Dr. breaks down NVIDIA's new soft-body simulation technique by Miles Macklin and Matthias Müller. It runs 80,000-element elastic simulations in 8 milliseconds per frame while preserving volume under extreme deformation, and the code is surprisingly simple to implement.
Dear fellow scholars, this is 2 Minute Papers with Dr. Today we are going to have a look at beautiful simulations from a quick paper. But wait, how can a research paper be quick? Well, it is quick for two key reasons. Reason number one, look at this complex soft body simulation. This is not a jumpsuit. This showcases the geometry of the outer tissue of this elephant and is made of 80,000 elements. And now hold onto your papers, away with the geometry and feast your eyes upon this beautiful simulation. My goodness. Tons of stretching, moving and deformation. Wow. So how long do we have to wait for a result like this? All nighters. Right. Well, about that quick part I just mentioned. It runs very, very quickly. 8 milliseconds per frame. Yes. That means that it runs easily in real time on a modern graphics card. And this work has some other aspect that is also quick, which we will discuss in a moment. But first, let's see some of the additional advantages it has compared to previous methods. For instance, if you think this was a stretchy simulation. No, no, this is a stretchy simulation. Look, this is a dragon. Well, it doesn't look like a dragon, does it? Why is that? Well, it has been compressed and scrambled into a tiny plane. But if we let go of the forces. Ah, there it is. It was able to regain its original shape. And the algorithm can withstand even this sort of torture test, which is absolutely amazing. One more key advantage is the lack of volume dissipation. Yes, Believe it or not, many previous simulation methods struggle with things disappearing over time. Don't believe it. Let me show you this experiment with gooey dragons and balls. When using a traditional technique. Whoa, this guy is gone. So let's see what a previous method would do. In this case, we start out with this block and after a fair bit of stretching. Wait a second. Are you trying to tell me that this has the same amount of volume as this? No sir. This is volume dissipation at its finest. So can the new method be so quick and still retain the entirety of the volume? Yes, sir. Loving it. Let's see another example of volume preservation. Ok, I am loving this. These transformations are not reasonable. This is indeed a very challenging test. Can it withstand all this? Keep your eyes on the volume of the cubes which change a little. It's not perfect, but considering the crazy things we are doing to it, this is very respectable. And in the end, look, when we let go, we get most of it back. I'll freeze the appropriate frame for you. So, second quick aspect. What is it? Yes, it is quick to run, but that's not all. It is also quick to implement. For reference, if you wish to implement one of our earlier papers on material synthesis, this is the number of variables you have to remember, and this is the pseudo code for the algorithm itself. What is that? Well, this shows what steps we need to take to implement our technique in a computer program. I don't consider this to be too complex, but now compare it to this simulation algorithm's pseudocode. Whoa. Much simpler. I would wager that if everything goes well, a competent computer graphics research scientist could implement this in a day. And that is a rare sight for a modern simulation algorithm. And that is excellent. I think you are going to hear from this technique a great deal more. Who wrote it? Of course. Miles Macklin and Matthias Muller, two excellent research scientists at Nvidia. Congratulations. And with this kind of progress, just imagine what we will be able to do two more papers down the line. What a time to be alive. This video has been supported by weights and biases. Being a machine learning researcher means doing tons of experiments and of course creating tons of data. But I am not looking for data, I am looking for insights. And weights and biases helps with exactly that. They have tools for experiment tracking, dataset and model versioning and even hyperparameter optimization. No wonder this is the experiment tracking tool choice of OpenAI, Toyota Research, Samsung and and many more prestigious labs. Make sure to use the link wnb me paperintro or just click the link in the video description and try this 10 minute example of weights and biases today to experience the wonderful feeling of training a neural network and being in control of your experiments. After you try it, you won't want to go back. Thanks for watching and for your generous support and I'll see you next time.
