AI revolves around robotics

When people envision the AI ​​apocalypse, they typically envision robots.android assassin terminator franchise.humanoid helper i robotCylon Army Battlestar GalacticaBut the robot takeover scenarios most often imagined in science fiction aren’t exactly looming. The recent explosion of advances in AI, along with the recent explosion of hype surrounding it, speaks to the existential risks posed by the technology. mainstream conversationHowever, progress in robotics—machines that can interact with the physical world through movement and perception—has lagged far behind. Eric Jang, his vice president of AI at humanoid robotics company 1X, said: Story At last year’s robotics conference, that was before ChatGPT came out.

A large language model writes screenplays, writes code, and jokes. Image generators such as Midjourney and DALL-E 2 art award and Democratizing interior design and production Dangerously convincing fabricationthey feel magicallyOn the other hand, the world’s most advanced robots struggle to open many kinds of doors. Like a real physical door. In the right context, chatbots can be mistaken for real humans. The most advanced robots look like mechanical arms attached to a rotating table.At least for now, our dystopian near future looks more like she than M3GAN.

The counterintuitive idea that it is harder to create an artificial body than an artificial mind is not new. In 1988, computer scientist Hans Moravec Observed Computers were already good at tasks humans tend to think complex or difficult (mathematics, chess, IQ tests), but they couldn’t match “the skills of a 1-year-old when it comes to perception and mobility.” Six years later, cognitive psychologist Stephen Pinker proposed a simpler formulation. This lesson is now known as the “Moravec Paradox”.

This paradox has become increasingly apparent over the last few years. AI research is advancing rapidly. Robotics research stumbles. One reason is that the resources in the two areas are unequal. Fewer people are working on robotics than AI. There is also a funding gap. “The flywheel of capitalism is not yet spinning fast enough in robotics,” Jang told me. “There is a perception among investors, based largely on historical data, that the return on investment in robots is not very high.” there is.On the contrary, in AI circles, open source is, or at least was-standard. There is also the issue of accidental corruption. If your AI experiment fails, you can reboot and try again. A mistake in a robot can cost you thousands of dollars in damaged hardware.

However, the difficulty of obtaining enough data poses even greater problems for robotics. Training AI requires a huge amount of raw materials. For large language models, this is text, an abundant resource (At this point). Recent advances in AI are greatly fueled by training large models with greater computational power on large data sets.

Read: GPT-4 May Be A Bloated, Pointless Chaos

Robotics bent on this approach who hope to apply the same machine learning techniques that have proven highly beneficial to large-scale language models face a problem. Humans generate vast amounts of text in the course of their daily work. Write a book, write an article, write an email, send a text his message. However, the kind of data that could be useful in training robots, such as the natural movements of human muscles and joints, is rarely recorded. Equipping large numbers of people with cameras and sensors is probably not a viable option. This means that researchers must collect data via robots either by manually controlling them or by having them collect data autonomously. Both options have their problems. The former is labor intensive, the latter falls into a kind of circular logic. To collect good data, the robot must be fairly sophisticated (it won’t learn much by just hitting walls over and over again). But to create a fairly advanced robot, we need the right data.

In theory, robots can be trained on data drawn from computer-simulated movements, but trade-offs must be made there as well. Simple simulations save time, but produce data that are unlikely to translate to the real world. Complex ones produce more reliable data but take longer to run. Another approach is to train the robot by watching thousands of hours of human motion videos pulled from sources such as YouTube. But even these data won’t provide much data about how fine motor control works, for example, Chelsea Finn, her AI researcher at Stanford University and Google, told me. In her talk, Chan likened computing to a tsunami lifting technology. AI is surfing the crest. Robotics is still standing by the water’s edge.

Some members of the robotics community are not particularly interested in riding the wave. A video of a dog and a humanoid robot he’s been going viral for over a decade in Boston. says Mr. apparently will change soon). The robot is generally not very adaptable. It excels at doing specific tasks in specific environments. As impressive as it looks, in this sense it’s far less advanced than some of the more modest robots that can open a wide variety of drawers.(Boston Dynamics responded to a request for comment.) did not.)

But the biggest obstacle for roboticists is the factor at the heart of Moravec’s paradox, that the physical world is much more complex than language. Running, jumping, and grabbing things may come naturally, but writing an essay, playing chess, or taking a math test generally doesn’t. “But in reality, motor control is in some ways an inherently much more complex problem,” Finn told me. “We’ve just evolved over the years to be good at motor control.” The language model has to answer queries made from an unimaginable number of possible word combinations. I have. Yet the number of possible world states a robot may encounter is much, much, much higher. Think about the gap between the informational content of a sentence or a few paragraphs and the informational content of an image, let alone a video. Imagine how many sentences would be required to fully describe a video and convey the exact appearance, size, position, weight and texture of every object the video shows, moment by moment.

Whatever the cause, lagging robotics can be a problem for AI. The two are deeply intertwined. Some researchers are skeptical that models trained on language alone or language and images can achieve human-like intelligence. Ernest Davis, a computer scientist at NYU, said, “Language leaves too much unspoken. There’s too much fundamental understanding of the world that isn’t specified.” He sees AI interacting directly with the world through a robot body. But unless robotics makes some major breakthroughs, I don’t think it will be possible any time soon.

AI improvements could help advance robotics. For years already, engineers have been using his AI to build robots. For more extreme and distant vision, super-intelligent AI can easily design its own robot body. But for now, Finn told me, AI is still a long way off. No android assassins. No humanoid helpers. Science He may not even be HAL 9000, the greatest AI antagonist in fiction.Set in the context of our current technological capabilities, a murderous exchange between HAL and Dave 2001: A Space Odyssey reading is very different.the machine does not garbage To help its human master. it simply cannot.

“Open the pod bay door, HAL.”

“Sorry Dave. I’m afraid I can’t.”