Wuhan Cooper has developed the only intelligent shrub disorder sorting line in China. The system can identify 26 kinds of shiitake mushrooms. The sorting speed can reach 0.8 seconds per piece, which is 2-3 times of artificial, and the precision rate can reach 99.7%. In the future, the intelligent disordered sorting system will not only be applied to the food industry, but the debugging of the equipment will not be dependent on programming, but will be given to independent learning.
Wuhan Cooper mainly develops robot operating systems. With the power of AI and traditional algorithms, robots become as simple and easy to operate as ordinary computers.
In March of last year, Cooper received a 40-yuan A round of investment in China. In December, the Bank of China completed RMB 102 million in Series B financing, setting a new high in financing. Next, in addition to 3C, food, automotive, logistics and other fields, the company will also explore the technology of medical and new retail scenes.
"Those things that others can't do, we did it, and the demand will come to the door sooner or later," said Li Wei, CEO of Cobot Technology Co., Ltd. in Wuhan. The only food-grade intelligent sorting system in China, the robot startup company established in May 2016 has made a name for itself.
By chance, my friend found Li Wei and asked him to help a Hubei mushroom company solve the problem. Traditional mushroom processing enterprises rely on manual sorting of different types of mushrooms, but recruitment difficulties have been plaguing mushroom processing enterprises. "Now the young people are 95, 00, and are not willing to do these boring and repetitive tasks." Li Wei said.
The manual sorting speed is also slow, and it takes about 5 seconds to sort a mushroom. Chinese people eat 1400 tons of mushrooms a year, and this sorting speed cannot meet the sharply growing market demand.
The sorted mushrooms will go to the downstream food enterprises, but the workers are prone to fatigue due to long-term low-head work, and the sorting quality industry is not guaranteed. If the downstream manufacturers cash back the impurities such as white film in the mushrooms, it will cause heavy punishment.
So, Cooper customized a robotic intelligent mushroom disorder sorting line for this company. The system can identify 26 kinds of mushrooms, the sorting speed can reach 0.8 seconds / piece, it is 2-3 times of artificial, and the precision rate can reach 99.7%.
Li Wei calculated the account: “A traditional mushroom sorting line usually consists of 8 workers. Only 4 robots and 2 workers work together on the Cooper intelligent line. The number of 1 hour sorting is 210 for the traditional line. %, the cost is only 75% of the cost of the traditional line. With the same output, an intelligent production line can save more than 200,000 enterprises per year."
Create a general-purpose operating system, use a computer, and use a robot
In Li Wei's view, the company's star product, the intelligent disordered sorting system, is more like a popular app.
"App fires, the value of the platform is highlighted." Li Wei said that no one is developing a fun application on the Android system, and this platform can't be fired.
Cooper's core competence is a self-developed robotic operating system. The system mainly includes two underlying technologies of vision and force control, which can be applied to sorting, grinding and flexible assembly.
From the robot operating system to the market, and Li Wei's "obsession" of robot software is inseparable.
Look at the sales figures for a group of industrial robots. In 2015, the global sales of industrial robots were only about 250,000, which is ten times the peak of sales of large computers. In comparison, the total sales of servers and PCs are about 10 million units and 300 million units respectively. It can be seen that, unlike computers that have already entered thousands of households, industrial robots have not really moved to the mainstream market.
Two key factors that constrain the industrialization of industrial robots are cost and ease of use. "With so many small and medium-sized enterprises in China, many of the needs are not met, and the operating threshold of the robots is not lowered. How to promote the upgrading of the manufacturing industry?" Li said.
However, improving ease of use is more difficult than reducing costs.
"Our survey found that even if you buy ABB, FANUC robots, users can not use." Li Wei said. “Companies are still looking for ABB people to do integration.†Several industrial robots currently used in China, Kuka, ABB, Yaskawa Electric and Fanuc, use their own control systems and interactive software.
Although applications such as ROS (robot operating system) and OpenCV (open source computer vision) simplify tasks and requirements, allowing robots to work and do something useful, they still need to be operated by a robotic expert with a PhD. It is very inconvenient to maintain and reprogram the bought robots. To change the company's robots, it is necessary to retrain the personnel. The subsequent labor cost is almost three times the price of the product.
In addition, the robot is also very stupid. Every action requires an engineer to perform fine programming on the back. Since the path is fixed or limited, you can only avoid some problems by manual adjustment. When the robot responds to complex scenes visually, manual adjustments are useless.
China values ​​robot hardware and looks down on integration. He feels that "it is not a tune", Li Wei sighs, in fact, the interlacing is like a mountain. "There are a lot of problems involved, such as engineering problems." Li Wei said, "There are scientific ideas, how to use software or even AI ideas to integrate them into the same platform. Let debugging become regular."
Deep learning has brought about a revolution in turning robots into learning machines. Without precise programming, robots can learn from data and experience over time and perform a variety of tasks.
Cooper's positioning is therefore very clear and clear: with AI combined software, relying on its own platform to achieve software to determine the hardware, improve the ease of use of the robot. For example, if a customer needs to grab a service, he can first purchase a robotic arm from a hardware manufacturer, and Cooper adds an AI+ software operating system to the robotic arm.
In order for the robot manufacturers to cooperate with the open interface, the Cooper family talked about cooperation. It will be very difficult at first, but Li Wei said, however, there are always points that can be broken. As long as there are orders in hand, domestic manufacturers are willing to cooperate with you.
Li Wei lamented that the underlying controller is the core technology of the robot giant and will not open source. "It is good to give you one or two APIs."
Borrowing deep learning technology to make robot operation easier is becoming a hot spot for entrepreneurship. Last month, former OpenAI scientists resigned to create the intelligent robot company EmbodiedIntelligence. The company's president also expressed a similar entrepreneurial idea: "The smart modules we provide can be connected to any robot on the market, allowing them to naturally learn new skills without having to write obscure code." Japan Mujin, PreferredNetworks, USA Strong artificial intelligence company Vicarious, etc., is also a hot startup in this field.
An analysis of 752 robotic startups in the global database of robot reporting sites shows that more than half of startups start with software. Not long ago, Wu Enda established the company Landing.ai, which aims to let deep learning go to the manufacturing industry.
Aiming at intelligent disorderly sorting, liberating human hands
AdilShafi, a late expert in robot vision, predicted that robotic disorder will be the mainstream in 2020. But in fact, robotic crawling has not been taken seriously for a long time. Li Wei is a Ph.D. student in engineering at the University of Technology, Lausanne, Switzerland, and a postdoctoral fellow at MIT. The old "Grasper" with eight years of experience in acquisition planning and intelligent control research is deeply felt. "The Chinese who do robots are basically graduated from foreign universities. Only 5 or 6 people have been doing 5-10 years or more." Li Wei said.
“Why are there fewer people to do? Because everyone used to laugh at the industry and think that workers can complete this action, there is no industry application prospect.†However, as the cost of labor climbs, many large companies have begun to try this field, such as Google, Amazon, FANUC, etc.
The commercial viability of automated sorting in an unstructured environment remains a formidable challenge. When the parts and goods are placed in the box in completely random form, the directions are different, and there is overlap or even entanglement, the imaging and grasping of the robot becomes difficult. In recent years, breakthroughs in deep learning have opened up new possibilities for identifying and capturing objects of various shapes and sizes in an unstructured environment. In 2016, in the famous Amazon PickingChallenge, the champion and runner-up took deep learning as the core algorithm behind their visual and crawling tasks.
As one of the few real cases in China that will be deeply studied in the industry, the accuracy of Cooper's intelligent disordered sorting system has reached 99%. The high-resolution industrial camera is combined with the image sensing algorithm to obtain the category and position information of the target object. Then, using the robot arm, end effector and dynamic grab control algorithm to achieve accurate sorting of objects. “The mushroom sorting system is still being trained now,†Li said. In order to train the mushroom sorting system, Cooper looked for a picture of more than a thousand T mushrooms, which is more difficult to identify than to identify dogs and cats.
There are many types of shiitake mushrooms, and mushrooms, mushrooms, and citron are three common shiitake mushrooms. The flower mushroom is further divided into flower mushroom, white mushroom, and camellia mushroom..... There are also thick mushrooms (shrimp mushrooms) and thin mushrooms. These mushrooms are different in size, shape and even texture. "The picture of cats and dogs can be crawled from the Internet. The pictures of mushrooms are not common. There are hundreds of images on ImageNet. We have to do the pictures ourselves." Li Wei recalls, "And mushrooms are the products of nature, China and Japan. The production is different. In addition, the mushroom iteration is very fast, basically a generation of three months."
Although deep learning is used in the visual aspect, in terms of power, it is still a traditional method, Li Wei told us. “Deep learning is not a panacea, and it should be used with caution.†At present, Cooper has achieved the sorting of more than 5,000 categories, and the introduction of new products can be completed in a few minutes.
In robotic grinding and 3C flexible manufacturing, Cooper is also exploring the use of machine learning. For example, how strong the grinding process is, how to design the force feedback control is better, and finally how to evaluate the quality of the grinding, which is still solved by experience in the industry. Cooper hopes to solve these three problems through machine learning.
In the 3C assembly process, it is more common to latch the hole into the hole. Is it possible to turn a hacking solution that exhausts all of its capabilities into a search-like problem? Search is an area where machine learning excels. How to build this database, how to conduct an effective search?
At present, deep learning is only a supplement to us, or a traditional method. Li Wei said repeatedly.
The scene is large, moving from industrial robots to service robots
On the choice of the track, Li Wei always has a clear understanding. Choose to do the crawling, and also consider that the giant will not do the crawling. Because of the giant's genes, Li Wei explained that the giants do robotic arms, but the arm does not have a grab function. They are used to changing interactions in the program, but there is no habit of changing physically. The industry itself is a relatively small industry, not so many people do, but the future robots need these things, so it is especially suitable for startups.
However, compared with the Internet giant, "we have technology, but lack of scenes." Li Wei pays great attention to the application scene. “You also want your things to be useful, right?†In addition to providing solutions for companies in the food industry, this year, Cooper has worked with ABB, JD. Some large integrators have reached a cooperation.
According to Li Wei, in the cooperation with the operator Prog, 4000 SKUs can be identified, which is completely realized by robots.
In the efficiency of logistics sorting, Li Wei said that the original 12 people / line can be reduced to 3 people / line, the sorting speed is increased from 12s / single to 4s / single, the peak value of a single system can reach 300,000 Single/day. "Basically, what the current manual workers with wages of 5,000 to 10,000 are doing in the field of potential robot replacement upgrades," Li said.
In December 2017, the company completed RMB 102 million B round of financing, and the amount of financing within the venture was new. After the B round of financing, Cooper said that the funds will be mainly used for market promotion and to explore new scenes such as medical and new retail.
It takes a lot of time to buy a bottle of water online, and the unmanned shelf content of the office building is not rich. In this case, “in the future, the convenience store may disappear, and it may be replaced by a large unmanned warehouse. This shopping model is completely sorted and transmitted by the robot in the background, reducing the cost of the enterprise and the cost of the consumer.†Road.
In the future, the company will expand from the field of industrial robots to the field of service robots. Cooper will also be committed to liberating people. This year, Cooper's sales revenue was 25 million yuan. In addition to the direct sale of software and hardware solutions, the company is also trying to charge by system call times in the form of leases.
The idea of ​​providing services rather than selling products has always been a good way to market untested products, and related startups have discovered the benefits of economies of scale.
Cooper believes that the robot-as-a-service (RaSS) model can greatly reduce the psychological threshold for users to pay and accelerate the company's market share. Now the company has about 70 full-time employees. Among them, 50 people are responsible for the development and iteration of the underlying algorithm of vision and touch, and the remaining 20 people are responsible for the landing of specific application scenarios. It is not easy to form such a team in a second-tier city like Wuhan.
"Although we don't have the top talent, the overall level of the core team is good." Li Wei said confidently. However, Wuhan is just the starting point for this robotic startup. In an old-fashioned job advertisement, Li Wei wrote: The coordinates of Wuhan within one year, the future coordinates of Shenzhen and Europe.