Zhidongxi (public account: zhidxcom)
Author | Li Shuiqing
Edit | Xinyuan
humanoid robot is setting off a new wave in the hard technology circle.
A few days later, the first humanoid robot of Tesla , codenamed Tesla Bot or Optimus (Optimus Prime), will be unveiled on Tesla artificial intelligence html on the 6th. Tesla CEO O Elon Musk has almost the magic of "turning stones into gold" in the technology circle, from smart electric cars to SpaceX and Starlink without exception.
Can humanoid robot take over and become the next industry trend? It has attracted attention and discussion in the domestic intelligent industry.
humanoid robot is also called humanoid robot . As the name suggests, it refers to a robot with human-like appearance, perception, decision-making, behavior and interaction capabilities. Just like the robot Tas in " Interstellar ", it can chat and tell jokes, run and save people, which represents the ideal form of robots in people's minds. This is a very high-tech and high-investment comprehensive field. From Assimo, Japan to Atlas, every robot is expensive and is a certain distance from commercialization. The entry of technology giants such as Tesla, , Dyson, and has undoubtedly brought new turning points to the industry.
Tesla robot diagram
What is the development of humanoid robots at present? What new changes will arise due to the entry of technology giants? Compared with the international level, what level is domestic humanoid robots in? Recently, we contacted Fu Chunjiang, the head of UBL Technology Humanoid Robot Innovation Center, and had an in-depth discussion with him on these issues.
, founded in 2012, has been developing humanoid robots for more than ten years. It is the first company in China to realize the commercialization of humanoid robots. Fu Chunjiang is a 10-year veteran of the humanoid robot industry. He once worked for the Honda Asimo (ASIMO) team, where the world's first bipedal robot that can walk independently and run independently was born.
Starting from Tesla’s humanoid robot map, Fu Chunjiang brings us more truth about humanoid robot technology and industry.
1. Tesla is a humanoid robot, just like IBM PCizing commercial computers
On Tesla's artificial intelligence day in August last year, a real-life robot dressed in white and black veil appeared. With a funny dance, it announced to the world that Tesla wants to "create people". A year later, this Tesla robot will be officially unveiled on the new Tesla AI Day on October 1st.
Tesla robot diagram
According to the trailer, Tesla robot has a slim figure, a height of about 172cm, and a weight of only 56.6kg. Its skin is smoother than humans. According to Musk's imagination, his humanoid robot was positioned in the early stage to replace people in boring and dangerous work, and in the future, it will enter thousands of households to cook, mow the lawn, and take care of the elderly.
In an interview in April this year, Musk directly revealed that Tesla robots are expected to be priced at 25,000 US dollars (about 170,000 yuan), which may be much cheaper than Tesla car .
"Most of what Musk does is grand and long-term, and it has increased in terms of technological difficulty and dimensions. It is often the direction of the layout of a country's Ministry of Science and Technology and Ministry of Industry. Developed countries lack labor, and Musk is a benchmark for the rich labor of , developing countries in ." Fu Chunjiang told Zhidongxi, "similar to IBM's launch of the world's first personal computer in 1981, pushing the computer from the government military market to the consumer market. Tesla entered the humanoid robot, which is expected to 'PC' humanoid robot'."
Robot Tesla robot parameter preview
As the jewel in the crown of robots, people have long expected humanoid robots to reach a higher level of intelligence. In order to meet the needs, the humanoid robot first does not need to re-adapt to the environment where the person is located, the second is to use the tools that the person uses, and the third is to have emotional interaction with the person. Obviously, Tesla's entry has allowed the industry to see a greater possibility that the robotic form is approaching this goal.
According to Tesla's preview, the Tesla robot is equipped with 40 motors , and the legs contain a force sensing device. The movement speed can reach 8 km/h , and it can carry items with a weight of 20 kilograms .
Vision and brain are key systems for humanoid robots. According to the trailer, Tesla robots will use Tesla's FSD (fully autonomous driving) computer modules and solutions, and use eight Autopilot cameras on the face, with the maximum monitoring distance of up to 250 meters; the "brain" uses Tesla's supercomputer system Dojo to train neural network through AI big data. The core is the "D1" chip with extremely strong computing power and extremely high bandwidth. The 25html three chips form a training unit, and the computing power can reach 9 trillion times .
Fu Chunjiang expressed his expectations for this. He believes that Tesla is expected to achieve greater improvements compared with its contemporary products.
Tras humanoid robots may be traced back to four years ago, when Musk talked about his factory being "overly dependent on robots." For example, in 2018, the fluff robot caused the mass production rate of Model 3 to slow down. This is because the fluff robot is mainly responsible for placing fiberglass mats made of fluff on the top of the Model 3 battery pack, but the robot is not good at picking up fluff, causing the product line to often fail.
Aaron Johnson, a professor of mechanical engineering at Carnegie Mellon University, said the fluff robot is “an interesting example” and it fully illustrates the point that autonomy is difficult to generalize, and that the deceptively simple task of handling soft fluff materials will have huge problems because it is not as predictable as rigid parts.
The robotic arm is assembling Model S
So far, Tesla has hired hundreds of robots designed specifically for the production of cars, but many mission robots perform far worse than humans. If Tesla robots can solve the problems of these robots' work, it is expected to promote the industrialization of humanoid robots to take a big step forward.
According to public information, Tesla is currently recruiting staff of humanoid robots and has released about 220 positions, including the design of key components of robots such as "drivers".
One of the job postings stated: "The high-quality code you write will run on millions of humanoid robots around the world for a long time." The matter revealed that Musk will deploy millions of humanoid robots, which is enough to show his ambitions for the field of humanoid robots and his strong ability to act.
This also explains why Tesla is expected to promote humanoid robots to consumers just like IBM PCs.
2. The price is lower than a car, behind it is the integration of three major technology schools
priced at 25,000 US dollars , which is about RMB 170,000. The preview price of Tesla robots is almost lower than that of a car. There are rumors that this price can be realized as soon as 2023 when is put into production.
In this regard, Fu Chunjiang judged that between 2025 and 2030, the possibility of achieving this price is relatively small.
Humanoid robot is a high-tech and comprehensive project. Compared with the popular quadruple-legged robot on the market, - humanoid robots basically have more than 40 joints, while four-legged robots may have only eight joints; humanoid robots have not been fully open sourced so far, and four-legged robots have related open source. There are already many companies making four-legged robots on the market, but productized humanoid robots are still the game for a few players. It can be seen that humanoid robots are much more difficult than four-legged robots.
The Asimo humanoid robot of Honda Japan was once the company's image spokesperson, but it has not been commercialized after more than 20 years of development. The humanoid robot R2, developed by General Motors and , for space exploration, has also been discontinued. Several robotics startups like Rethink Robotics have also gone bankrupt because they failed to commercialize their products.
humanoid robot research and development requires learning human body structure , which mainly involves three levels: human skeletal system, human muscle system, brain and nervous system . Fu Chunjiang said that at present, most robots are focused on simple imitation of the human skeletal system; the human muscle system involves driving, and now it mainly uses servo , and it may use flexible imitation of human muscles in the future; in the field of control, humanoid robots have a lot of reference for the brain and nervous system, but it is still a black box at present.
Review the development history of humanoid robots , which can be traced back to 1967. At that time, Professor Ichiro Kato of Waseda University in Japan released the bipedal robot WL-1; in 1992, Honda launched the humanoid robot P2, which opened a new era of highly integrated humanoid robots in the system. In 2000, it launched the ASIMO robot that could go up and downstairs and dance, representing the highest level at that time. Entering 2011, the new generation of Asimo can run at a speed of 9km/h, and just two years later, in 2013, Boston Dynamics released Atlashml6, which ushered in a new era. It adopts hydraulic control to push high dynamic movement to the extreme, and standing on one foot or walking on gravel is not a problem. Since 2020, we have continuously launched high-difficulty actions such as triple jump , forward roll , and inverted handstands.
Boston Dynamics Atlas Robot
The development of technology often has a process from special to general, and the development of humanoid robots is no exception. Fu Chunjiang talked about From the perspective of software algorithms, humanoid robot players can be roughly divided into the following three technical schools:
1, control-led. The typical representative of is Honda, which was the first to create a self-made humanoid robot. It is mainly based on simplified analytical model control methods, quickly decomposes various actions and action combinations, and supports humanoid robot movement. In the era of limited computing power, this method can quickly and in real time.
2, optimization leading. The typical representative of is Boston Dynamics. It did not find a decomposition model that can be analysed, but found a decomposed motion control regular experience. However, this method is difficult to quantify and combine with action, so Boston Dynamics was silent for many years after launching the first humanoid robot, and did not return to its peak until it later adopted a direct optimization method.
3. Hardware hybrid and soft learning algorithm. Typical representative of is Agile Dynamics, the American company, which did not adopt the control-led and optimization-led path, but tried new hardware control programs to achieve control through a hybrid structure such as serial and parallel connection, hard and soft.
Fu Chunjiang tells us that is now moving towards integration of humanoid robot technology, which is also the technological driving force behind Tesla's entry into humanoid robots.
According to industry speculation, Tesla's humanoid robots may use a set of neural networks to build a unified architecture, unify visual perception, cognition and even motion control systems, and rely on large computing power and visual perception capabilities accumulated in smart electric vehicles. Many of its control frameworks may be based on model prediction control, and the lower level is based on full-body control, and then use learning or simulation to solve the scenario-based problems, thereby achieving better results.
In other words, Tesla will likely launch a cross-generation product that will integrate the research results of the past few decades.
Arizona State University Human Systems Engineering Professor Nancy Cooke believes that to succeed, Tesla needs to let the robot complete a variety of scriptless actions. Without clear instructions, it is difficult for humanoid robots to move freely in reality. But Tesla can leverage its expertise in artificial intelligence and key chips to develop and produce smart humanoid robots at scale at a lower cost to promote the commercialization of humanoid robots.
Fu Chunjiang also has similar expectations. The neural network experience, architecture, AI algorithms, and data accumulated by Tesla's development of autonomous driving can be migrated to humanoid robots. "Compared with robots, the tasks of autonomous driving processing require higher real-time requirements, and the speed is redundant in performance when putting it on humanoid robots; but the interaction between robots and human bodies and the environment is very different. It is curious how to adjust the past architecture when dealing with touch problems."
" Tesla has the ability to build a large-scale development team of in a short time, increase development in action diversity and increase investment in computing power.At the same time, when developing autonomous vehicles, Tesla developed various chips and algorithm platform capabilities of and created an integrated production and manufacturing method , which can all be applied to humanoid robots; humanoid robots share the industrial chain with autonomous driving in servo, speedometer, battery thermal management , etc., and Tesla's supply chain control can also be reused. In addition, the material advantages accumulated by Tesla in aerospace satellites can also be used in humanoid robots. "
In other words, in addition to the forward-looking nature of the technology route, many technologies accumulated by Tesla can be reused, making it the best candidate to shoulder the banner of humanoid robot innovation.
3. The domestic industrial chain may usher in "spring", and humanoid robots still face four major challenges
" The entry of giants such as Tesla and Dyson is a major benefit to the humanoid robot industry. "Fu Chunjiang told Zhitiao, "We know that in the past, in addition to technical reasons, the related costs and selling price issues were also difficult. : Large companies such as Tesla are expected to greatly drive the development of upstream and downstream industrial chains by building hundreds of thousands of robots of the order of millions. "
Humanoid robots involve motors, reducers, structural parts, batteries, sensors, robot shells and other industrial chain links. Taking humanoid robot joints as an example, the number is as high as 25-50, and the cost accounts for 50% of the whole machine. As the core driving component in the joint, the motor affects the output force and motion performance of the joints. According to the CICC report, the mass production of humanoid robots is expected to drive the rapid explosion of motor demand, and the new motor market size is expected to be nearly 100 billion yuan, which will inevitably promote domestic manufacturers to produce more affordable and cheap products. .
In fact, in the face of the trend of technology integration in the humanoid robot industry, domestic companies and scientific research talents have also seen new industrial opportunities in China. With the requirements of domestic economic strength, scientific research results, and manufacturing capacity, many scientific and technological talents have chosen to return to China to develop. Fu Chunjiang, who once worked for the Honda Asimo robot team, is one of them.
Fu Chunjiang (left) and Boston Power Robot Atlas
" have formed a huge reservoir in terms of domestic scientific research results, improving the overall industrial potential. For example, now we want to find any cutting-edge technology that can basically be found in the country. At the same time, there are also a large number of talent resources in China, so that people from all categories can support the development of innovative industries. " Fu Chunjiang said. In the past 10 years of research career, Fu Chunjiang mainly focused on the stability, robustness of movement, as well as environmental adaptability and movement diversity of humanoid robots.
coincided with the eve of the news of Tesla robots in 2021. Fu Chunjiang returned to China and then chose to join UBIT Technology, which has more than ten years of R&D experience. Data shows that UBIT Technology has laid out core technologies including high-performance servo drive , mechanical transmission, motion planning and control, computer vision and perception, intelligent voice interaction, SLAM and navigation, human-computer interaction and hand-eye coordination, and also launched the robot operating system application framework ROSA. In 2022, its large humanoid robot Walker is the generation of China's hard technology Table, the China Pavilion at the Dubai World Expo has been working for more than 1,200 hours, and has completed 1,000 reception tasks.
UBIT Panda Robot Youyou revealed at the China Pavilion at the Dubai World Expo
Fu Chunjiang that UBIT Technology continues to focus on humanoid robot related technologies, and the company is working hard to optimize the whole machine and reduce costs.
But he also admitted that the development of the domestic humanoid robot industry also faces challenges that are visible to the naked eye.
First of all, humanoid robots need to pass the technical products level. The biggest difficulty lies in allowing the robot to flexibly contact objects or environment . In terms of stability, naturalness of movement, and diversity of movements still need to be improved.
In terms of stability, this requires robot companies to set multi-level stability strategies, such as gait cycle adjustment, sole contact surface adjustment, fall and climbing mechanism, etc.; in terms of movement nature, high-dimensional movement exploration needs to be explored through dynamic models; in terms of movement diversity, it is necessary to imitate human movements or refer to relevant basis for doing movements, to promote humanoid robots to explore more difficult movements.
action generalization is another problem that large humanoid robot products need to overcome. . Simply put, it is how to let the robot learn from one example and apply it to others. From the perspective of the global humanoid robot development pattern, one solution is to build a large enough library like Boston Dynamics to call various actions; the other solution is to depend on the development of large general AI . Through general AI, robots can recognize tasks and realize various actions, so there is no need to build large libraries.
At present, one of the restrictive conditions for humanoid robots is the development of general AI technology, and the other is that needs to return to cost.
When will the cost of a humanoid robot be reduced to a price close to that of a car? This determines whether it can actually enter the family. At present, the cost of most well-known humanoid robot products is more than 1 million yuan. Ubishop is the first company in China to reduce hardware costs to US$100,000 per Taiwan, but Fu Chunjiang believes that it is expected to continue to drop to US$50,000 per Taiwan in recent years.
In the cost map of humanoid robots, the most expensive ones are mainly servo machines, including motor reducers, encoders, motor structural parts, etc., accounting for 1/3~1/2 of the cost; the other part is various electronic sensors and computing power boards, accounting for nearly 1/4~1/3; in addition, there are structural parts of the entire machine. Fu Chunjiang said that with mass production, the cost of the entire structural parts and servo of the humanoid robot will be greatly reduced, and the structural parts need to be upgraded by materials and processing technology. He believes that it is entirely possible to reduce the cost of the whole machine by 50% after mass production.
Finally, Fu Chunjiang talked about the importance of ecological cooperation.
"The aspects involved in humanoid robots are very complex, and they are not something that can be done by one person or a team. Therefore, they need to cover ecological cooperation between industry, academia and research. Cooperation is needed in terms of technology, funds, experimental platforms, basic interfaces, etc. to avoid repeated creation of wheels."
Conclusion: Grab the humanoid robot track and test long-term technical accumulation
let the robot help us do repetitive, monotonous and dangerous work, and is the sea of stars for the development of human science and technology. If Tesla robots really amazed the industry this time, then the stars and seas are one step closer to us. "By 2025, humanoid robots will assist some jobs in semi-structured scenarios; by 2030, human resources in many positions can be replaced, allowing people to liberate themselves to do creative and interesting things," said Fu Chunjiang.
At present, with the development of my country's economic strength, scientific research strength and manufacturing industry, the accumulation of technology reservoirs and entrepreneurial enthusiasm have attracted a large number of scientific and technological workers to return to China to invest in new careers. Humanoid robots are an extremely complex and systematic project, which requires enterprises to have comprehensive scientific and technological strength, and also requires the entire ecosystem to gather strength, and even promote it by the state, insisting on solving key issues in cost, general intelligence and other aspects through long-term research and development.
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-