Embodied intelligence

hello and welcome it is July 17th 2024 we're in active inference guest stream 83.1 with Josh bongard thank you for joining this should be very exciting we'll have a presentation and then some discussion so if you're in the live chat please feel free to write any questions and thank you again Josh looking forward to this yeah thank you Daniel and thanks to those of you that are attending online um so my name is Josh bongard I'm a professor of computer science uh here at the University of Vermont and um my bread and butter in my lab is the study of Robotics and Ai and obviously we're in the middle of the uh a current AI summer um so what I wanted to do today is uh show a couple highlights from my group things that I've worked on in the past and that we're working on at the moment um that I hope in the long term will help us realize sort of the long-term vision for a lot of those trying to create intelligent machines which is to create machines that are helpful but but also safe um we're part of the way there but as anyone who's worked with chat GPT or stable diffusion NOS or even has a robot vacuum cleaner at home there's some limitations to our current technology it's hard to create machines that are autonomous and useful and safe all simultaneously so what are the things that we're missing um that's what I wanted to sort of um seed the pool with today and hopefully we can move on to an interactive discussion about it so uh I'm going to leave this slide up and just sort of talk over this slide for a few minutes that'll hopefully generate some food for thought and then questions um this is a snapshot from uh some of the projects I've worked on over the years um first thing you'll probably notice is there's a lot of different robots that have very different structural properties they not only act differently but look very different and that is a fundamental foundation in everything we do uh in my research Group which is to try and understand how the body facilitates uh cognition uh years ago with my PhD advisor we wrote a popular science book called How the body shapes the way we think and we can mean that literally or figuratively um we wrote that book a while ago we made some arguments about how the body shapes the way we think and since that time my group and others have formulated uh other uh Arguments for why or how the body shapes the way we think and I'm hoping to survey some of those uh today so as I mentioned you can see a lot of different robots here very different structure they're they're they've got very different Form and Function but across each of the um pair of videos that you see here you'll notice that there's also a common pattern which is on the left side you tend to see something that's virtual and on the right side of the video pair you tend to see something that's physical and this illustrates the basic approach that my group takes to understanding how the body shapes the way we think which is to create AI that creates robots creates embodied AI so what do I mean by that uh what I mean is that in all of the projects that you see here um we create an AI that searches the space of all possible robots that could solve the tasks that we want the robots to solve um most people that are familiar with AI and Robotics and autonomous cars and drones have a rough understanding that AI is somehow optimizing or tuning the brain of the autonomous car or the Drone or the the robot what have you there's an underlying assumption in all of that current in in most current robotics which is that the AI Tunes the brain but does not tune the body Tesla cars are dreamed up mostly by humans and an AI Tunes their their brain or their control policy but of course nature doesn't work that way nature produced us and all the other intelligent organisms on this planet by carefully tuning body and brain simultaneously certain bodies facilitate certain kinds of behaviors and certain kinds of intelligent behavior and other bodies don't…