Researchers rush to build robots working in Amnesty International. But will they work? : NPR

Chelsea Vin (left) and Mo Jin Kim perform a demonstration with a robot at Stanford University.

Mo Jin Kim/Stanford University


Hide the explanatory name

Switch the explanatory name

Mo Jin Kim/Stanford University

Artificial intelligence can find a recipe or generation of an image, but it cannot hang a picture on the wall or cook dinner.

Chelsea Vin wants to change. The Finn, an engineer and researcher at Stanford University, believes that artificial intelligence may be on the threshold of operating a new era in robots.

“In the long run, we want to develop programs that allow robots to work smartly in any situation,” she says.

A company that has already participated in its founding has shown a robot of artificial intelligence for general purposes that can fold laundry, among other tasks. Other researchers have shown the capabilities of Amnesty International to improve the ability of robots to do everything from the package sorting to the drone race. and Google revealed the veil A robot works for Amnesty International can pack lunch.

But the research community is divided on whether the tools of the gynecological intelligence can convert robots the way it is around some online business. Robots require data in the real world and face strict problems than Chatbots.

“The robots will not suddenly become a science fiction dream overnight,” says Ken Goldberg, a professor at the University of California at Berkeley. “It is really important for people to understand this, because we are not there yet.”

Dreams and disappointment

There are fewer parts of science and engineering that have a larger gap between expectation and the reality of robots. Karel Abyek, the writer of the word “robot” itself, was drafted in the twenties of the twentieth century, she wrote a play that imagines human -like creatures that could implement any task of its owner.

In fact, robots have faced a major problem in performing trivial functions. The machines are at their best when they perform very frequent movements in a carefully controlled environment – for example, on the auto collection line inside the factory – but the world is full of unexpected obstacles and unfamiliar things.

In the Finnish Laboratory at Stanford University, the graduate student Mu Jin Kim explains how at least the robots of the Acting have the ability to fix some of these problems. Kim developed a program called “OpenVla“Which means vision, language and work.

“It is one step towards ChatGPT for robots, but there is still a lot of work to do,” he says.

Moo Jin Kim has prepared a robot that works with a detect at Stanford University.

Moo Jin Kim has prepared a robot that works with a detect at Stanford University.
Hide the explanatory name

Switch the explanatory name

The robot itself looks somewhat noticeable, but rather a pair of mechanical weapons with sounds. What makes it different is what is inside. Ordinary robots should be carefully programmed. The engineer must write detailed instructions for each task. But this robot is supported by an educational nervous network. The nerve network runs how scientists believe that the human brain may work – the internal “nodes” in the network have billions of connections together in a way similar to how to link neurons in the brain together. “Programming” like this network revolves around enhancing the connections that concern, weakening those that do not do so.

In practice, this means that Kim can train the OpenVla model on how to do a group of different tasks, simply by displaying it.

Hanging on the robot is a pair of control stick that controls every arm. To train it, the human operator uses “brides” control clips on the robot as it does the required task.

“It is essentially similar to any task you want to do so just to do this over and over again like 50 times or 100 times,” he says.

This repetition is all that is required. The connections between the nervous network are strengthened by the artificial intelligence of the robot each time they are displayed on the procedure. Soon you can repeat the task without brides.

To demonstrate, Kim shows a tray of different types of corridor mixture. She has already taught her how to pull. Now I want some mixture of green M & MS and nuts, and all I have to do is the question.

Mo Jin Kim/Standford University

“Some of them are green with nuts in the bowl,” write. Slowly shivering the robot arms at work.

On the video summary, OpenVla places a star over the correct container. This means that the first part of the model, which must take a text and explain its meanings visually, work properly.

Kim says he does not always do. “This is the part in which we breathe.”

Then slowly, hesitantly, communicate with his claw, pick up the journalist and get a trail mix.

“It seems that he is working!” Kim says enthusiastically.

It is a very small scoop. But a scoop in the right direction.

Any thing robots

Stanford researcher Chelsea Vin participated in the establishment of a company in San Francisco Physical intelligenceWho seeks to follow this training approach to the next level.

It imagines a world in which robots can quickly adapt to simple functions, such as making a sandwich or storing groceries. Unlike the current thinking on robots, you suspect that the best way to get there may be to train one model to do a lot of different tasks.

“We actually believe that trying to develop public systems will be more successful than trying to develop a system that does something very good,” she says.

Physical intelligence has developed an Amnesty International nervous network that can fold the washing, coffee pills and assembly of a cardboard box, although the nerve network that allows it to do all these things is very strong so that it cannot be physically on the robot itself.

“In this case, we already had a workstation in the apartment that was calculating the procedures and then sending it over the network to the robot,” she says.

But training data in the following settlement of Robot Ai-is a more difficult task than just collecting text from the Internet to train Chatbot.

“This is really difficult,” where is it. “We do not have open internet data from robot data, and often it comes to collecting data ourselves on robots.”

However, the Finn believes it can be done. In addition to the human trainers, robots can also try to try again and again to carry out tasks alone and build the base of their acquaintances quickly, she says.

Data dilemma

But Ken Goldberg of Berkeley is more doubtful that the gap in the real world can be blocked quickly. AI Chatbots has improved over the past two years because they have a large amount of data to learn from it. In fact, they swept the entire internet to train themselves on how to write sentences and draw pictures.

Ken Goldberg's founder, co -co -Robotics professor and professor at the University of California, Berkeley.

Ken Goldberg’s founder, co -co -Robotics professor and professor at the University of California, Berkeley.

Niall David Cytryn/Ambi Robotics


Hide the explanatory name

Switch the explanatory name

Niall David Cytryn/Ambi Robotics

You only need to build internet data in the real world for robots will go more slowly. “At this current rate, we will take 100,000 years to get a lot of data,” he says.

“I would like to say that these models will not work in the way they are trained today,” Bolkt Agarawal, robotics researcher at the Massachusetts Institute of Technology, agrees.

Agrawal is a defender of simulation: mode the nerve network of artificial intelligence runs the robot in a virtual world, allowing them to repeat tasks over and over again.

“The simulation strength is that we can collect large amounts of data,” he says. “For example, in three hours of simulation, we can collect data of 100 days.”

This approach has succeeded well for researchers in Switzerland who recently trained a drone how to race by placing his brain that works in Amnesty International in simulating and operating it through a pre -determined and repeated course. When I entered the real world, he could faster and better than a skilled human discount, at least part of the time.

But simulation has its faults. The drone worked well for an interior. But she could not deal with anything that was not simulated – wind, rain or sunlight – can throw the drone.

Airlines and walking are relatively simple tasks for simulation. Goldberg says that picking up things is actually or performing other manual tasks that humans find it completely clear to be repeated in the computer. “There is mainly no simulation that can accurately treat treatment,” he says.

Absorb

Some researchers believe that even if the data problem can be overcome, the deepest issues may be Amnesty International robots.

“In my opinion, the question is not so, do we have enough data … it’s most framing the problem,” says Matthew Johnson Roberson, a researcher at Carnegie Mellon University in Pittsburgh.

Johnson Roberson says all the amazing skills that Chatbots offered, and the task required of her to do is relatively simple-look at what kinds of human users and then try to predict the following words that the user wants to see. Robots will have to do more than just compose a sentence.

He says: “The best prediction word works well and is a very simple problem because you only predict the following word,” he says. The transition through space and time to carry out the task is a much larger range of variables for a neurological network to try to treat.

“It is not clear now that I can take 20 hours of Go-PRO shots and produce anything reasonable for how the robot in the world moves,” he says.

Johnson Roberson says he believes more basic research should take place in how nerve networks do better. It warns that the field needs to be cautious because the robots have been burned before-by race to build self-driving cars.

“A lot of capital rushed quickly,” he says. “People have motivated promises on a timetable they could not present.” Then he left a lot of the field capital, and there are still basic cars without a driver who is still without a solution.

However, even skeptics believe that robots will be changed forever by artificial intelligence. Goldberg participated in the founding of the Tours Al-Hazm company called Ambi Robotics, which has launched a new AI-1 system known as Prime-1 earlier this year. Artificial intelligence is used to determine the best points for an automatic arm to pick up a package. Once the selection point set by artificial intelligence is placed, the arm, which is controlled by more traditional programming, makes the seizure.

He says that the new system dramatically reduced the number of times that the packages decrease. But he added, “If you put this thing in front of a pile of clothes, you will not know what to do with this.”

Once again in Stanford, Chelsea Vin says she agrees that expectations should remain under control.

“I think there is still a long way to move to technology,” she says. It also does not expect global robots to replace human work – especially for complex tasks.

But in a world with the population who progresses in aging and the lack of expected labor, you believe that the Acting Robots can block some gap.

“I imagine this will be something that increases people and helps people,” she says.

Leave a Comment