Googles New SELF IMPROVING AI 'Robocat' Takes Everyone By SURPRISE! (Now ANNOUNCED!)

So Google's specialist AI division, Google DeepMind has once again raised the bar in terms of robotics. They've introduced something called Robocat, a self improving robotic agent. And like most times, people have missed the mark as to why this is truly revolutionary. And it's no surprise because DeepMind, if you aren't really sure as to what they've done before, this is the same company research team that has come up with many different instruments before that have truly changed our world. They're responsible for this project called AlphaFold, which is a program that can accurately predict 3D models of protein structures and is accelerating research in nearly every field of biology. They are also the team that crafted AlphaGo, which is the first computer program to defeat a professional human Go player and the first to defeat a GO world champion, and is arguably the strongest Go player in history. And they also created Wavenet, which creates more natural sounding speech for products used by millions of people around the world. So the reason I gave you the backstory of that team is because you need to understand that this team is, if not the best, one of the best when it comes up to introducing breakthroughs in the world of artificial intelligence. And Robocat is no different. So now a quick disclaimer before this video does delve into some more detail. DeepMind didn't actually create the robot. What they've actually done is, is created the system that powers the robot. So this, what you're going to be looking at is essentially an artificial intelligence program. And I know some people might get confused as to what the robot actually is. The robot is a Sawyer robot, which is a high precision robot which is used in many different factories. So the reason that this new Robocat is honestly so impressive is because of what it able to accomplish. So the blog post starts out by explaining that robots are particularly becoming newly integrated into our lives. But there is a large problem in robotics which many people have started to realize and it's something that we are all so familiar with when looking at large language models. And that is the problem bottleneck of training data. So in order to have a robot perform something effectively, it it needs to essentially be trained on a set of data that it can then understand exactly what to do. Now oftentimes what we do do is we'll have a robot understand or watch a million or hundreds of thousands of different examples before being deployed and before the robot artificial intelligence manages to perfect the said task. Now, what DeepMind have done is they've actually managed to get this robocat to self Generate new training data to improve its technique and significantly cut down the time to make itself understand what specific tasks to do. And for the specific number that DeepMind talk about, they say that Robocat learns much faster than other state of the art models. And it can pick up a new task with as few as 100 demonstrations because it draws from a large and diverse data set. And this capability is going to help accelerate robotics research and is going to reduce the need for human supervised training and pretty much is going to step us a giant leap in the general purpose robot AGI era, which we are quickly embarking now. Another thing that I did find truly interesting about DeepMind's robocat was the fact that this robot was essentially programmed to not only work across one specific hand, but to pretty much work on any hand that there was. Which is definitely quite scary because it's quite like those movies where you get an AI system that sort of infiltrates a system and understands immediately how to use everything, even though it's never seen it before. And that's what we did see with Robocat. The DeepMind also did publish these specific examples of where these robots were tasked with a goal image and then obviously asked to recreate said image. Now, these robots did perform the task exceptionally well. And what DeepMind also showcased was that even when shown a task that they hadn't seen before in training, and even applied to a robot which hadn't been seen in training before, the robot artificial intelligence program was able to do this very effectively. And this is why this is such a large breakthrough. And there's honestly a lot to that this paper does cover, which we'll explore later. There is also an area in which we see the robot artificial intelligence program react to disturbances in the environment. And this is something that we did see in earlier papers from Google before, but this was still nice to see, even on a artificial intelligence program, which is still in relatively early stages. Which means that these robots are going to be very effective at real world scenarios. Because, as you know, the real world isn't just a test facility where we have a few objects. There are always going to be things that happen that don't go according to plan. And it's important for these robots to be able to quickly and robustly adapt to these scenarios, which is what we see demonstrated here. Now, the most interesting thing for me when doing research on Robocat was the fact that Google has this, their own multimodal model called Gato, which is essentially Spanish for cat, also pronounced Gato. Now, essentially, why this is so incredible is because we know that Google's DeepMind is currently working on a product called Gemini. Now, the Gemini artificial intelligence program, as you may know, is set to rival OpenAI's GPT4, GPT5. And they are working very, very, very hard on this. This GATO multimodal model, which can process language, images and actions in both simulated and physical environments, seems to be already very, very effective. And we all know that multimodal models are going to be the models of the future. So the reason robocat can actually improve itself autonomously is because they manage to combine gato's architecture with, with a large training data set of sequences of images and actions of various robot arms, solving hundreds of different tasks. So what we have here is robocat's training cycle, which was boosted by its ability to autonomously generate additional training data. So step one was to collect 100 to 1,000 demonstrations of a new task or robot using a robotic arm controlled by a human. Then it was to fine tune Robocat on this new task arm, creating a specialized spinoff agent. Then the spin off agent practices on this new task arm on an average of 10,000 times, generating more training data, and then incorporate the demonstrational data and self generated data into robocat's existing training dataset. Then you can train a new version of Robocat on the new training data. Now, what I did find interesting is that Robocat training data differs from the traditional approach. You see, when you're training a robot, essentially you're going to be using five common forms of training data, which is number one, sensor data, which the robot arms are often used various different sensors to perceive the environment. This sensor data can include information from the cameras, depth sensors, force sensors, joint encoders, or any other relevant sensors. Then also we have state information, which is the state of the robot arm, such as joint angles, joint velocities and effector position or orientation can be recorded as training data. Then of course we have demonstration data, which involves a human operator or expert performing the task using the robot arm. Then we have reinforcement signals, and reinforcement learning algorithms require additional feedback in the form of reward or reinforcement signals. And, and these essentially can indicate the success or failure of a particular action or state guiding the learning process. And simulation data, which is of course what we see here, where these are synthetic generated simulated versions, where you simply have sensor readings that are simulated robot configurations, which are all simulated in an environment like Nvidia's Isaac Sim. But what made this very, very good was we had Two additional versions of training data, which was of course V videos and this was able to be done by their multimodal feature. So I'm guessing it just simply had the video of it and was able to learn via that. Then of course we had the self generated training data, which is of course another method of training your data as well. So I do find that this is all very, very groundbreaking stuff. Like we stated earlier, Robocat's diverse training data was so vast that it essentially was able to train itself on arms that it hadn't seen before and able to adapt to these different controllable inputs. So on the left you can see a robotic arm that Robocat learned to control. And then you of course you can see a video of Robocat using this new arm to pick up the gears. The paper further goes on to say that after a thousand, only just a thousand human controlled observations collected in hours, Robocat could direct this new arm with enough grip to pick up the gear successfully 86% of the time. With the same level of demonstrations, it could adapt to solve tasks that combined precision and understanding, such as removing the correct fruit from a bowl and solving a shape matching puzzle, which are necessary for more complex control. And remember, the key to why this robot is such a large statement from Google's DeepMind is the fact that this robot is able to continuously self expl improve. And this is all down to its virtuous cycle of training. The more new tasks it learns, the better it gets at learning additional new tasks. So the initial version of Robocat was successful just 36% of the time on previously unseen tasks, after learning from 500 demonstrations per task. But the latest Robocat, which had trained on a greater diversity of tasks, more than doubled this success rate on the same exact task tasks. So the ability to independently learn skills and rapidly self improve, especially when applied to different robotic devices, will help pave the way to a new generation of more helpful robotic agents. So to conclude, Robocat demonstrates the ability to generalize new tasks and robots in both zero shot and through adaptation, using a relatively small number of examples. And remember, zero shot is simply where you haven't seen the task whatsoever. And this adaptability and generalization make Robocat highly versatile and efficient in learning new skills. And Robocat can generate its own training data for subsequent iterations, creating a unique self improvement loop. And this iterative learning process contributes to its autonomous improvement and enhances its capabilities over time. So overall, Robocat's combination of visual goal conditioning and adaptability, generalization and self improvement sets it apart from previous AI robotic systems, making it a groundbreaking development in the field. And I do think that the future applications of DeepMind's Gato multimodal, multitask, multi embodiment, generous policy can have significant impacts on further applications.