Does Artificial Intelligence Need a (Virtual) Body?
Embodied cognition might hold clues about what AI is currently missing to bridge some of the cognitive gaps that's holding it back
Embodiment
Your body determines how you exist as part of the world and how you experience your environment. That, in turn, affects the way you see things. The world looks different for someone who’s five feet tall than it does for a seven-footer. Reaching shelves, finding a place on public transport, buying furniture, the way people react to your presence… It’s all going to be different.
This is a silly example, of course, but it shows how having a specific body with specific characteristics shapes the way you experience the external world. Your cognitive processes will be influenced by that. Mind and body are parts of a whole. Down with Cartesian dualism!
How you perceive things, how you move, how you speak… All of that is partially ‘encoded’ in/by your body. That’s the general idea behind the embodied cognition theory.
The theory comes in a few different flavors that make stronger or weaker claims. This family of theories is sometimes referred to as ‘the four E’s’ - embodied, embedded, extended, and enactive cognition. (Although embodied cognition can also be used as an umbrella term to cover all four of these.)
If you want to dive deeper into this, check out the entry in the Stanford Encyclopedia of Philosophy. The key point for our discussion, as noted in the encyclopedia entry, is that:
…embodied cognition is better characterized as a research program with no clear defining features other than the tenet that computational cognitive science has failed to appreciate the body’s significance in cognitive processing and to do so requires a dramatic re-conceptualization of the nature of cognition and how it must be investigated.
So, there may be open questions and fuzzy edges, but having a body (probably) matters when it comes to cognitive processes.
Thank you for stopping by! If you like reading Thinking Ahead, please click one of these buttons - it helps me silence my inner critic and justify the time I put into this. Thanks!
Coding cognition (send in the robots)
What about artificial intelligence, though? If the only ‘body’ it has is a static, sessile, sensory-deprived, humming server stack, can it ever be truly ‘intelligent’?
You’ve probably heard about Google’s AI LaMDA (short for language model for dialogue applications) that made the news recently due to one of the company’s engineers who claimed the program is sentient. The AI research community was quick to point out all the flaws in his reasoning and in the program.
Here’s my take on it:
Likewise, the recent unveiling of the multi-purpose machine learning model Gato by Deepmind led some people to claim that all that’s left to do before we hit upon the holy grail of general AI is scale things up. Here too, a pushback rapidly followed. (A bit more about it in the May nuggets.)
(Read more? I’ve been on an AI kick lately. If you missed some of my earlier writing about it → baby steps of a giant, algorithmic culture, and (non?) creative AI.)
What if AI is missing a body? A sense of ‘being in the world’? To be fair, the Gato model does get permission to play with a robotic arm. Why stop there, though? Why not give it an entire robot?
I bet many of you had a similar image in mind as I had: a humanoid robot. How self-centered we meatbags are. (Interesting tangent: did you know most human-shaped robots are white? Bias is everywhere.) But robots come in different sizes and shapes; from AIBO over Sophia to ATLAS, robots imbued with machine learning are nothing new.
The real interesting question is: if AI needs a body, how do different body shapes shape its cognitive processes?
Let’s return to the robot arm. It’s pretty stationary, so I imagine that knowing its own reach becomes very important. Perhaps it will develop an uncanny sense of proprioception and distance to best manipulate its nearby environment?
Of course, we need to be careful not to attribute too many human-like cognitive traits to robots and/or machine learning systems. Only days ago, I came across this paper that considers:
…humanwashing of AI-enabled machines as a specific anthropomorphization notion.
There’s something else, something more fundamental, missing in my ‘AI needs a body’ thought experiment, though. If we already have robots with (some) machine learning capacity, where is the real AI hiding?
Teaching AI as a child (or a gamer?)
The missing thing relates to an observation I mentioned in a previous (not yet) creative AI newsletter: like AI creativity, machine learning in robots is task-constrained.
AIBO will learn to respond to verbal cues with specific behaviors, Sophia will attempt to mimic certain human behaviors and speech patterns, and ATLAS will learn to move over different terrains and withstand locomotor challenges (which is a nice way of saying ‘being kicked around’).
But beyond that, meh, nothing impressive to report. Perhaps that’s because these robots come pre-loaded with programming that has specific goals and endpoints. What if we leave it much more open-ended? A ‘go forth and explore’ kind of directive? This is somewhat similar to the idea that it might be worthwhile to instill a few basic instincts into a machine learning system and then basically teach it like a child to hone those instincts into a flexible and adaptable way of thinking and to interpret the wider context of the world.
Of course, we don’t exactly want curious robots roaming around and taking things apart to see what’s inside. (“Hey, what’s this mush inside this human?” Robbie thought while holding the guts of an unsuspecting passer-by…)
So what about a virtual environment? Specifically, a sandbox environment like the one used in open-world games. An AI training ground, if you will. Give the system an avatar with ‘physical’ constraints within the virtual world, a set of ‘physical’ laws it has to adhere to, and a varied environment with other agents but without specific goals beyond ‘explore and learn’.
And look, just as I am writing this, I come across new research by Deepmind, in which they use curiosity-driven learning in a virtual environment to:
… achieve[s] superhuman performance on the ten hardest exploration games in Atari while having a much simpler design than other competitive agents.
Perhaps that’s the job of the future: virtual teacher of budding general artificial intelligences. Slap on your VR goggles, step into the haptic suit, and teach a class of rambunctious baby dragons, centaurs, and unicorns. Yes, I get carried away sometimes, that’s my fiction writing side.
(Do read Ted Chiang’s brilliant, heartbreaking story ‘The Lifecycle of Software Objects’, in which digital entities or ‘digients’ learn, grow, and develop their own personalities through interactions with humans and each other. They even get robot avatars…)
I am aware that I have glossed over important points and that there are several holes waiting to be poked in what I’ve written. That’s good. That’s the point of a thought experiment.
I’m really curious to hear what you think about this (or, if you know anyone who’s into AI stuff, nudge them to read this. The more, the merrier.)
As always, thanks for reading! See you (virtually) soon.