NVIDIA researchers cut contact computations from 16,000 to 300, making physics simulations 10,000 times faster. Robot arms trained in these virtual environments reached 85% success rates on assembly tasks in just 1.5 hours, with 100+ simulations running on a single GPU.
Today we are going to talk about video games, reality and robot arms and how the three come together. So how? Well there are these robotic contraptions in many factories that help us build stuff cheaply and efficiently. Here is a solution that works extremely well. But there is a problem. What is the problem? Well, it is extremely expensive and tailored to the task at hand. Now, from a different research paper, here comes the cheaper solution. General purpose robot arms. These are much cheaper. However, they are also much more unreliable and slow. So what is the answer? Well, hopefully the answer is training in simulation. That means that the robot would try to assemble all kinds of objects in a video game and reuse that knowledge to build things for us in reality. However, not so fast. There is a problem here. If we try to simulate just one nut and bolt to be able to screw it in, the resolution of this geometry in the simulation has to be so fine, just one bolt would need about 50,000 elements coming colliding with each other and unfortunately it typically looks like this, uh oh, not like reality at all. So no virtual factories for us, I guess. Right. Well don't despair. Here is a paper that might be able to help us. Dear fellow scholars, this is two minute papers with Dr. Zsolnai. But the troubles don't stop there. I mean, how do we even represent the geometry of a simple bolt? Even if we take a convex decomposition, collision checks are expensive. And still that is not an accurate piece of geometry at all. What about triangle meshes? These are used for detailed geometry in video games too. Well yes, that would work. However, unfortunately detecting collisions between the nuts and bolts takes too long. And this paper proposes a way to represent these geometries in a way that, that can give us a super detailed bolt. Ok, that's good. Now let's look whether we can run the nut and bolt simulation. And oh my, look at that beauty. And it runs in real time. Absolutely amazing. So, are we done? Well, not quite, not at all. In fact, due to computational and memory requirements, we could simulate at most about 20 nuts and bolts with which is not sufficient for a large scale simulation. We need to go faster. You see, we still need to compute the interaction between 16,000 contact points. But hold on to your papers fellow scholars, because this work also proposes a way to reduce that by about 98% down to about 300. Whoa. Now we are at tens of thousands of nut and bolt interactions per second. That is insane. That, dear fellow scholars, is finally a real factory. And if you can't believe this paper is more than 10,000x faster than previous techniques. I gotta tell you, it gets better. No one said it can only simulate nuts and bolts. It works well on other things. And finally, we can sort those nuts with a resonating feeder mechanism. That's why everything is so jumpy. Grab it and there we go. It is so fast. We can run more than a hundred of these environments in parallel. Fantastic. We can even run a full assembly board of commonly used parts. And it works beautifully. And everything is simulated down to the last millimeter. Yes, you can even plug in that virtual usb. A adapter in a virtual world. Accurately. A feat almost impossible to mere mortals. But wait, we are still not done. Something super important is still missing. What is that? Well, of course, teaching the robots to use these objects properly. This work also proposes a way to do that. And after an hour and a half of training, what can it learn? Well, it can learn to pick up that mat and screw it in with a success rate of what? Above 85%. That is incredible. Just imagine you doing it where first you have to align the mat properly, must not get jammed over time, and so on. Even the worst case of this problem has an approximately 75% success rate. And when compared to humans, it has a narrower range of force application, making it a really nice and consistent solution. And just imagine how good this will be just two more papers down the line. So we we went from an absolutely impossible problem to something that can be learned to be solved 75% to 100% of the time. And all this not only 10,000 times faster than in previous works, and not even in real time, but several of these factories can run at the same time on just one graphics card. I am completely stunned. And all this is possible through research papers and through sheer human ingenuity. What a time to be alive. And what would you fellow scholars use this for? Let me know in the comments below. And if you feel that like me, you have developed a little addiction here, subscribe and hit the bell if you wish to see more amazing papers like this. And. And if you wish to run your own experiments in the cloud, make sure to check out Microsoft Azure AI. Azure AI is a powerful cloud platform that offers you the best tools for your AI projects with responsible AI built in. And here comes the best part. You can even try it out for free through the link in the video description.
