Okay, So I'm curious because one with the IEEE, some assistant professor at the Department of prospects engineering. So we, I tried to make it as robotics as possible. But again, I'm like banners or more of the control theory side. But we have some cool work on motion planning and I'll be happy to show it to you. Okay. So I'm the Director of what we call less intelligence. Have a fiscal systems laboratory. So we work a lot on learning-based control with stability and Optimality guarantees. We also apply game theory to model disturbances, adversarial attacks, and so on, so forth. So if I have to describe my lab, so we work, I will say in the big picture of cyber physical systems, but on the control side. And then we design the resident controllers to secure actuator sensors and communication from different kinds of attacks. So we tried to predict attacks or mitigate, isolate, attacked components and so on, so forth. Okay, So I guess this slide is very busy, but don't worry too much about it. Okay, So why we work on online learning? Again, this is, this goes back to full autonomy, where we want everything to work AS we plug and play something and then we let it do whatever it has to do. Now pretraining. So usually, since we want to do learning-based control, we don't require any pre-training, So everything happens in real time in a plug-and-play framework. Okay, So then we go back to what is called full autonomy. Okay? So we want to learn from mistakes, okay, so we want to prove performance efficiency. And we want to learn from other agents that interact with them. Okay? So of course, if we can achieve all that, then our world will be changed. Okay, so again, we have lots of work to do here. And I think lots of the wonderful people here work on that. So hopefully sometime we will be alive to see fully autonomous things running around. Okay? And then we have, We have done some magazine articles. I can send them to you if you're interested. Okay, so there's some connections of learning and control, okay? So I think also Frank talked about optimal control, the thing. So lots of people, I think also panels talk about optimal control. So optimal control has to happen offline, okay, but the good idea about it is that it optimizes performance, okay? So I give you a performance which can be energy, time, or fuel. And then based on that, you compute your Ricard is you solve your 100 Jacobi Bellman equations and then you take this controller and you put it back to your system. Okay, But these habits of lines, so what happens when, for example, you burn fuel and your system changes mass? So then what you have computed before, it doesn't really work for the system. Okay, So what we have to do is we have to solve optimal control problems online. Okay? So in other words, to learn the optimal control solution. So if you had to think about it, is like, I'm designing an optimal controller that is adaptive. So some people call it off my adaptive control. And this has discussed connection with reinforcement learning. And then gets this beautiful picture from my friend Debbie. And so you see that reinforcement learning is in the middle of many things, which is optimal control machine learning operations research, game theory, and classical operand conditioning. And you will see that actually impede this talk. You will see that all these things interact in different ways. So again, I think reinforcement learning in a feedback loop can significantly improve many things in a tunnel. Okay, So of course, learning was not something new. So you see the father of information here, Information Theory. He demonstrates tissues. Okay, So and this actually learned how to navigate around the maze. Okay, So that was, that happen 1952. But of course, there's many issues nowadays for one of the most important one is the constitutionality. And this actually happened when bellman, who was actually working for and cooperation develop dynamic programming. Dynamic programming is not really programming, but because his boss didn't like anything that had to do with applied math. So she got to name it as programming. Okay, so Washington, his boss, is happy. Okay, so Guzman tonality is one of these things, of course, when you do learning, you have to, to check what you have to optimize. Some people follow the slope. Competition probabilities in optimization, okay, some people work on stochastic and deterministic constraint, so forth. And of course, everything then boils down to some, pitching some parameters, okay, optimizing these parameters. And then just let the algorithm to the rest. Okay? So this is actually what learning is. And Samuel actually defined that as learning is the ability to learn without being programmed. So anything that has to do with pre-training your program to do it. So that's why we tried to do fully autonomous. Okay? So another thing that actually are going to saw here is learning and morality. Okay? So morality in human autonomy, okay? So of course, do the right thing, right? But what is morality must be metonymy. Okay? I guess do what you're told, right? And okay, but I wanted to show you this one, so this is a very important thing. Okay, so what you do here, I guess, if you find what you do here, you can get millions. But some people say, okay, just run them over and then come back and run him over two. So that's one thing we can do there. But so to be fair. Okay. And I'm sure some of you, you also have to sum the Roombas. Okay. So if you, let's say you spill the milk that there's a fire there. So again, if you want to really clean the milk, then you will burn your Roomba. Okay, so let's go now to start with a small introduction and then I want also to differentiate between automatic, adaptive and autonomous control. Okay, so autonomous is the time that you and you introduce some, some cognition. Then we're talking more about autonomy. Okay? So now let's say what we do in my lab. So, and actually I like that the Frank when before. So now you're familiar with dynamic motion planning. So okay, so let's say you have an autonomous system and you want to navigate in an unknown environment, okay, So they're obstacles, you don't know they're there, they're moving obstacles. And you have also to predict the movement of the obstacles. Let's say you're in a mall, okay? And then you see that people, they don't cross because they predict where it's personable go. So and we do that with what is called, as you'll see pretty soon, a cognitive reality which is part of game theory but doesn't happen equilibrium. Okay, so I'm sure you have seen the movie, a Beautiful Mind Witness. Okay, So these cups and equilibrium. So, but some, let's say some robots or some humans, they're not really rational. So that's why we want, we don't apply equilibrium there. Okay? So some motion planning is the brainstem problem, okay? Because you have dynamic constraints, you have temporal constraints. Okay? So I went to navigate and go from one place to the other place and specific time. And of course, I want to do that based on a blunt, but of course the plant, if I want to be fully autonomous, had already know the plant, okay. So I want to have it completely model-free. And in other way you can also have some disturbances. And so you also have to encounter for safety in the presence of disturbances where this disturbance also can be different attacks on your session actuators. And everything has to go up and as I said, real-time. Ok. So Of course, yeah, this is just some news coverage. Don't worry too much about it. Okay? So how you do safe learning? So safe learning happens when you incorporate some temporal logic constraints. So let's say I want to go from point a to point B. But before I go to point B, I have to bus from another point. Okay, so you see here, how about trajectory? And I want to go to omega 2, but before that I have to pass four megawatt. Okay? But while I do that, I always have to stay inside the omega-3 that you say here. Okay? So you can always do that by combining learning, reinforcement learning and barrier functions and Deborah constraints. And we have the framework for this one. Also. Another IDS, it comes from behavioral psychology. So we, we work on being inspired by actually Skinner with who develop operand conditioning. We want to see how all these ideas apply back to control systems, engineering systems, robots, which can be any ground flying robots are underwater vehicles as you'll see later. Okay, so operand conditioning. So you can think that I get the reward not every time I do something good, but I will be rewarded. Not surprising effect. Okay? So. So for example, I have to do five good things to get rewarded. Okay? So after I get five good things that get rewarded or either good things, but I don't know when I will be rewarded, so I keep doing good things. Okay? So this is the basic idea of operant conditioning. And actually we have, we have applied successfully to engineering systems and we find something similar with what Skinner found in humans. Okay, so let's see how actually don't worry too much about the math. Okay, So this since I'm a control person, I have to put some mother. I want to go to the cool stuff. Okay. So let's say you are in an environment, okay, So you're going to go from point a to point B and you have unknown obstacles, Oliver. Okay, so let's say you're in a forest, okay? And you fly, but also there are other, other, like birds fly around and so on, so forth. So let's see how this power. So, so you see I'm the Vigo over there and you see things appear randomly. I don't even know where they are. And you'll see that my system will actually navigate in real-time and will replan while going to the correct direction. Okay. And all these things, they happen in real time without requiring any knowledge of the system. Um, okay, so here as you see, what we have done here is that the obstacles for not moving. Okay? So if we make the obstacles move, so probably we will require what I said about a non-equilibrium game theory. Okay? So game theory lets us be robust to something that somebody else is doing. Okay? So what we will do is we will predict how the other obstacles or agents will move in real time. Okay? So, okay, so for example, let's say I'm, I'm not that smart. So what I will do, I'm a level 0, okay, So in psychology, you actually model intelligence with different levels, which is called cognitive yet. Okay? So let's say level-0. You are a random player into the system, okay, So if your random plane to the system, what you will do, I will just move straight biting you that there is no obstacle. Okay? Then if I am a level one, then I will think that the other eight and is level 0, okay? And then it goes straight, Sobel avoided, and so on, so forth. Okay. So let's see how this can happen. So these things actually there had been fighting by nasa. So we want to apply these things in urban air mobility where you have flank taxes, okay? So here what you see on the picture on the right is you see four agents. And then you can think that they're UAVs or ground vehicles and so on, so forth. So one player is in gray, okay? And the goal state is in green. You see two level 0 ADC yellow, okay, you see them over there. And you see some play obstacles that they are actually black. And them. And you see that as I move around, I recognize the movement of the other agents. I predict their movement, recognize what level of intelligence they are with a significant probability, and then navigate around without collisions. Okay? Okay. And let, these are some applications we have done general, so we have built robotic hands. We have a project that we want to do firefighting with drones. We have built a ground vehicle with the support of nato. We work with drones. We actually recently got interested in doing shape morphing with learning to find the optimal CEP of wings. We also work with underwater vehicles, specifically for security. So we want to see what happens when you do GPS spoofing. And then also we want Finally, that's my last slide to do human autonomy interaction. So we can build some tools that actually they're being used currently on the Pentagon to train security analysts. So and so you can think that I have a mission which the missions is somewhere in, I don't know, in Afghanistan. And I want then I have my drones and I think I'm out of time. Thanks. And then of course, I want to thank my students. Without them, nothing is possible and of course my sponsors and anything that you want to find this in my web page. Or if you want any copies of my books, I will be happy to share.