Skip to content
Google's AI Frontier: RT-2, the Self-Teaching Robot Redefining Learning from the Internet episode artwork

Jan 21, 2024 11 min

Google's AI Frontier: RT-2, the Self-Teaching Robot Redefining Learning from the Internet

from The Daily AI

In this episode, we explore the frontiers of artificial intelligence with Google's RT-2, a self-teaching robot that redefines the learning process by autonomously acquiring and executing tasks from the internet. Tune in for an in-depth analysis of the potential impact on robotics. Invest in AI Box: https://Republic.com/ai-box Get on the AI Box Waitlist: ⁠⁠https://AIBox.ai/⁠⁠ AI Facebook Community Learn more about AI in Music Learn more about AI Models

Episode details come from the podcast's RSS feed.

Embed this episode

Ready to play

Google's AI Frontier: RT-2, the Self-Teaching Robot Redefining Learning from the Internet

0:00 11:23
of matches

Transcript Auto-generated, may contain errors

Today we have some breaking news out of Google's deep mind on AI mix with robotics. So let's jump in. Essentially, Google's deep mind team is really pushing the boundaries of AI and robotics. And they're bringing a new vision language action model.

So this is the robotics transformer to, which is the RT two. And this cutting edge AI model was trained on a massive amount of text and images collected from the internet. And this enables it to generate robotic actions, which is a massive step forward from traditional AI based chatbots like chat GPT or Google Bard, which essentially just generate texts, snippets for you, right? So this is really bridging the gap between web knowledge and robotics control.

So robots, unlike chatbots, need a very firm grounding in the real world to, you know, truly be able to be helpful to humans. And I think this has always been a very large task that they have to, you know, comprehend and carry out really complex tasks in a lot of very diverse and often very unpredictable environments to date. We have a lot of factories that have robots that are operating doing very simple tasks. But as soon as you want them to be able to not just perform a repetitive task over and over again, but kind of do new things and learn and adapt, it becomes very, it becomes much more challenging task to undertake.

And because of this, I think the training of models like RT two represent a much more intricate undertaking compared to, you know, training just a large language model or an LLM. So I think it's not just about, you know, a robot recognizing an apple, for example, it has to understand the context, it has to differentiate it from a red ball. I mean, it has to perform various tasks related to that object. So I think historically practical robot training demanded billions of data points about, you know, the physical world, but RT two has ushered in a more streamlined approach.

So building on our two RT ones ability to essentially just generalize information across systems are to actually create a single model capable of complex reasoning with just a for action of the amount of, you know, robot training data that is typically required. So let's talk a little bit about RT two's kind of novel ability. So what makes RT two unique, I think, is its ability to transfer knowledge from a, you know, vast corpus of essentially web data and to be able to navigate complex situations and human made requests. So it can comprehend and execute tasks like, you know, deposit in a piece of garbage, even without explicit programming for that specific action, right?

So normally when you have a robot, you literally just train it on how to do a task. A lot of times it's very automated. This is not very automatic. And so now, you know, you're telling to go pick up that piece of trash and throw it in the garbage.

And it's based off of everything that's learned and pictures it's seen, we'll go and try to perform that task, which is absolutely a massive step, right? It's the difference between, you know, perhaps you could train a robot to drive a car from, you know, point A to point B because you just, you know, have it automatically programmed that every inch it's going to turn right or left. It's a difference between that, which is what previously has been done to something like Tesla's self-driving, where, you know, that it's actually looking at the environment around it and making adaptations to how it drives, you know, it could avoid a pedestrian or it could stop at a stop sign. So it's very, very, it is a massive step.

And so of course we have something like Tesla self-driving, which is follows roads, which have, you know, rules and signs and a lot of information like that, which is specifically designed to do. This is a whole new step where this is a robot that essentially has to be able to do tasks in the world where there's an infinite amount of actions that can accomplish and there's an infinite amount of ways to do that. So I think this is really, really impressive. And I think that really kind of demonstrates the model's capability for learning beyond its initial training, which I think is key.

So Google engineers put RT2 through the paces with over 6,000 robotics trials and in tasks based on the training data, RT2 performed comparable to its predecessor, which is RT1. However, when faced with novel unfamiliar scenarios, RT2 doubled its success rate from RT1's 32% to a, you know, I think it's around a 62% success rate. So I think this improved adaptability is a monumental advancement in the AI model's capabilities. And you know, I can already hear a lot of a lot of doubters or haters or whatever you want to call them, right?

Skeptics, perhaps is a better word saying, Oh, look, this thing's only 62% efficient. It only has a 62% success rate. But like, look, RT1 was not put out that long ago and between then and now we've already doubled our success rate from 32% to 62% for completely unfamiliar scenarios. Meaning you plop this thing down in a park, you say, go pick up that piece of trash, go through in that garbage can, it's never been to the park before.

It doesn't know necessarily what that garbage can is or what type of trash that is. And it's able to have a 62% success rate executing that task. And imagine, you know, they ran this through 6,000 robotic trials. So imagine how many of those 6,000 tasks it was able to do.

I'm absolutely blown away. And for the critics or skeptics out there, I think this is incredibly impressive and especially the speed of adoption and improvement we're seeing here, I think is incredibly, incredibly impressive. So the RT2's model is an advancement on its predecessor, obviously, right? The RT1, which was trained on multitask demonstrations, which was essentially in an office kitchen environment and that happened over 17 months.

So the data collected by 13 robots informed RT2's development, which allowed it to demonstrate improved generalization capabilities and semantic and visual understanding beyond just the robotic data it was exposed to, right? So it's actually able to see the world around it, actually able to learn and it's able to improve based off of, right, its experience with everything around it, not just the data that was originally given, which is very, very impressive. The sort of backbone or, you know, the most important part of the RT2 is the adaptation of high capacity vision language models, which are called VLMs, but essentially they have been successfully trained on web scale data. So these models are brilliant at kind of recognizing visual or language patterns and kind of operating across different languages.

But to control a robot, these models need to be trained to output actions. So this challenge is addressed by essentially representing actions as tokens in the models output similar to language tokens, right? When we have an LLM like JTBT, it's not that it knows, you know, what word necessarily goes next. It just knows what token goes next in a sentence, which is a fraction of a word.

And by making it for actions of words, it's actually a lot more accurate. And that's essentially how these LLM are able to generate sentences. So they've done the same thing except now they're calling actions or parts of an action tokens, right? So you might need a hundred tokens to complete a specific task.

And it doesn't have to know exactly how to do the whole task. It just knows how to like do micro movements putting it towards that task, which is absolutely fascinating, in my opinion, and well done to the D-MIND team on doing that. So D-MIND actually performed a series of qualitative and quantitative experiments on RT2 models. And in this process, they categorized skills into three areas.

So they had simple understanding, reasoning, and then human recognition. So tasks requiring knowledge transfer from web pre-training essentially demonstrated it's, you know, pretty remarkable potential, but with significantly improved generalization performance compared to previous models. And I think that that's absolutely important. So on tasks seen in robot data, RT2 retained performance while improving on unseen scenarios by almost double, which essentially was showing, you know, a really massive benefit of large-scale pre-training and it outperformed multiple baselines on unseen tasks and achieved a success rate of 90% on the open source language table suite of robotics tasks in simulation.

Furthermore, RT2 actually proved capable of carrying out more involved commands that required reasoning about intermediate steps. So this actually showed that RT2 can plan from both image and text commands and essentially is, you know, enabling visually grounded planning, which is really, really impressive. I think RT2 showcases how vision language models can evolve into really powerful vision language action models. And that can really directly control a robot.

So I think this is promising and this whole development really signifies the dawn of a new area in robotic control. And I think it really raises hopes for the creation of general purpose robots capable of reasoning, problem solving, and performing a diverse range of real-world tasks. So while Google admits there's still work to be done, the success of RT2 is a testament to the potential advancements in general AI and LM technology. And I think the leap from RT1 to RT2 really just kind of underscores the immense strides being made in the field.

Now, I will say, of course, this is incredibly exciting, incredibly interesting, but I will say that there are some concerns here, right? If you really think about it, as these robots essentially are sitting here and learning how to take in the environment around them and do certain tasks, it brings up the question of robotics ethics and a whole Terminator-style event essentially happening where the robots are able to learn how to do something, right? Maybe they learn how to drive a car and maybe they learn how to rob a bank or maybe they learn how to fly a fighter jet or maybe they learn how to commandeer a freighter ship, right? Like there's all sorts of things that these can essentially learn how to do.

And I know a lot of people think like, oh, this is so far fetched, but in my opinion, based off of things I've seen from chat GPT or reflection AI, I think there's all sorts of dangers that definitely we need to be aware of. I think the never one most important thing is making sure that the preservation of human life is the top priority for these robots. And then of course, that begins at AI model, which is kind of the brains of a lot of this. So it's going to be really interesting to see how this evolves.

And I definitely think that is a major area of focus we need to continue to be very vocal about in whatever respective fields we have.

Frequently asked questions

How long is this episode of The Daily AI?

This episode is 11 minutes long.

When was this The Daily AI episode published?

This episode was published on January 21, 2024.

Is there a transcript available for this episode?

Yes, a full transcript is available for this episode. You can read the complete transcript on the episode page.

Can I download this The Daily AI episode?

Yes. Use the download control on the episode player to save the publisher-provided media file.