Why it’s simpler for AIs to write about art than for robots to do your laundry
A robotics CEO tells diginomica why teaching an AI about the Renaissance is trivial compared with sorting your dirty socks.
A sanctuary is a refuge from danger or persecution. So, ‘Sanctuary AI’ is a bold name for a robotics and artificial intelligence company that aims to build a new generation of AI-powered, general-purpose humanoid robots to work in factories, warehouses, and other industrial settings.
That’s because Sanctuary AI robots appear to be what science fiction writers have warned us about: intelligent machines that look like us and take away our jobs. Indeed, this narrative trope has been with us for as long as the word ‘robot’ has (coined in the 1920 play, Rossum’s Universal Robots, in which an uprising of artificial workers leads to humanity’s extinction).
But the Canadian robotics company, which has kept a low profile until recently, sees the world very differently to those heralds of sci-fi apocalypse, as it begins to train and roll out its machines into our fragile human world.
First, the ‘robots in, humans out’ scenario has always been a dramatic oversimplification. Robotics, AI, automation, the IoT, autonomous transport, and other Industry 4.0 technologies create new jobs, services, markets, and companies – like Sanctuary AI – as well as automate tasks.
Second, human employment in highly automated countries has generally remained high – the world’s current economic problems notwithstanding. Ironically, many of the West’s recent large-scale job losses have been in the IT sector itself, rather than across broad swathes of industry. (That may change, of course.)
Third, demographics are certainly changing in mature economies, where ageing populations in ageing cities mean that extra assistance will be needed to keep people alive, cared for, and supplied with goods and services.
And fourth, humans don’t want the jobs that Sanctuary AI’s robots are designed to do, claims co-founder and CEO, Geordie Rose.
At present there are 10.8 million unfilled jobs in the US alone, according to March 2023 figures from the Bureau of Labor Statistics. So, as a general principle, staffing unfilled roles with robots would not put millions of humans out of work, he explains.
But surely the large number of vacancies in a given month doesn’t prove that people don’t want those jobs? Rose says:
The question of why there are millions of unfilled jobs is complex. One of the most powerful forces is demographic change. Birth rates are decreasing worldwide. In some parts of the world, they're shockingly low, like in South Korea and Japan [two of the world’s most highly automated countries].
So, the number of people who are of working age declines as a fraction of the population, but at the same time the amount of work that's needed to support that population increases.
One of the main side effects of that is choice. If I'm a person of working age today, I have the kind of choice that no one in human history has ever had in terms of how I spend my time working. So, there's an intersection with demographic change: a lot of work that’s out there, people just don't want to do.
Most of our customers are large companies that employ tens of thousands of people. And there is a hierarchy of tasks in those places. Some are just not desirable; they’re the things people avoid doing. They're not glamour jobs, they don't pay a lot, and they’re not fun to do.
Those are the ones we're going after. And we do it empirically by asking the workers in those sites, ‘What do you not want to do today?’ Often these tasks are repetitive, boring, dull… the same things over and over. But those tend to be the easiest things for machines to do well.
A good point. As Dr Anders Sandberg of Oxford University’s Future of Humanity Institute once observed, if you can describe what you do for a living in simple terms, then your job can – and will – be automated.
A synthesis of human knowledge
However, the notion that AI, robotics, automation, and other Industry 4.0 technologies will save humanity from boredom, thus freeing up workers to focus on being creative, has become an industry cliché. And there’s a big problem with it.
The surge of interest in ChatGPT and generative AIs, for applications such as content generation, writing, editing, illustration, design, photography, video, and even music, has proved it to be a fallacy almost overnight. Companies are falling over themselves to use AI for tasks that, until now, have been the preserve of talented, expert, creative humans.
The reasons for this have been explored in a recent diginomica report. But the critical one is that the arts have been low-hanging fruit for AI’s designers. The likes of OpenAI have been able to scrape vast amounts of data off the Web with impunity and use it to train their systems.
But what has this got to do with robots? Rose makes a point that is equal parts clever, depressing, and demonstrably true:
It might appear that simple things like picking up a cup would be easier to achieve technologically than generating an essay on the Renaissance. But it turns out that's not true, and that's counterintuitive to most people. The reason is humans can pick something up without thinking about it, but what we're doing is very difficult for a machine.
Humans have a long lineage of evolutionary help to be able to do things that we all take for granted: the simple act of looking around you, knowing what things are, and being able to pick them up and use them. But those are beyond the state of the art for AI and robotics at present. And that is difficult for people to understand.
This is because, however advanced or dextrous they may be, robots and AIs have no concept of what anything in the human world actually is. To a robot brain and camera, a simple object like a cup is just a collection of pixels. And the same cup viewed from different angles is yet more pixels. And then there the millions of cups that look different, thanks to human ingenuity.
You get the picture. But robots don’t: that’s the point. Teaching a robot that a cup is a cup, let alone training the machine to put it on a shelf, is an incredibly complex engineering and programming task. And that’s in spite of all those videos of Boston Dynamics robots running, jumping, and doing somersaults.
AIs have no concept of what the Renaissance is either, of course. But millions of humans do, and have shared their knowledge about it online. In this sense, epochal AIs like GPT-4 are neither artificial nor intelligences, but merely a synthesis of the vast amount of human knowledge that is out there on the internet.
So, writing an essay or a speech is easy for a machine, though humans find it hard. But the reverse is true when it comes to the manual tasks that we all take for granted, such as using tools. The very thing that first made us human. Rose says:
The automation of all the things that people can do without thinking is not within sight of modern technology. Not now, nor in the near- or medium-term future. But, for example, the movement of goods from place to place is more like low-hanging fruit. Those are the kinds of tasks that you can automate with technologies that are within sight.
The delivery of work
For Sanctuary AI, industrial and business-to-business services are the focus, rather than, say, the domestic market targeted by Tesla for its Optimus robot, for example.
Eventually this type of robot will be adopted more broadly. But domestic applications are not the right things to focus on first. You will fail if you go after home robotics now, simply because the types of things you would need to do with these technologies are far beyond the horizon of what the industry can do in practice.
Then he adds:
Sanctuary AI has been able to deploy a robot in the wild, in a store – an environment that we didn't control – and the technology worked just fine. But it’s still just a store, and they have certain properties: wide-open floors that are accessible and everything is on the same level.
But houses and apartments are different. And the kinds of things that people might want robots to do in their homes are among the hardest things that people do, but without thinking about them.
Doing your laundry is much harder for a robot than anything you've heard of an AI doing, including all that apparent GPT ‘magic’. Those things [GPT’s achievements] are trivial compared to building a robot that could do your laundry on its own. That's a hard thing to accept, but it's true.
The robot deployment to which Rose refers took place earlier this month in a branch of Mark’s, the industrial clothing chain owned by the Canadian Tyre Corporation. A robot trained at Sanctuary AI’s own offices worked, experimentally, in the client’s premises and apparently performed well.
Now the company’s strategic aim is to create what it calls “human-like artificial intelligence” and put that into robots that are capable of carrying out complex, repetitive tasks. But what does Rose mean by that? He explains:
It's more specific than artificial general intelligence. Our conception of how to approach the problem of building a human-like mind is to create robots that are designed to do work. And that means work in general, and not just the sorts of things that would typically be automated. All the interstitial things we take for granted.
For humans, our conception of the world is a model, an inner world. There's a thing that exists inside you that is a model of the real world, and it contains all your beliefs, memories, understanding, and so on.
So, for a robot to have a human-like mind requires an understanding of the world that persists over time, that it builds as it learns. We want it to be able to deliver work – everything from sorting socks to making plans about where to put things in stores, to being able to sweep the floors.
All of the things that civilization actually runs on from day to day require enormous amounts of cognitive capability.
An unclear timeline
However, the videos that are currently online of Sanctuary AI’s robots in action – impressive-looking machines that are humanoid in appearance, but with an industrial solidity – generally show them being tele-operated by remote human ‘pilots’.
Those are significant engineering achievements, but a far cry from the claims of autonomy and human-like intelligence that Sanctuary AI makes. So, at what point does the teleoperated robot become an autonomous robot with human-like intelligence? In short, when will the human step out of the loop?
It happens gradually. One analogy is cars, and the transition from driver-assistance technology to autonomous driving. In the beginning, the car is a passive machine that's operated entirely by a person. Then a series of changes happen in the underlying technology that make it easier for the person to drive. Then ultimately, performing the entire task under the direction of the person occurs, and the car can do it autonomously.
But in the intervening time, there is a series of steps. That's the type of augmentation that comes from an AI.
In the systems we build, that analogy is pretty accurate. Teleoperation of the robot is like driving a car, with the robots being passive recipients of the pilot’s actions. But over time, aspects of that get automated, so it becomes easier for the person to direct the robot.
The end of that journey is the person instructing the robot simply by speaking to it – ‘Arrange all these objects on that shelf’. And the robot will be able to understand what that means, and be able to execute the task to your satisfaction.
How long will that future take to arrive?
That’s not entirely clear. Some things are difficult, and it's unclear if some of them even can be automated, because the target we've set ourselves is to build something that's very much like a person.
Some of the things that humans do are magical, if you like, and are still not well understood. So, there's a gradient between the two. Between the work that can be done and the magic of the human mind.
Humanoid robots have long been a problematic concept: a sci-fi idea that has slowly become a reality, of sorts, in the century since it appeared. Today, some are brand ambassadors or pop-culture icons, but others are little more than engineering achievements that lack utility: examples of our hubris as a species, perhaps.
Why some robots need to be humanoid in appearance can be answered by the need for our future machine helpers to exist in a world designed for, and by, people. But what such robots might actually be for is another question entirely; one that Sanctuary AI has come closer than most to answering.
Time to pass the cup?