AI circles around robotics

When people imagine the AI ​​apocalypse, they usually imagine robots. The assassin androids of the terminator franchise. The humanoid helpers of I robot. The Cylon armies of Battlestar Galactica. But the robot takeover scenario most often envisioned in science fiction isn’t exactly looming. The recent and explosive advances in AI, along with the recent and explosive hype surrounding it, have made the existential risks posed by the technology a hot topic. to integrate conversation. Yet progress in robotics, that is, machines capable of interacting with the physical world through movement and perception, is lagging far behind. “I can’t help but feel a little envious,” said Eric Jang, vice president of AI at humanoid robotics company 1X, in a talk at a robotics conference last year. And that was before ChatGPT came along.

Great language models write scripts, write code, and crack jokes. Image generators, such as Midjourney and DALL-E 2, are win art prizes And democratize interior design and produce dangerously convincing fabrications. They feel as if by magic. Meanwhile, the world’s most advanced robots still struggle to open different types of doors. As in real physical doors. Chatbots, in the proper context, can be – and have been – mistaken for real human beings; the most advanced robots look even more like mechanical arms attached to rolling tables. For now, at least, our dystopian near future looks a lot more like Her that M3GAN.

The counter-intuitive idea that it’s harder to build artificial bodies than artificial minds isn’t new. In 1988, computer scientist Hans Moravec observed that computers already excelled at tasks that humans tended to regard as complicated or difficult (mathematics, chess, IQ tests) but were unable to match “the skills of a one-year-old child in perception and of mobility”. Six years later, cognitive psychologist Steven Pinker offered a more concise formulation: “The main lesson of thirty-five years of AI research,” he writes, “is that hard problems are easy and hard problems are easy. easy are difficult”. This lesson is now known as “Moravec’s paradox”.

The paradox has only been accentuated in recent years: research on AI is advancing at full speed; robotics research stumbles. This is partly because the two disciplines do not have the same resources. Fewer people work on robotics than on AI. There’s also a funding disparity: “The flywheel of capitalism isn’t spinning fast enough in robotics yet,” Jang told me. “There’s this perception among investors, based mostly on historical data, that the return on investing in robotics isn’t very high.” And when private companies invest money in building robots, they tend to accumulate their knowledge. In AI circles, on the contrary, open sourcing is – or at least was-Standard. There is also the problem of accidental breakage. When your AI experiment fails, you can simply reboot and start over. A mistake with a robot could cost you thousands of dollars in damaged hardware.

The difficulty of obtaining enough data, however, creates an even bigger problem for robotics. Training an AI requires large amounts of raw material. For a large language model, this means text, an abundant resource (for the moment). Recent advances in AI have been fueled to a large extent by training larger models with greater computational power on larger datasets.

Read: GPT-4 might just be a bloated, useless mess

Roboticists inclined to this approach — hoping to apply the same machine learning techniques that have proven so successful for large language models — run into problems. Humans generate an immense amount of text as part of our day-to-day business: we write books, we write articles, we write emails, we send text messages. However, the kind of data that could be useful for training a robot – drawn, for example, from the natural movements of a person’s muscles and joints – is rarely recorded. Equipping masses of people with cameras and sensors is probably not a viable option, which means researchers have to collect data via robots, either controlling them manually or having them collect data autonomously. . Both alternatives present problems: the first is labor intensive, and the second gets bogged down in a kind of circular logic. To collect good data, a robot needs to be pretty advanced (because if it hits a wall over and over again, it won’t learn much), but to make a robot that’s advanced enough, you need good data.

In theory, a robot could be trained on data drawn from computer-simulated movements, but here too you have to make compromises. A simple simulation saves time but generates data that is less likely to translate to the real world; a complicated version generates more reliable data but takes longer to run. Another approach would be for the robots to learn by watching thousands of hours of videos of people in motion, taken from YouTube or elsewhere. But even those wouldn’t provide as much data on, say, how fine motor control works, Chelsea Finn, an AI researcher at Stanford University and Google, told me. In his speech, Jang likened computing to a tidal wave lifting technologies: AI surges to the top of the ridge; robotics always stands at the water’s edge.

Some members of the robotics community aren’t particularly concerned with catching the wave. Boston Dynamics, whose videos of canine and humanoid robots have been going viral for more than a decade, “uses virtually no machine learning, and a lot of it is sort of manually tuned,” Finn said (although that apparently will change soon). His robots are generally not very adaptable. They excel in performing a specific task in a specific environment. As impressive as they are, in that sense they are much less advanced than some of the more modest robots capable of opening different types of drawers. (Boston Dynamics did not respond to a request for comment.)

But the biggest obstacle for roboticists – the factor at the heart of Moravec’s paradox – is that the physical world is extremely complicated, much more so than language. Running, jumping and grabbing things can come naturally to people, unlike writing essays, playing chess and taking math tests. “But in reality, motor control is in some ways an inherently much more complex problem,” Finn told me. “It’s just that we evolved over many, many years to be good at motor control.” A language model must answer queries made from an unimaginable number of possible word combinations. And yet, the number of possible states of the world that a robot could encounter is still much, much greater. Just think of the gap between the informative content of a sentence, or even a few paragraphs, and the informative content of an image, let alone a video. Imagine how many sentences would be needed to completely describe the video, to convey at every moment the exact appearance, size, position, weight and texture of every object it shows.

Whatever its causes, robotics lag could become a problem for AI. The two are deeply linked. Some researchers are skeptical that a model trained only on language, or even on language and images, will ever achieve human intelligence. “There’s too much implicit in the language,” Ernest Davis, a computer scientist at NYU, told me. “There is too much basic understanding of the world that is not specified.” The solution, he thinks, is to have AI interact directly with the world via robotic bodies. But unless robotics makes serious progress, that’s unlikely to be possible anytime soon.

Improvements in AI could boost advances in robotics. For years now, engineers have been using AI to build robots. In a more extreme and distant vision, super-intelligent AIs could simply design their own robotic body. But for now, Finn told me, embodied AI is still a long way off. No android assassins. No humanoid assistants. Maybe not even HAL 9000, the greatest of science fiction’s AI antagonists. Set in the context of our current technological capabilities, HAL’s murderous exchange with Dave from 2001: A Space Odyssey would read very differently. The machine does refuse help his human master. He just can’t do it.

“Open the mod bay doors, HAL.”

“I’m sorry, Dave. I’m afraid I can’t do that.