Embodied Intelligence Mission
December 2, 2022
November 4, 2022
Leslie P. Kaelbling
All Captioned Videos Advances in the quest to understand intelligence
Leslie Kaelbling - MIT EECS, MIT CSAIL, MIT Quest, MIT CBMM
Nancy Kanwisher - MIT BCS, MIT Quest, MIT CBMM
Mehrdad Jazayeri - MIT BCS, MIT Quest
JAMES DICARLO: As I said at the beginning, and Josh outlined as well, we're trying to execute on this vision. We've organized some of our big bets around things we call Missions which are these long-term collaborative projects that are rooted in a foundational question of intelligence around some domain of intelligence.
And you'll hear about several of them today. First you're going to hear from a team of people, Leslie Kaelbling, Nancy Kanwisher, and Mehrdad Jazayeri, about the status of something we call the Embodied Intelligence Mission. So Leslie, go.
LESLIE KAELBLING: All right. Thanks very much. So I'm Leslie Kaelbling and I'm an engineer. So that was a little bit of a change of emphasis and focus. This is a talk that I'll be giving with my colleagues Nancy and Mehrdad. And we are going to talk about one of the Missions of the Quest, and the Mission is Embodied Intelligence.
So the fundamental question is, how is it that our bodies actually affect the way that we think about the world, and how do we make intelligent agents that are-- that demonstrate their intelligence via their interaction with the world?
So if you look at what people do, they do things with their bodies and their hands all the time. They build houses, they make food, they care for people, they solve problems in interesting, new ways. And fundamentally, intelligence shows up in all of these interesting ways that people interact with the world around them.
And so if we think about robots-- so if I come from engineering, I come from robotics, I think about, well, how could I make robots that are intelligent? And what does it mean-- how is it that the intelligence really helps a robot solve a problem? So when we think about robots, often we think about very stereotypical robots solving stereotypical problems, but our interest in this Mission is in actually solving problems of the kind that humans solve.
So how could you take a robot and put it in something like a disaster site and have it react appropriately to all the wild and crazy things that could happen? How could you take a robot that could really help a person in an open domain house with all kinds of new objects and problems to solve? To work on a construction site, where as we all know, nothing goes exactly according to plan. Or to work in a-- say a supply depot that has, again, a great variety of things to do.
So what do we have to do to make this happen? Well, we have to build up from the engineering, but we also have to really fundamentally understand the nature of intelligence. And Josh just slightly stole my thunder here, but I'll take it back. I'll take it just right back and we'll say, well, OK, so how are we going to think about building these intelligences? And what is the underlying science?
And of course, in brain and cognitive science, for years people have thought about this. They've developed models of neurons and neural circuits, they thought about how you could build on those circuits by understanding the mental processes that run on top of them. And how ultimately they give rise to behave in humans and other animals, and how people who study neuroscience and cognitive science try to build theories of this kind of intelligent behavior.
At the same time on AI, we've been replicating or cross-pollinating. And so we've built a system with gates and with circuits and thinking about how algorithms run on top of them and how that can generate behavior and robots and AI theory is a thing. So there are two parallel processes.
There's another process which is coming up from theory. So there's math and theory of information and computation and so on. And what we're really after is a thing that's missing in this picture. And it's the thing that connects all these parts together strongly.
And so we're putting our bet on this notion that there's a general idea of intelligence, and one kind of intelligence is what you find in nature, there might be other versions of intelligence that you find in artificial systems, but by studying them collectively, we can arrive at an understanding of intelligence as a particular process that happens in many different contexts.
So OK, so let's just take the robot perspective for a minute and think about, how are we going to make intelligent robots? And there's a bunch of different strategies, people are placing their bets differently. One category is reverse engineering humans. It seems like if we could understand exactly how humans work, we could replicate them and we could make intelligent robots that way.
Another strategy is to try to replicate evolution, to do an enormous amount of work offline, searching for good solutions to the problem, often maybe at this moment by training neural networks and so on, but doing an offline search for a very good behavior.
And another strategy is to just use the standard principles of engineering, of understanding the system the world that we live in, building in constraints, designing systems using classical methods. So all of these methods have been successful in the past in various ways, but the current research culture is kind of a monoculture.
And almost everybody is working on trying to build systems that search offline for a good solution-- like evolution search is offline for good solution that we can then deploy in the world.
Our bet is that actually, you do better by combining all of these three methods, and in particular, that you can learn from natural systems and that you can learn from engineering principles good kinds of structure that form a substrate for doing the learning and that can make it much more efficient.
And so if you think about how much work it takes to arrive an intelligent system, our bet is that by building in strong priors, we can learn pretty quickly as a function of time or energy or CPU cycles or GPUs or whatever. And that the approach that builds in less bias might take a lot longer, a lot more data, a lot more compute, a lot more time.
Although it's possible in principle that because it has less bias built into it, it could actually eventually learn to be better than the somewhat more structured and designed system, although we're not sure. And we can't really predict where these curves are going to cross and we're not even sure if they will cross.
But it's a really interesting thing to keep in mind. We think that both things could work out well. We don't mean to say that you should do only one or only do the other, but that a monoculture is not such a good idea.
OK, so why does embodiment matter? Well, if you ask, how do you measure whether something is intelligent? You look at its behavior, and its behavior comes through the interaction of the creature in this world.
So our strategy is to define a broad range of naturalistic tasks, study humans and animals solving them, build robots that can solve them, too, and our prediction is that if we're asking these systems to solve similar distributions of problems, similar kinds of tasks, that we'll arrive at similar kinds of solutions.
OK. So an important point about our approach is that it's integrated in a bunch of different kind of ways. Maybe the most important kind is that we observe in our fields, both in neuroscience and cognitive science and in computer science, that researchers, people tend to study little pieces of a big thing.
It's easy to put your hands around a piece and study the heck out of it, but it might be hard to figure out how your piece and my piece fit together. So in brain science and in computer science, we observed that no individual module by itself generates behavior, and that to understand behavior, really, and therefore, intelligence, you have to understand whole-behaving creatures.
And if we do them independently, we'll make assumptions that don't match. And so we have to talk to each other all the time and be sure that our parts and pieces fit together.
Another kind of integration is integration of methodologies. So there's all kinds of different engineering methodologies, different natural science methodologies, different ways of designing structures. And we think there's value in lots of these, and that again, we have to just keep a dialogue going so that we can take the best parts and put them together in a good new way.
We have a team that's very interdisciplinary. We have different technical strategies, we have different foci, but we have a good property that we like each other and we talk to each other and we work together. So the arcs on this graph are existing collaborations that we had even before this thing started. So we're a tight team that covers a lot of territory.
Why now? Well, Josh and lots of people have talked about this already, there's been major advances in neuroscience, cognitive science, computer science, robotics. And these advances have been going on independently, but we feel like they have put us in a place where we're ready to put it all together and see where we can get.
OK. So we have really great reason for optimism about this whole thing. And now my colleague Nancy is going to take over and talk a little bit about a particular example that's been very successful.
NANCY KANWISHER: Hi, everyone. So as Leslie mentioned, I'm going to get much more concrete and talk about some specifics of some of the things that have already come up here. So I think one of the great successes that several people have mentioned in recent AI is the ability of artificial neural networks trained on lots and lots of labeled data to basically be able to tell you what categories of objects are present in an image.
And that's a real triumph for AI, but it's also very cool because it gives us concrete, image-computable, computationally precise hypotheses for what might be going on in the brain. And even cooler than that, data that's already come along suggests that there are, in fact, striking similarities between how ANNs and how brains solve the problem of object recognition.
So just as one example, in a seminal paper eight years ago, Jim showed that you can actually use ANN models to make to predict how a single neuron in a monkey brain that's part of the process of computing what objects is in front of the monkey, how will that neuron respond to a totally new image that you haven't presented to that neuron before.
So in red, it's the predicted response of that neuron to several hundred different images that the neuron hasn't seen before. And in black is the actual observed measured response, and you see a remarkable ability to predict that response, and that's amazing.
And more recently, Jim's lab and my lab have been collaborating-- actually, by Ratan Murty, who did this beautiful study over here-- he's sitting over there-- where he used similar methods to show that you can predict the magnitude of response of a patch of the cortex in humans, a patch that's involved in visual recognition of faces and objects and scenes, you can predict how strongly that patch will respond to a novel image.
And what that shows is that in fact, brains and artificial neural networks have arrived at least somewhat similar solutions to the problem of object recognition. And I think that's just kind of astonishing. It so did not have to be.
And to me, it's just a beautiful demonstration of something that David Maher said over 40 years ago-- I forgot to put the quote in here, but it's something like, "The nature of the computations that underlie perception will be determined much more by the structure of the task that has to be solved then by the hardware that's implementing that solution," and I think this is a beautiful demonstration of that idea.
And it tells us that, as people have said before already today, that neuroscience and artificial intelligence have much to learn from each other because of this convergence of solutions. So as Josh and others have said, this is simply pattern recognition. This is a relatively simple problem, a teeny part of the problem of intelligence.
So how do we build on this success to take the next steps? That is, to not just know the categories of things that are in front of us, but to do stuff in the world? And so to do that, we need physical scene understanding. So what do I mean by physical scene understanding?
Well when you look at a scene like this, the point is that you see not just the categories of objects that are there. You see that the glass is supported by the table, that it contains orange juice, that it could be picked up. But might spill, staining the placemat. And you see all of that in just a few extra hundred milliseconds after that first feedforward pass delivering the categories of objects.
So that's what I mean by physical scene understanding. And so to act on the world, you need that. But then you also need to take this rich, physical understanding of the scene in front of you and figure out, what sequence of actions you can take to complete some goal like setting the table.
And that entails answering some really hard questions like, how can I move the glass without spilling? If I pull on the placemat, will the glass come along? How should I look for a plate? Start at the closest place? Does the human I'm serving even still want breakfast? And what should I do about that alarm I hear?
So these are really, really hard problems. But encouragingly, our team has already laid some of the foundation for this kind of work. So Leslie and Tomas Lozano-Perez have made this cool robot here. It does cool stuff by representing the belief state of the world and making these long horizon plans under uncertainty.
Josh Tenenbaum has devised these computational models of physical scene understanding in which you start with an image, extract a representation of the physical structure of the scene, use a forward physics model that predicts time point by time point what's going to happen next. And he's tested that model against humans performing the same tasks and found remarkably good fit between his model and what humans do.
Vikash Mansinghka has developed these amazing computational models of 3D scene perception that can segment objects in the image, figure out their 3D shape, and figure out what's touching what.
In collaborative work between me and Josh, we've identified some candidate regions in the human brain that seem to be implicated in physical scene understanding, telling us where to look in the brain. And Mehrdad Jazayeri, who'll be speaking in a couple of minutes, is looking at the neuronal basis of object tracking and physical intuition.
So to take the next step together, we're going to tackle a concrete but broad class of problems that's known as generalized fetch. That means find and retrieve or rearrange specified objects in an unknown environment doing whatever it takes. And why is that a good problem class? Well, for a number of reasons.
One, it's actually a huge space of problems, and we want to have one solution that solves all of them. Second, it's a classic embodied real-world problem. You have to go all the way from perception through to action. You can't cheat and skip a step. And third, it far exceeds the current state of the art in both what we understand about brains and what we can do in machines.
And finally, it has a useful, simple space to start from which to build. And the simple space we're going to start from is a task called "Get the Grape". And this is a very natural task for humans, for monkeys, and for robots. So we're going to do this in all three of them. And we'll start with a hierarchy with some very simple tasks, scaling up from there, from trivially simple to more complex.
And for the human version of this task, we're using a simulation platform devised by engineers here at MIT called ThreeDWorld, and it can create these beautiful very realistic renderings of worlds, and it can simulate physically what's going to happen in those worlds.
So for example, if you perform an action on it, it will give very realistic forward physical simulation. But also importantly, it interfaces with virtual reality so that we can put a human in a VR space and give them an action task and record all of their behavior from eye movements through trajectories of hands. And we can measure how people perform these find the grape tasks.
And so we're starting with the simplest and most rudimentary versions of this "Get the Grape" task. And to do that, we're going to show you a very simple example of this, and we're going to ask Jim to get the grape, performed by Aryan Zoroufi who's going to put him in VR and get him to find-- see if we can find-- he can find the grape.
So we designed it so there's a couple of very basic rudimentary tasks and then a more complex one. We'll see if Jim can do that. That he hasn't done before.
JAMES DICARLO: Yeah, this is scary. I'm scared that the director of the Quest for Intelligence is going to look not intelligent.
What am I doing?
NANCY KANWISHER: So this is what Jim is seeing.
JAMES DICARLO: Wow.
NANCY KANWISHER: So he's got to get the grape.
JAMES DICARLO: Grape is somewhere. The thing doesn't move.
NANCY KANWISHER: So he tries moving, the first thing, it doesn't move. He was there-- great. All right. Go, Jim! Woo!
OK. OK, that was pretty basic. Let's do another one.
JAMES DICARLO: Basic? Oh, I thought I succeeded. OK.
NANCY KANWISHER: No, that was the easy one, Jim.
JAMES DICARLO: All right. That was the easy one? OK.
NANCY KANWISHER: Yeah.
JAMES DICARLO: Do I clear this?
NANCY KANWISHER: What's happening? It's taking-- there we go. Here we go.
JAMES DICARLO: Ooh, there's stuff. OK. All right. Are you guys seeing all this?
NANCY KANWISHER: Jim, you're making a mess. Throwing stuff all over. Clean up your room!
JAMES DICARLO: I thought I was supposed to be like a four-year-old. OK. Oh, there it is. Yeah, that's no problem.
NANCY KANWISHER: All right. Go, Jim!
NANCY KANWISHER: Yeah. All right.
All right. The next one we designed to be more tricky. It involves some significant innovation. So who knows if Jim will be able to do this. He hasn't seen it before.
JAMES DICARLO: OK. Ooh. Ah. OK. Oh, that's-- all right. Ooh I think I know what you're trying to get at here. Aw! Yeah, yeah, yeah.
NANCY KANWISHER: Oh, smart. Woo! All right.
All right. Here we go.
JAMES DICARLO: --another one?
NANCY KANWISHER: No.
JAMES DICARLO: Oh, it's over there.
NANCY KANWISHER: OK. Great. Thank you, Jim. That was wonderful, you did great. A-plus. So the question is, how did Jim do that? That was amazing, that was awesome. But there's a huge space of possible computational models for how he might have done that.
And so by collecting rich behavioral data, we'll be able to prune this tree of possible computational models, but only partway we'll still be left with lots and lots of possibilities of what algorithms might be running in his head when he's doing this task. And that's why we need Mehrdad and his monkeys.
MEHRDAD JAZAYERI: --machine learning in limited domains. But when it comes to real-world tasks and real-world environments, we don't have good solutions currently. And the monoculture is such that we might not actually easily find a solution if we just continue on that path.
Well that said, we know that there exists one guaranteed path to a solution, and that's the brain. We know the brain offers the kind of intelligence we're looking for. So if we were to understand how the brain solves it, we would have the solution in our hands. In my lab, we work on non-human primates, macaque monkeys to study their brain to try to understand how they develop intelligence.
Now for those of you who are not familiar with this kind of work, you might think like, what do monkeys have to offer to this process? And I want to share with you three points. Point number one is that these monkeys actually are quite intelligent.
Now I'm going to show you a little fun video of monkeys trading objects with folks that are nearby. You will see them handling objects, but more importantly, you see that they are indeed intelligent.
Now what you will see is that these monkeys steal the glasses and various objects of people, and they go somewhere, and they wait until the person figures out, OK, I need to give this monkey something. They don't like the glasses, of course, but they want food, they want bananas, they want something else.
And what happens is that they wait for a while, and until they basically-- they're testing the human intelligence there. That, are going to figure it out that I want your banana? And at some point people figure it out and they give their banana and then they quickly get it and give back the object.
And they do that-- they wait until they get what they want. So if you give them what they don't like, they refuse. It's like, no, I'm holding the glasses. I want that other thing, I want the candy. And they do that and they get it.
So you get a sense that they're quite intelligent, they understand social interactions, and they can handle objects. So that's point number one, they're quite intelligent. So we can study intelligence in these animals.
Point number two is that we think we have found the secret sauce of how to-- well, this was in their natural habitat. But we think we have found a secret sauce of how to bring this kind of intelligence into the lab setting. Any guesses on what that secret sauce is?
Well remarkably, it's actually embodiment. So what we have found is that when you put these monkeys in front of a computer monitor, a screen and a joystick and such, they often find contrived solutions to the problems that they don't generalize well. They're not quite flexible and they're not able to really take advantage of their knowledge and intelligence.
On the contrary, when you put them in front of a tabletop with real objects, they quickly reveal that they understand objects, they understand the relationship between objects, and they know what they can and cannot do to those objects. So they very quickly express what over the evolution they have learned to do in the settings that are real.
So we're going to have a version of "Get the Grape" for them with physical tables and physical grapes and physical objects, and what we hope to do is to use that setting and the animal's natural embodied intelligence to ask questions, fundamental questions about intelligence like how does the brain represent objects? How do they make causal inferences? How does it make plans to interact with the objects and revise those plans?
And the plan is to, of course, while the animals are doing this, to record from the brain. But the setup of "Get the Grape" for monkeys introduces some major challenges. First of all, as Leslie and Nancy mentioned, this is an integrative problem. It requires multiple faculties-- perception, memory, cognition, action, and that means that many parts of the brain are going to be involved. So we need to be able to record from multiple parts of the brain simultaneously.
In addition, a hallmark of intelligence is that you learn very rapidly and you can very quickly generalize to very novel conditions. And what that means is that we ought to be able to record from many, many neurons simultaneously within single trials. We need to capture the animal's mind as it's thinking about a new condition right as it is doing that.
And that brings me to point number three. As you see, in early 2000 when I was a student, we could record from a handful of neurons simultaneously. In the past two decades, we have come a long way. That has increased by two orders of magnitude. And currently in my lab we're setting up to record from a few thousand neurons simultaneously.
And we think that puts up puts us in the right ballpark to be able to begin to understand the basic building blocks, algorithms that the monkey's brain might be using to handle objects within the tabletop.
So what do we want to do? We want to record from these neurons because we think we can then analyze them and extract from them certain various hypotheses and possibilities of how the brain solved this.
For example, if we think about the object representation, we can look at the neural activity to distinguish between whether the object representations are retinotopic, which means the objects are tied to space, or object files, which is inspired by theories in cognitive neuroscience and computer science for representing objects.
And for example, similarly, we can look at the structure and dynamics of neural activity as the animals thinking and interacting with the objects to infer the strategies that they might be using for making causal inference.
And if we were to find one of these turns out to be the situation-- so for example, for object representation, if it turns out that the monkey's brain uses a representation that is more aligned with object files compared to retinotopy, that will allow us to further prune, further constrain the search to subbranches of this tree and help the larger trees that are trying to figure out how to build models out-- how to integrate to make a much more efficient search and hopefully make the curve that Leslie showed you, bring it further down and down so quickly we get to the answer.
And with that, I guess I'll pass it on back to Leslie to tell you how we hope to use this information to engineer next-gen robots.
LESLIE KAELBLING: All right, so here we go. What do we do? We started with a big hypothesis space. We pruned it based on behavioral data, let's say, from humans. Now we can prune it further based on actual neural recording data for monkeys. So maybe now we're down to a smaller set of hypotheses about how things might go. Now we can try to test some hypotheses in a robot.
And so for instance, I might take a particular class of algorithms-- in this case, I'll talk about algorithms that plan in belief space, that where we divide the problem of thinking about the world into some building a representation of what we know about the world and actually explicitly our own uncertainty, and then the problem of taking that uncertainty and deciding how to behave. So that's a class of controllers for robots.
And then I can actually go further and take a specific one, so a particular implementation or a particular instance of that category of algorithms and gather data. So here, we have the robot in the same setup. It's looking for a grape. It tries to move that object and it can't. It tries to move this other object thinking that the grape might be underneath there. It moves it carefully so that it doesn't disturb whatever is underneath.
And it looks to see now, we simulate looking right now at the moment, simulated and we find another object there. Not the grape, but the object could still maybe contain the grape, so the robot is reasoning about where-- what parts of the space has it not yet seen. And it explicitly plans to look in places that it hasn't yet seen.
And these glosses that we have in the movie are actually-- we can find nodes in the search space. We could build visualizations that would help actually understand what the robot's thinking while it's doing this. So it finally picks up the last little cup and it finds the grape, too.
The robot is probably not as happy as Jim or the monkey about having found the grape, we have to work on that part, but still, if we tell it to look for a grape, it will. OK. So we can then, having run that, we can test different variations on this. We can test strategies that prioritize looking in different kinds of ways,
So we can study, we can generate more constraints, make more refined hypotheses, take those hypotheses and use them to design better and different experiments in the humans and the monkeys and help refine our understanding of, for instance, how one might approach this problem.
OK, so that very simple get the grape is very simple, but it gives us just a nucleus from which we can expand a lot. So we might be interested in more general kinds of manipulation. So we see animals doing rich kinds of manipulation, using tools, building tools, studying memory and learning what happens if we show you the grape and then take it away-- take the scene away for a while and come back. So all of these things we can study even just in the context of this "Get the Grape" problem.
And there's lots of really interesting underlying algorithmic choices. They're interesting to me as a computer scientist and they're interesting in general as questions of intelligence in cognitive and brain science. So things like the question we were just asking about was maybe is the representation in memory image-centric or object-centric?
But we can study things like how do we organize the search for an object? Is it systematic? Is it greedy? Is it random? When should we gather more information versus when should we try? Is it worth looking around at the back of something to see what it's like before we try to move it or should we just move optimistically?
Or how frequently should we reevaluate our higher-level plans and intentions? So in our computer implementation of the controller for that robot, it has a whole hierarchy of plans, and given evidence, it might decide to just reconsider its whole strategy. But how often should it do that, for instance?
How should we do data association? When we see a new object, how do we decide whether that's really a new object or actually just a new view or a slightly moved version of something before? How do we organize our representations of space? How do different aspects of behavior improve with experience?
So the list of questions is just enormous and it's super exciting because we're poised to study these questions now in all different modalities. So now we can scale up even further. So finding the grape in a tabletop is interesting, but hardly the whole world of embodied intelligence.
So one of the things that we're working on also right now in parallel is increasing the spatial scale of the problems we think about. So we have an initial collaboration with Ila Fiete who's going to be speaking after me, studying navigation and how it is that in rats and in robots, we can build maps that have a hierarchical structure and the degree to which that might help navigation be more efficient.
We can compare that to other kinds of methods that build maps that are more globally accurate, more metrically accurate, and this is work with Nick Roy. So we can extend to bigger spatial scales. We can also just extend this problem in a myriad of ways.
So we can study the amount of knowledge that we have in advance, we can make the objects be non-rigid, we can really study the question of how partial information changes your strategy, increasing the spatial size, making the task complexity go from finding a grape to assembling something, building IKEA furniture, putting away the groceries. We can vary the variety of situations and goals that we ask for.
We can start to connect to our colleagues working on language and think about how it is that we interact with humans and other agents and how can we interact linguistically and non linguistically in useful ways. And again, the questions just grow from here.
So fundamentally, though, the important thing is that we don't want to solve each subproblem individually. That is absolutely not what we want to do. We want to understand how to build one general purpose solution to solve this whole class of problems because we think if we could do that-- again, we'll come back to that bet we're making.
But if we could build systems, robot systems, let's say, that can solve such a big class of problems, then it's most likely that they would share some underlying structure and technique with the way that humans solve those problems, too.
So let me just close by kind of recapping the kind of impact that we hope to have. So there's practical impact in building robot systems that can go beyond very prescribed factory environments and out into the messy, messy world that we all live in. There's scientific impact. There was the grand rainbow of scientific impact that I talked before, but there's also a more near-term scientific impact.
In particular, the hope here-- and clearly, actually, the reality is that by having this close interaction between the cognitive science, brain science, computer science people, we can really come up with new classes of system models as potential hypotheses about mechanisms for all kinds of human cognition, including perception beyond the very first few layers, mental representations of physical scenes, memory organization, how you reason about the physical world, how you make plans at different levels of abstraction and so on.
And the idea is that once we have these hypotheses, then we can design more focused neural and behavioral experiments. Instead of just doing an experiment and looking at the data and going, hmm, I wonder what about that, we can be really much more focused in terms of trying to discriminate between the hypotheses.
And we can maybe improve our computational implementations, and those computational implementations will demonstrate a certain kind of behavior that we can also go look for in the natural systems.
So we think that this process, this virtuous cycle of going around between constraints that we get from experiments in nature constraints we get from implementation, new hypotheses and so on will give us really new insights in all of these different realms.
We also hope to have a real impact on society. One important type is the idea that if we build robots and software agents that have, in their underlying structure and processes, a similarity to humans, that these systems might be easier and more natural for humans to interact with.
We also think that if we build these systems that have some amount of engineered structure in them, it will be easier, actually, to understand and verify them, to trust them, to believe that we understand what they're going to do in more complicated situations, because we can actually look inside the implementation and see what's going on as was true with our robot that was looking for the grape.
The robotics people like to talk about tasks that are dirty, dull, and dangerous. So that's a watchword for a category of jobs that humans find unpleasant or difficult and that maybe we can make robots take over some amount of that work, subject to a bunch of concerns that Jim laid out earlier.
Building assistive technology that could help individuals at a variety of different levels in a variety of ways, going all the way to brain-machine interfaces that can really fundamentally help people who are fundamentally challenged at sensing and control.
And so with this, with impact in the practical world, in the scientific world, and in social world, we think that embodiment will really move the Quest for Intelligence forward. Thank you.