. Make sure to visit them through WNB MEGD or just click the link in the video description. Our thanks to WeightsAndBiases for their long standing support and for helping us make better videos for you. Thanks for watching and for your generous support and I'll see you next time.
