Why is it so difficult for robots to grab an object?
Core hints :Ken Goldberg, professor of robotics at Berkeley, California, was sadly looking. he kept rubbing his coffee cup in his hand and murmured in his mouth, " how can it be so difficult for robots to master such data? " nowadays, artificial intelligenc

Release date:2018-07-17

Browse number:47

Ken Goldberg, professor of robotics at Berkeley, California, was sadly looking. he kept rubbing his coffee cup in his hand and murmured in his mouth, " how can it be so difficult for robots to master such data? " nowadays, artificial intelligence can easily handle complicated cognitive tasks such as assisting legal and medical research, but for robots, it is still impossible to pick up clothes falling from the ground. Berkeley, Cornell and other universities and Amazon, Toyota and other companies are making continuous efforts to make robots as agile as human hands.


诚然,机器进入我们生活已经几个世纪了,但它们能做的工作较为有限。“它们都被安置在固定地点,重复的做着各种机械的任务。”Goldberg 说道。不过,一旦出了工厂,在一些非结构化的环境中,如乱糟糟的房间和繁忙的仓库,机器就会变得束手无策。

Grasp the trick

" grabbing an object sounds simple, and humans can easily do it without even thinking, but this action is very subtle and elusive for robots. " Goldberg said. But if we carefully consider it, this process actually depends on the very complicated network in our brains. Take the mugs as an example, the human brain will automatically calculate how to hold the mugs most stably. it even specifies the place where each finger should be placed.

Through evolution, the human brain already has its own highly customized processing conventions. " although I have never seen the pen on the table, I know I can easily pick it up. " Goldberg said. " during the process of taking the pen, the brain regained the previous similar experience and handed back the processing method to both hands. " now Goldberg is teaching robots to learn this trick with his students. For this reason, they have specially built a network database named dexterity network, where about 10,000 3d virtual objects are stored, and the scale of storing virtual objects in the database may gradually expand to millions in the future.

When I visited Goldberg's laboratory in September last year, he placed many strange 3d printing models in front of me. Goldberg asked me to try to pick up one of them, but I found that these things had no handlebars to hold, so for the first time a model slipped out of my hands. Goldberg called this shape hostile. he believed that if his own database could handle these shapes of objects, the dexterity of robots would be better than that of human hands.

For this reason, dex - net database has specially developed an algorithm. for each virtual object in the database, it will try to grab 1000 times in 1000 different ways. Three months later, I visited the laboratory again. here I met Jeff Mahler, Goldberg's favorite pupil, who is now in charge of operating the database and has completed the connection between the industrial robot yumi and the database. " industrial robots are good at doing repetitive work, but under the condition that the environment is constantly changing, robots need to continuously adapt to the new environment they feel. this is a huge challenge. " Goldberg said.

With Alexa, Mahler asked the robot to put those strange 3d printing models in the box. When the manipulator touched the object that gave me the power to dismount, it slipped. However, mistakes can also lead to new experiences. if you can get hundreds of machines to test together, you can find the trick to grab this item. one robot has learned, and all the connected robots have learned.

Amazon also has its own set of robots. In 2015, the e-commerce giants launched the Amazon robot competition project, and the winning robots may enter the shipping center for service in the future to completely replace human workers. In 2016, the winner of the competition was Delft university in Holland. their robots took 12 pieces of goods out of their big bags and put them in different boxes. When grabbing goods with smooth surfaces, robots use suction cups, while others use machine claws. Although the whole process is very accurate, the speed is really too slow.

Berkeley university has not participated in the Amazon competition yet, but this year they will participate in the long-established domestic robot contest. In the competition, the robot has to finish the task of vacuuming, delivering food or cleaning the room. However, there are many restrictions on the robots participating in the competition. teams can only use Toyota's human support robot or soft silver cute pepper pepper.

So how attractive is the domestic robot? " if you spend 2,000 dollars to keep your home clean and tidy as new, I will not hesitate to buy one. " Goldberg predicted. This kind of robots can not only deal with littering Xiong Haizi, but also help disabled or elderly people do housework. in the future, they may also take on the task of going shopping.

Seven dreams, a company, has already made laun droid, a folding robot. after years of research and development, this product will be officially launched in March this year, but it moves very slowly and only makes one movement of folding clothes.

Light pipe at Cornell university

Give the robot a pair of dexterous hands

Looking back at the 1973 original " westworld" movie, the robot only installed a slightly deformed hand. However, even after a few decades, the current robots have made little progress. The yumi robot developed by Goldberg's own laboratory has two rigid fingers, which can open and close like a great white shark. If they can use their full hands, it will definitely be much easier to grab things.

However, the main task of the Goldberg team is to improve the existing industrial robots. to create smart hands for future robots, the methods used can be completely different.

In this respect, Cornell organic robot laboratory has a certain leading edge. their robots have five fingers like human beings. these five fingers are made of silica gel instead of hard metal. " in simple terms, each finger is like a balloon, while compressed air drives them. its principle is similar to that of our common paintball gun. The bottom of the artificial finger can hardly move, but after being filled with air, the top of the artificial finger can bend inwards to simulate human hand movements. "

Theoretically, you can even shake hands with Cornell university robots smoothly. this technology can be used to create bionic hands in the future.

With soft fingers, the process of grabbing objects will be much simpler. " our bionic hand can deform according to the shape of the object, so it can grab any object without using the algorithm. " Cornell university expert shepherd said.

The bionic hand of Cornell university is not a unique product in the industry. be bonic and open bionics already have dexterous manipulators with very high degree of completion, but they still need to be operated by human beings. only when the upper part of the severed limb of the disabled person is connected can accurate electrical signals be collected. In addition, the cost of building these robots is huge and the average person cannot afford it at all. Next, the goal of open bionics is to reduce the selling price of its own robot products to about us $ 3,000.

Shepherd is very optimistic about his silicone bionic hand. once it is mass produced, its cost is only about 50 us dollars. However, the real killer of Cornell university is the sensor. they are very accurate and simple to mass produce.

Cornell university has implanted three polyurethane tubes in each finger of the bionic hand. researchers call them light pipes. they can work like fiber optic cables. Two ends of each light pipe are respectively provided with led and photoelectric detectors, and the light passing through the light pipe will gradually dim along with the bending of fingers. Afterwards, the robot can obtain the position of each finger and the degree of contact with the object by integrating the data read by the radio and television detector. Because you can sense external pressure, your fingers will still have pain in the future.

Long way to go

To have the same dexterity as human hands, robots still have many mechanical and computational problems to overcome. " people are very complicated. they are the product of precise cell-level evolution. our current work is only a low-level imitation. In the future, the number of sensors on the bionic hand of Cornell university may increase from 3 to 100, but we need thousands of sensors in order to achieve a nerve density comparable to that of human hands. " shepherd said.

It should be noted that simply adding sensors cannot perfectly solve the problem of dexterity of bionic hands. " we are advancing very slowly on the processing of sensor data. " Goldberg said. " therefore, it is not enough to have sensors alone, and the algorithm with them must keep up. And to be honest, we are still far from the ideal requirements. " this is bad news for housewives who are troubled by housework, but it has saved the rice bowls of many people who rely on their strength to eat.



Friendship link    | International Artificial intelligence

News Exhibition Video Product Enterprise Picture Technology
Liaoning ICP No. 234234234 -1 | Liaoning No. 1102342342030919 public security