ColdFusion - Google Just Put an A.I. Brain in a Robot [Research Breakthrough]

Starting point is 00:00:00 Hi, welcome to another episode of Coldfusion. These were the winners of an AI competition in 2020. Frankly, they're nothing groundbreaking and what they're trying to depict is barely coherent. Looking at some more AI-generated images from back in 2020, it's amazing that in just two short years, we have photorealistic imagery. While AI-generated images have obvious visual signs of improvement, what many people don't realize is that in the past two years, years, AI language models have also drastically improved. In a previous episode, we saw how a Google

Starting point is 00:00:39 language model conversed so well that one of Google's engineers claimed that the AI had become sentient. In another infamous case, Redditors were fooled by an AI for months on end. Over in the other corner, we have Boston Dynamics, and in a smaller sense, the Amica robot. They've both shown us that robots can indeed be made with human dexterity. So you probably know where this is going. What if we happen to combine AI's new understanding of language with a physical robot body? Well Google has just done this and the results are very interesting. In this episode, we'll learn all about it and see what could be in store for the future. Google Research Labs, along with another alphabet subsidiary called Everyday Robots,

Starting point is 00:01:29 have teamed up to make a new kind of robot. You see, instead of carrying out pre-programmed tasks, this new robot can understand a spoken task in natural language and then do that task for you. This method removes the obvious limitations of pre-programmed hard coding and it opens the door to a massive amount of possibilities. To achieve this mission, the team put one of Google's most advanced AI brains in a robot body, and when I say brain, I'm not using that word loosely. A recent 2022 paper, published in the journal Nature, found something interesting. When he human brain patterns were measured in experiments and compared to AI language models, there were some uncanny similarities. Particularly in the way both systems process language, specifically in the way that they predict words from context.

Starting point is 00:02:22 In addition to this, their neural activations mirror each other. The AI's internal activities, quote, significantly map onto human brain responses to written sentences. So it sounds weird, but according to this research, we're beginning to learn that AI language models are behaving like human brains. Kind of surprising, but maybe it's just the fact that technology is starting to help us understand biology. So back to the robot. In summary, the physical robot acts as the language model's hands and eyes, while the language model supplies the knowledge about the task at hand.

Starting point is 00:02:59 To give an example, a user can ask a task, I've spilled my drink. The robot interprets the sentence using the language model. It comes up with a bunch of solutions in its software, but only selects what it's capable and best at doing. It then analyzes its environment to see how to physically proceed. Okay, so what? This couldn't have been that hard to do, right? Well, there's a lot of careful engineering that had to be done to make this possible. If you ask an AI language model a question, it will give you a sensible and reasonable answer,

Starting point is 00:03:31 but that answer may not be feasible or useful to the robot. For example, if you say, I've spilt my own. drink the language model may tell you you could try using a mop but if there is no mob in the scene or if the robot is incapable of using a mob the task is impossible Google researchers managed to create a way for the robot to understand what it can and can't do but more on this later AI computer vision was used to pair what the robot is seeing with its vast knowledge base from the internet the robot can understand and complete a task its complex as I spilled my orange soda

Starting point is 00:04:07 can you throw it away, bring me something to help clean it up, and then bring me a replacement? So in all of this, the key here is that the machine ends up being pretty open-ended. It's general. Industrial machines, for example, have to be precisely programmed to do a very specific task. So how is understanding achieved? Google Research is using their palm model in the robot. So what on earth is palm? Palm stands for Pathways language model.

Starting point is 00:04:39 It's an algorithm that's been trained on countless amounts. of text from the internet, we're talking hundreds of billions of parameters and petabytes of text. The result of all of this training is its ability to understand language. It's important to note that these types of massive language learning models are pretty new, only around two years old. The vast levels of training on billions of parameters makes them vastly different to services like Siri or Alexa. In addition to the Palm model, the researchers used another system of code to let the robot know what it can and can't do. This code, called world grounding,

Starting point is 00:05:17 trims out unsafe or nonsense actions. Machine learning is one of the few things that we have that really makes it possible to get computers to behave intelligently in unstructured environments by using experience, using data, using things that the computer has seen in the past to figure out patterns. Reinforcement learning is one of our big bets.

Starting point is 00:05:41 It's a machine learning algorithm where each time the robot attempts to do a task, we can reward it if it succeeds, tell it that it didn't succeed if it fails, and then the things that lead it to succeed get rewarded more and more often, and it gets better and better at the task. And this is something we can do across many robots at the same time to share that learning and that experience. To speed up the learning process, the software was trained in a virtual world so it could gain the benefits of knowledge through a simulated environment.

Starting point is 00:06:09 But, without having to speed it. spend mass amounts of time in the physical world. We can really multiply our impact by having the robots also practice how to do things in simulation. We can do things like simulate many other buildings and many other types of objects and many other types of lighting conditions and backgrounds. And what we're being able to do with the simulator is significantly reduce the amount of real-world data that we've needed on the order of going from about 500,000 attempts all the way down to only needing 5,000 attempts to learn grasping. grasping.

Starting point is 00:06:41 We've seen this method before in the episode about OpenAI and their dexterous hand robot. I thought it was a very clever idea then. So finally, the robot's world grounding was tied into another section of code that turns high-level instructions into low-level sub-tasks that can then be performed. Once all of this is done, you have a robot that can understand and then perform tasks without human intervention, and it does so better than anything else out there at the moment. In their paper, Google claims that the language model and robot selects the right sequence of skills 84% of the time and executes them successfully 74% of the time.

Starting point is 00:07:21 This is a 50% improvement over the previous latest research. The computer scientists also state that the performance keeps improving when the language model does, so this could get much better pretty quickly. The team stated that the mobile robot could accomplish a large set of abstract instructions in a real kitchen with objects it's never seen before. The significance here is that this is the first time a large-scale model has been used in a real-world robot. It can be seen as a hallmark in the field of robotics. These findings suggest that the field of robotics can benefit from the improvements in language modeling.

Starting point is 00:07:59 This could bring the two areas of research closer together in the future. As for the robots, 100 of them are doing autonomous tasks around Google's offices. It has to wander around buildings, find the trash cans, look in them, understand what's in them, and if something's in the wrong bucket, look in those buckets, robot, and if you can, move it into the right bucket. The sorting experiment is really pushing us to try and write an application that improves by itself just through practice. Is that going to work? I hope so. The way we made it, the tools and recipes we used to create this pick an object, out of a bin capability, we'll be able to use that same recipe to learn a new skill.

Starting point is 00:08:46 The more we give the robots a problem and they get better at it just by trying to get better at it, that's the thing that will convince us that these robots are now ready to go solve lots of problems for people in their everyday lives. So that was kind of a lot to get your head around. But what does this mean for the future of robotics? Although it is a major milestone, right now, Google's robot, is pretty slow at completing tasks and can only grab basic objects. But then we have to ask, what about in 15 to 20 years?

Starting point is 00:09:27 Will we have to worry about physical automation for entry-level jobs like fast food? Taking into account the current pace of advancements in both robotics and the new AI that's starting to power them, this could be a strong possibility. But on the flip side, there are plenty of other use cases too. Language model robots could be useful in anything from aged care to growing food or cleaning up dangerous waste and pollution. Robots like Spot have already proven the latter possible. In any case, the future is definitely going to be interesting. As for me, in summary, I wouldn't exactly want one of these Google robots in my house,

Starting point is 00:10:04 but I definitely do see the value in the research for certain situations. As a side, if you're confused and asking yourself, why is this all suddenly happening now? Where did AI even come from? And who exactly invented AI? You're in luck, because I've done a whole documentary on it, and you can check it out in the description below. I just think it's an important story that adds a lot of context to what's happening today.

Starting point is 00:10:31 Anyways, that's about it from me. My name is DeGogo, and you've been watching Cold Fusion. If you did enjoy this episode, don't bother subscribing. But instead, have a browser of this channel. There's plenty of interesting stuff on you. All right, so I'll see you again soon for the next episode. Cheers, guys. Have a good one.

ColdFusion - Google Just Put an A.I. Brain in a Robot [Research Breakthrough]

There aren't comments yet for this episode. Click on any sentence in the transcript to leave a comment.