快猫短视频

WHO鈥橲 in charge, your brain or your聽body? The answer may seem聽obvious, but there is plenty of聽evidence to suggest that our physiology has聽dramatically affected the way we think. This idea of embodied cognition could hold聽important lessons for those trying to聽build genuinely intelligent machines聽鈥� artificial intelligences that learn and think and聽can generalise their knowledge to all聽manner of聽tasks, like humans.

, a roboticist at the University of Vermont, is among those who insist AIs will聽only fulfil their promise if they can directly experience and interact with the physical world. That is a far聽cry from AIs like ChatGPT, whose only interactions with the world come via the abstract medium of language. But the聽field of embodied AI is pushing for the convergence of聽artificial intelligence and robotics, and Bongard is聽at its forefront.

He reckons we need to rethink our approach to both disciplines. Simply integrating an AI chatbot with a robotic arm, as Google has done with its PaLM-E system (pictured below), may not be enough. Instead, Bongard focuses on 鈥�evolutionary robotics鈥�, which leverages the principles of natural selection to rapidly iterate through robot designs, many of them made of soft materials. He is also part of a team using living cells to form simple biological robots, known as xenobots, that not only perform basic tasks, but can interact with and respond聽to聽their environment too.

Here, Bongard tells 快猫短视频 how this work is suggesting entirely new ways to think about embodied cognition, and what counts as a robot, which could transform our approach to building intelligent machines.

How AI is shifting the limits of knowledge imposed by complexity

From weather to the structure of proteins, some things are predictable in theory, but too complex to figure out in practice. But the rise of artificial intelligence is changing that fast

Edd Gent: How does the body influence the way we think and learn?

Josh Bongard: There鈥檚 little in the way humans understand the world that isn鈥檛 directly or indirectly influenced by our physiology. We use our bodies to push against the world and we use our sense organs to observe how the world pushes back. This feedback loop is how we learn about the world. It also means that different bodies generate different action-reaction results and thus learn about the world differently. For example, a big, land-based animal should see high cliffs and water as dangerous.

But flying animals may see a cliff as a good launching point, and insects see water as something that can be walked across.

Some of the simplest clues to that connection between the body and the way we聽think are hidden in language. Since humans have forward-facing eyes, when we walk, we are looking at objects and situations that are about to become our future.

This means that humans tend to 鈥渓ook ahead鈥� to events weeks or years in the future, or 鈥渓ook back鈥� with regret or nostalgia on past events. Even in abstract ways of thinking like mathematics, we talk about manipulating an equation. The root of the word manipulation is about hands, and a lot of mathematicians report visualising themselves pulling apart and recombining mathematical concepts with their hands.

What lessons does that hold for people trying to聽create artificial intelligence?

That connection between the body and the way we think is something that is not always obvious to us, and something that we haven鈥檛 really built into our machines. In my opinion, that is one of the missing ingredients from all these very powerful, yet surprisingly flawed, intelligent technologies: the lack of a body and the lack of understanding of the connection between the real world and the world of intelligence, abstraction, reasoning and cognition.

You can sense that some of these AIs have fundamental flaws. There are gaps in their knowledge and understanding about the real world. Those of us that work in embodied AI, we point to the fact that ChatGPT has never had聽the ability to push against the physical or聽social worlds and see how they push back.
Also, current AI is good at generating things such as images and videos and responses to questions, but it can鈥檛 make anything physical.

It has no access to the real world. It doesn鈥檛 know what works and what doesn鈥檛 work. This聽is an important aspect of intelligence. It鈥檚 relatively easy to dream up new ideas, but the value of聽intelligence in humans is that you can demonstrate that they work in the real world.

Is it enough to hook up a large language model聽(LLM), the kind of AI system behind the聽likes of ChatGPT, to a robotic body, as聽Google聽has done with PaLM-E?

I think nobody knows at the moment. There is this growing awareness that we need to embody LLMs or we need to bring AI into contact with the real world somehow. PaLM-E is one way to do it, which is to basically make a huge brain in a vat and then hook it up to a complex physical machine聽 [in this case an articulated arm with a kind of grasping claw capable of picking things up]. Whether or not that鈥檚 going to work in the long run is anyone鈥檚 guess, but it鈥檚 definitely not the way Mother Nature does it.

She starts simple. Most animals begin as single cells, which then divide and become specialised such that the animal itself grows聽gradually more complicated, both on聽the brain side and the body side.

And there聽are good reasons for that. When we鈥檙e teaching a child to聽ride a bike, we don鈥檛 put them on a full-sized bike to start with. We put聽training wheels on the bike, we make the聽task easier for the learner. Your own development is your training wheels. Making聽very sophisticated bodies and brains聽and then聽mashing them together and聽crossing your fingers, I don鈥檛 think that鈥檚 going to be a聽recipe for long-term success.

What kind of bodies should we give robots?

This is part of the fun of this next stage of creating intelligent machines. If the body is important for acting intelligently and safely in the real world, what body is the right body for a given task? In the very early years of aviation, most of the ideas were: 鈥淪hould we make a large thing that flaps its wings or should we make a small thing that flaps its wings?鈥�. Of course, the Wright brothers鈥� answer was neither. Nature has been incredibly creative in all the physical forms she鈥檚 discovered, but we don鈥檛 necessarily have to stick with what she came up with.

I work in this space called evolutionary robotics, where we ask AI to design not just the brain of the robot, but the physical structure of the robot聽鈥� the right brain-body combination for the job at hand. Most robotics folks are building the products of evolution, but in evolutionary robotics we鈥檙e trying to build the聽process of evolution. And that process produces new kinds of robot bodies and brains unlike anything we鈥檝e seen in nature before.

Evolution takes a long time, so you rely on simulations to accelerate it. That works fine for聽simple robots and simple environments, but聽is聽it聽feasible for more complex problems?

We don鈥檛 necessarily need to do evolution the way it happens in the real world. Biological evolution is blind: it makes progress through random mutations. But we published a in October showing that you can reduce the error in the trial and error. You can create an AI that watches its robot designs in simulation and when the design doesn鈥檛 do what we want, the AI can backtrack into the body of the robot and identify exactly where things went wrong.

We showed that this can speed up robot design in simulation from two weeks on a supercomputer to 30聽seconds on a laptop. The resulting algorithm is still designing relatively simple robots, the sort that kind of shuffle around a little bit. But if we can design simple robots in 30聽seconds on a laptop, now we can start to think about designing more complex and capable robots using this process.

What might those robots look like?

A lot of my colleagues work in this relatively new field known as soft robotics. So, instead of making robots from metal and ceramics, why don鈥檛 we make them from rubber and silicone and soft things? Why don鈥檛 we build them with hollow chambers inside, where we can inflate those bladders to increase the size of the robot?

Soft robotics is a really promising real-world聽technology that allows us to make robots聽that are safer, just by the simple fact that聽they鈥檙e soft. But being soft also means you聽can deform and change shape and change size. You can extrude arms and legs and fingers and tools, and then retract them when they鈥檙e no聽longer needed. So, I think there are a lot of聽advances that are occurring in material science that allow us to make robots from more exotic materials, mixtures of soft and聽rigid components.

In recent years, you鈥檝e also been working on聽building robots out of living materials. How聽is聽that work related to embodied AI and聽evolutionary robotics?

I鈥檝e been working with at Tufts University in Massachusetts, who has shown that animal body plans are more open to suggestion than we previously thought. What I mean by that is that you can surgically rearrange tissues in an animal embryo and, in many cases, it will grow into an adult organism that looks and acts very differently from the wild type. For example, Mike showed that you can surgically rearrange the eyes in a frog embryo so that it grows into an adult frog where the eyes are on the back.

About six years ago, I teamed up with the Levin lab to ask: 鈥淚f a human biologist can come up with a way to rearrange living tissue to produce a new kind of organism, could we get an AI to dream one up?鈥�. We asked the AI to聽put frog skin and heart cells together in a聽particular pattern that would produce something that walks along the bottom of聽a聽Petri dish. The heart tissue increases and decreases in volume, so it can act like little pistons or little motors. We set the AI to work for about two weeks on a supercomputer and it聽tried out millions of different arrangements.

At the end of that two-week process, it sent聽us back a short video of a virtual frog bot. We passed that video to a microsurgeon in Mike鈥檚 lab, who spent a few hours painstakingly putting cells where the AI said they should go.聽Then he let go of his little creature and this聽xenobot walked along the bottom of聽the聽Petri dish.

That鈥檚 an impressive feat, but are robots built out聽of biological materials anything more than聽a聽research curiosity?

This is still very basic science. That being said, I鈥檇 be surprised if AI-designed organisms do not eventually find applications in certain areas. They are very small.

There are a lot of use cases in which very small robots that go, look, remember and come back would be very useful, but it鈥檚 hard to make very small machines out of traditional components. They鈥檙e also biocompatible and biodegradable, so they would be a very eco鈥慺riendly technology.

What can they tell us about the role morphology plays in intelligence?

They鈥檙e already telling us many things about the nature of embodiment and intelligence. They show some of the rudiments of intelligence: they seem interested in certain things in their environment and they shy away from other things. But they鈥檙e just frog skin and heart cells, there is no brain.

The individual cells have never been in this particular arrangement before, but somehow they figure out what to do. They communicate with each other electrically, chemically, mechanically, they push and pull on each other. And they seem to work together to do what the AI wants them to do.

The unique promise of 鈥榖iological computers鈥� made from living things

Biologists are finally beginning to corral molecules, cells and whole organisms to carry out complex computations. These living processors could find use in everything from smart materials to new kinds of artificial intelligence

The intelligence resides in the cells. This is a machine made up of intelligent machines. And so it鈥檚 suggesting completely new ways to make intelligent technology. People often ask us, 鈥淗ow do you program these robots?鈥� and we kind of smile and say, 鈥淭hat鈥檚 the whole point.鈥� There is no programming. There鈥檚 no little biological computer inside these robots.

There鈥檚 no distinction between the controller of the machine and the machine itself. So maybe we鈥檝e been going about it wrong this whole time. It may be that what we typically refer to as the brain is a support structure, it鈥檚 not the command-and-control centre.

Mammalian brains are one of the most recent inventions Mother Nature came up with. Bodies have been around for a very long聽time, so it鈥檚 possible that most of the intelligence resides in the body, and brains are聽just there to facilitate or enhance that latent intelligence

Edd Gent is a technology journalist based in Bangalore, India