Google's robots are undergoing a process of "humanization."

 Google's robots are undergoing a process of "humanization."

Google's researchers use massive language models and artificial intelligence studies to train simple robots to make decisions and carry out more complex jobs.

SCHENECTADY-CUPERTINO, N.Y. & MOUNTAIN VIEW, Calif. Researchers at Google's lab recently tasked the robot to construct a burger using various plastic toy materials.

The robotic arm was intelligent enough to know that the robot should add ketchup after the meat and the lettuce. Still, it misunderstood the proper procedure and poured the entire bottle into the burger.

Google engineers reported a major accomplishment on Tuesday, and while that robot will not be working as a line cook anytime soon, it is indicative of the field as a whole. The researchers claim they have created robots that can assist humans with a wider variety of common chores using recently developed artificial intelligence software known as big language models.

Robots may now reply to whole requests, more like humans, rather than just a laundry list of instructions, guiding each robot's motions individually.

One instance occurred last week when a researcher asked a robot, "I'm hungry. Can you grab me a snack?" The robot then hunted for some chips in the cafeteria, opened a drawer, located some snacks, and brought them to the human.

According to Google's executives and academics, this marks the first time language models have been embedded into robots.

"This is very fundamentally a different paradigm," says Google research scientist and co-author of a new report outlining the company's accomplishments on Tuesday, Brian Ichter.

The use of robots is already widespread. There are millions of them working in factories all around the world. However, they all do the same repetitive activities, such as pushing products along an assembly line or welding two pieces of metal together exactly as instructed. Competing to create a robot that can learn on the job and do a wide variety of common activities is significantly more difficult. For years, large and small tech firms have toiled to develop such versatile robots.

Corporations use poor data to train AI. Therefore, researchers set out to find more reliable information.

For language models to function, large amounts of text must be uploaded to the internet and then used to train artificial intelligence software to predict what kinds of replies would follow particular inquiries or comments. Nowadays, interacting with a model is a lot like having a conversation with an enlightened computer, as the models have gotten quite adept at guessing the correct response. Google, OpenAI, and Microsoft, among others, have invested heavily in developing more sophisticated models and training them on ever-larger bodies of text in various languages.

There is debate about the work. In July, a Google employee who said he thought the software was conscious was terminated. Despite widespread agreement among AI researchers that these models lack sentience, many are nevertheless wary about their potential for bias after being taught with so much raw, human-created material.

Upon being presented with the appropriate comments or questions, certain language models have proven to be racist, sexist, or easily influenced into uttering hate speech or lies.

Deepak Pathak, an assistant professor at Carnegie Mellon specializing in artificial intelligence and robotics, has claimed that language models could provide robots with knowledge of high-level planning procedures. However, robots will still be missing some crucial details without those models, such as the appropriate amount of force to use while opening a refrigerator. He remarked, "It only addresses the problem at the strategic level."

This Google bot is practicing its catching skills using a lacrosse net. Image from Monica Rodman / The Washington Post.

Google is still pushing forward and has already integrated language models with its robots. Researchers may now have conversations with robots in plain English rather than needing to program them with detailed technical instructions for each activity. In addition, the new software improves the robots' capability to interpret multi-step instructions independently. Now robots can make sense of orders they have never heard before and behave accordingly.

AI was used to program these robots. As a result, they turned racist and sexist.

Zac Stewart Rogers, an assistant professor of supply chain management at Colorado State University, has speculated that language-model-capable robots could revolutionize the administration of factories and warehouses.

Today, "the most effective team is the one that includes both humans and robots," he remarked. Robots can perform physical labor that humans find difficult or dangerous. Nuanced problem solving is best left to humans.

With smarter robots, distribution centers might need fewer people to do the same work they currently do for hundreds of humans. That might lead to a decline in employment opportunities. However, Rogers notes that declines in other sectors typically offset job growth if automation causes a contraction in one.

In addition, it is probably still rather far off. Robots have been trained using various AI methods, including neural networks and reinforcement learning. There have been some significant advances, but overall growth is modest. Google's robots are far from marketable. The company has frequently stated in interviews that it is merely a research facility and has no ambitions to commercialize the technology at this time.

A Google robot challenges a human engineer to a game of table tennis. Image from Monica Rodman / The Washington Post.

However, it cannot be denied that Google and other large tech giants are keenly interested in the field of robots. Numerous Amazon robots work in the company's warehouses, the company is testing drone deliveries, and just last month, Amazon agreed to pay $1.7 billion to acquire the company that makes the Roomba robot vacuum. (Jeff Bezos, founder of Amazon, also owns The Washington Post).

The 'nice' robot Tesla claims to be developing will reportedly be programmed to do basic duties without resisting or attacking.

Tesla is working on both autonomous driving features for automobiles and general-purpose robots.

Boston Dynamics, the creator of the internet-famous robot dogs, was one of several robotics startups acquired by Google in 2013. However, the executive in charge of the initiative was later accused of sexual misconduct and departed the organization. Softbank, a major telecom and technology investment firm based in Japan, purchased Boston Dynamics from Google in 2017. The excitement surrounding more intelligent robots created by the world's leading technology corporations has subsided.

Researchers from Google's Everyday Robots division collaborated on the language model project. Everyday Robots is a wholly owned subsidiary of Google that focuses on developing robots capable of performing various "repetitive" and "drudgerous" jobs. In certain Google cafeterias, robots are already busily cleaning tables and disposing of garbage.