What is the reason for the delay in the robot revolution?

Let’s first conclude: In my opinion, the reason why the “robot revolution” in our minds has not arrived is because human beings are too confident in their own attributes, so that in the process of development, they have unknowingly found the wrong direction. So I took some detours.

Recently there have been rumors that “The Matrix” is going to shoot the fourth. As a very classic sci-fi movie, the image of the robot in the Matrix also represents a large part of people’s imagination of the robot: head, five features, and limbs. Except for the strange organs on the face, people can see at a glance that they are robots, and other features are similar to those of people.

In fact, there are now many excellent technology companies that focus on scientific research on “humanoid robots.” For example, the famous Boston Dynamics [1] has made a great breakthrough in the exploration of “humanoid robots”. It has previously released its “jumping obstacles” and “backflips” videos, and recently released it “identification”. The video of “Complex Terrain” and “Parkour” is amazing.

But I don’t know if you realize that although these robots reach or exceed the average human level in some actions (such as most people do not backflip), in terms of the fluency of the behavior and the scope of the environment, the distance The real human beings are still far apart. While marveling at its rapid development, we should also clearly understand that it is still far from being able to compete with real humans.

why? Because the structure of this robot is “imitation”. Therefore, in the process of manufacturing robots, the most energy-consuming part is not in its “smart” part, but in its “imitation” part. For example, in order to make it act more like humans, people spend a lot of thought on their mechanical construction and driving methods.

Found a problem? We know that in terms of transportation efficiency, the car does not know how much higher than humans; in terms of durability, the tank does not know how much higher than humans; in terms of balance, the limbs do not know how much higher than humans… but we are still willing Regardless of its performance, it takes a lot of effort to “manufacture humanoid robots” just to satisfy its dream of “robots”.

Of course, human beings have long recognized the problem, so many “robots” are no longer rigid in human form: for example, “sweeping robots” that have been commercialized for a long time are generally just flat cylinders, such as one or two years ago because of victory. The “artificial intelligence” of the top Chinese Go masters AlphaGo is simply a bunch of computers, and people have to work hard to do it… But looking at the history of “artificial intelligence,” humans have indeed made mistakes because of some obsessions.

Dr. Wu Jun [2] once gave a lecture at Tsinghua University. He mentioned that the development of artificial intelligence is mainly divided into three stages:

• The first phase was before 1966;
• The second phase is from 1970 to 2000;
• The third phase is from 2000 to the present.
According to computer ancestor Alan Turing [3], is there a test method for “machine intelligence” called “Turing test”: suppose there is a computer and a person behind a wall, then ask them a question and give an answer. When you can’t judge whether it is a person or a machine, you can say that the computer has the same intelligence as people.

So how do you make the machine owner’s intelligence? In the first phase of the development of artificial intelligence, the idea of ​​mankind at the time was to let robots learn humans. At first, people thought that through the exact words and logic, the machine can be intelligent.

But people quickly discovered a problem. In 1966, well-known computer scientist Marvin Minsky [4] gave a vivid example of the existence of this problem.

He said two sentences:

• The pen was in the box.
• The box was in the pen.
The first sentence is easy to understand. How to understand the second sentence? It turns out that pen has another meaning, which is “fence”. However, you can’t judge whether the pen here is a pen or a fence through pure logic. What we rely on is “life experience.”

Because of his fame, the Foundation stopped supporting his artificial intelligence after the American Nature Foundation wrote such a report.

The second phase of artificial intelligence was about 1970-2000, which was mentioned by the natural language processing master Frederick Jelinek [5] in the 2000 IEEE ICASSP speech. Jelinek was born in communication and came to IBM in 1972. One thing to do at the time was to make the machines smart. They were thinking, what can prove that the machine is intelligent? One solution is: first, it recognizes the human voice; then, translates the language into another language that can be understood by the machine; and finally, answers the question.

Jelinek doesn’t know anything about artificial intelligence, so he thinks about it from a new perspective: he sees speech recognition as a communication problem. Communication has a source and a channel. If the source and channel are coded separately, just decode. It is known that the content of the transmission; he believes that the speech recognition is also the same, assuming that the brain is a source, the language is propagated through the channel, and the communication method, using a large amount of sample data, can correspond to the language with the exact meaning – – Because its accuracy has a lot to do with the number of samples, this is called a “data-driven” approach.

At this point, human beings have gradually discovered that the key to realizing “machine intelligence” is not to imitate human language logic, nor to imitate human speech content. These original breakthroughs have proved to be futile. The real breakthrough lies in “Digital” and “sample size” – collect enough data samples and quickly calculate and then respond.

However, “collecting large amounts of data” was a very difficult thing before the advent of the Internet era. “Rapidly calculating large amounts of data” was a very time-consuming task in the past when computer performance was not high enough.

In fact, it was not until 2005 that we could more easily use the Internet to collect large numbers of samples; until around 2010, the use of computers for complex large-scale neural network operations gradually broke the bottleneck in computing speed. Similarly, until recently, with the improvement of the computer system ecology, robots have gradually begun to “open source” and can receive more “data” as “learning materials.”

For example, the domestic robot unicorn UBTECH has taken the lead in making a ROSA robot operating system. As an “open system”, ROSA has achieved relatively complete functions in voice, visual, motion and emotional control, device interaction and scheduling through “modular design” and “hierarchical architecture”, enabling the “service robot” to have Expressiveness and sense of life. ROSA (official website) provides a rich interface, so that ordinary programmers can also develop in Java language, use their own way to DIY machine movement, expression and language functions, so that their robots become “unique” in a certain sense. robot.

Of course, from the current situation, the ROSA-like programmable interface used on the robot is not flexible enough, and there are more restrictions, but it is considered to be “the first crab.”

As I mentioned before, “machine intelligence” has become a real opportunity to form a large-scale breakthrough, but it is only a matter of the last ten years, and the development of new things is always a trot, PC era, Windows, mobile Internet There are Android and iOS in the era, what will the robot era be, this is a future worthy of yearning.