Whether you accept it or not, AI technology has begun to reshape the real world.
In the field of consumer electronics, mobile phone and computer manufacturers are implanting different types of AI models into various terminal products. These companies generally believe that AI can greatly improve the status quo of industries that are stuck in innovation bottlenecks and re-energize users' purchase demand. In the automotive industry, Tesla pushed the official version of FSD (Full Self-Driving) to 1.7 million car owners across the United States in early April, and the end-to-end neural network AI system makes driving decisions more like those of human drivers, such as steering across 4 lanes in a row. More importantly, so far, Tesla's FSD has not been exposed to any major accidents.
Where will the next stop for AI? Venture capitalists, accustomed to catching trends, are beginning to build consensus in the humanoid robotics industry.
In China's primary market, in October 2023, the humanoid robot startup Leap Power completed nearly 200 million yuan in angel and Pre-A round of financing, in December of the same year, Zhiyuan Robot, which was established less than a year ago, received 600 million yuan in financing funds, in January 2024, Xingdong Era, which was established less than half a year ago, announced the completion of more than 100 million yuan in angel round financing, and in February 2024, Unitree Technology completed the B2 round of financing, with an amount of 1 billion yuan.
"The development of China's robot industry has experienced several rounds of ups and downs. In 2013-2014, the investment in industrial robots began to take off, and in 2016-2017, there was another round of investment boom in the field of collaborative robots. Since 2022, general-purpose humanoid robots have become the focus of the industry. ”
Yan Qianhang, vice president of Frees Capital, told Titanium Media APP that the market penetration rate of China's domestic industrial robots has reached about 1/3, and the entire robot industry is gradually maturing. The qualitative changes produced by the AI large model have made everyone realize that the intelligence of robots will become higher and higher, and they will gradually become generalized.
As for when the general humanoid robot will be able to really go to the production line and enter the home, the startups in the industry have different judgments. In terms of Jiji Power, it will take 5-8 years for general-purpose humanoid robots to replace the fine operation of humans on the production line, and it will take 8 to 10 years to really go to the home market. Wang Xingxing, the founder of Unitree Technology, told Titanium Media APP, "By the end of 2025, more generalized humanoid robots will appear, and I feel that I have seen the direction." ”
Musk, all the calls
What ignited the "fire" of humanoid robots? Almost all respondents gave a unanimous answer, and that was the founder of Tesla, Elon Musk.
In February 2022, Tesla completed the manufacturing of the Optimus development platform, and seven months later, at Tesla's Artificial Intelligence Day 2.0, Musk unveiled a prototype of the Optimus Prime robot developed by the Optimus platform, which is already able to walk and carry items independently. At the end of 2023, the second-generation Optimus was officially unveiled, with a weight reduction of 10 kg, a 30% increase in walking speed, and a more dexterous hand and a more free neck.
After Musk entered the game, the entrepreneurial tide of humanoid robots was completely detonated.
Tesla's humanoid robot Optimus
Since 2023, a number of domestic humanoid robot products, including Unitree H1, Zhiyuan Expedition A1, Fourier GR-1, Xingdong Era "Xiaoxing", Step-by-Wheel Power CL-1, and Xiaopeng PX5, have been launched. In the secondary market, UBTECH, known as the "first stock of humanoid robots" in China, rose more than 88% in intraday trading, although the company's main source of income is not humanoid robot products.
In overseas markets, in May 2023, 1X, a Norwegian humanoid robot startup, announced the completion of a $23.5 million Series A2 financing round led by OpenAI. Almost at the same time, Figure, an American humanoid robotics company, received $70 million in Series A financing. In January 2024, 1X once again closed a $100 million Series B funding round from investors including EQT Ventures and Samsung NEXT. A month later, Figure announced the closing of a $675 million Series B funding round with investors including Microsoft, OpenAI, Nvidia and others.
"In 2022, OpenAI has not released ChatGPT yet, but Musk may have seen the capabilities of GPT before the industry. ”
Wang Xingxing told Titanium Media APP that Musk has proved his success in both the automotive industry and the commercial aerospace field. Therefore, when Musk began to make humanoid robots, the government, the market, and capital institutions all believed that they must speed up the entry, and they could not wait for Tesla to really make it before everyone chased it. Of course, the more essential reason for the attention of humanoid robots is the emergence of large AI models.
According to Wang Xingxing, Unitree Technology did not have the idea of setting foot in the humanoid robot track before, because the humanoid robot is too complex and cannot be controlled by traditional algorithms. However, the current development of AI technology has far exceeded its expectations. For example, it used to take one to two years for a humanoid robot to learn to walk, but now it can be trained in a month using AI algorithms.
"The training algorithm of traditional humanoid robots is equivalent to relying on some intelligent human brains to write some mathematical equations, and then solve this equation to formulate the movement trajectory of the robot. However, these equations have great limitations, and once the environment changes, they may not be usable, and new equations need to be redesigned. ”
Wang Xingxing further explained that this kind of training method will lead to a very large amount of code, and when the system is complex to a certain extent, it is impossible to maintain the system by human power alone. However, for AI, as long as the model is built well enough, and then the AI is constantly fed data and computing power, AI can continue to trial and error. Using the reward mechanism in the reinforcement learning algorithm, AI can automatically keep the good training results and throw away the bad ones, and the training efficiency can be qualitatively improved.
Relying on the efficiency improvement brought by AI, Unitree Technology only took half a year to launch its first humanoid robot product. In the "finale" of the 2024 GTC conference, NVIDIA CEO Jensen Huang appeared with nine humanoid robots. Among them, the second humanoid robot from the left is Unitree H1, a subsidiary of Unitree Technology.
Image source: Nvidia's official website
It should be noted that this wave of humanoid robots has even forced Boston Dynamics, the originator of this field, to make changes.
Boston Dynamics is an American engineering and robotics design company founded in 1992. In 2013, Boston Dynamics unveiled the humanoid robot Atlas in the U.S. Department of Defense competition. After multiple iterations, Atlas can perform complex actions such as fast runs, 360-degree spin jumps, and jumping over obstacles. For motion control, Atlas uses the traditional algorithm of "solving a large number of equations" and is powered by hydraulics.
"The previously disclosed cost of the Atlas is around $2 million. At present, the humanoid robot on the market, the price of Unitree Technology's products is about 600,000 yuan, and Fourier is about 1 million yuan. Xi Yue, co-founder of Xingdong Era, told Titanium Media APP that this is the huge gap in cost between Boston Dynamics and the new generation of humanoid robots.
On April 16, 2024, Boston Dynamics announced the official "retirement" of the hydraulic version of the Atlas. Boston Dynamics then introduced a new all-electric Atlas that, like all current humanoid robots, is powered by batteries. In the next control algorithm, Boston Dynamics will most likely also use AI models with higher efficiency.
Three unsolved problems: the brain, the cerebellum, and the proprioception
"The current heat of humanoid robots is equivalent to a small flame, which has just begun to burn. If AI and hardware continue to iterate every year, the industry will be very disruptive to the real world. ”
Wang Xingxing said that by the end of next year, at least one company in the world can develop a more general robot model. This basic large model is like a complete building block, and the large language model is only one of the parts, and other components include visual perception, force perception, decision-making, and interaction.
However, such a judgment has not yet reached a consensus in the humanoid robot industry. The more mainstream view is that for humanoid robots to achieve a greater degree of generalization, they need to make breakthroughs in the brain, cerebellum, and ontology at the same time, which is almost impossible to accomplish in a short period of time.
The so-called brain refers to the robot's comprehension ability, that is, the robot's understanding of human instructions and environmental perception. The cerebellum refers to the robot's fine motion control ability, and the ontology refers to the various parts that make up the prototype of the humanoid robot, such as joints, limbs, heads, etc.
"The emergence of large models has mainly improved the brain ability of robots. Liu Pengqi, executive director of Frees Capital, told Titanium Media APP.
Yan Qianhang said to the titanium media APP, but just like the "brain in a vat", the current large model is just a brain that inputs and outputs language or multimodal information, and exists independently of the machine or ontology. What kind of body should the large model be connected to in the future in order to fully play the generalization function? At present, whether it is investors or entrepreneurs, everyone is in the process of exploring.
In terms of the cerebellum, current humanoid robots have made great strides in walking upright, both on flat ground and on rugged mountain roads. In the subdivision scenario, Figure 01 became the first humanoid robot to "pick up an apple", and the Stanford team's Mobile ALOHA showed good ability to stir-fry and tidy up.
Image source: Figure Official
However, these advances are not enough for humanoid robots to become fully generalized. Whether it's taking apples or stir-frying, it reflects the robot's ability to imitate learning, that is, to learn a single skill by imitating human actions over and over again.
"It's hard to get high-quality data for robots to interact with the physical world, so imitation learning has its place - by using people to teach them, to accumulate some numbers. However, the current imitation learning simply teaches the robot to copy the actions of humans, but it does not understand what the driving factors of each action are. Or rather, the robot doesn't understand why it's doing what it's doing. Yan Qianhang said that if the robot is asked to complete complex human operations such as "serving a cup of water and adding some sugar", imitation learning may not be possible.
"The introduction of vision sensors may make robots no longer blind. But there are many more dimensions of perception that today's robots lack. Yan Qianhang said that although there are sensors such as tactile and force sensors on the market, they are not currently popular in the field of robots, mainly because these sensors are low integration, expensive, and relatively large for humanoid robots.
"Reinforcement learning is a process of trial and error, and it is more generalizable than imitation learning. ”
Xi Yue, co-founder of Xingdong Era, told Titanium Media APP that similar to the training method of autonomous driving, reinforcement learning can allow robots to train in a simulated environment in real scenarios by building a simulation environment, and optimize their behavior through continuous trial and error. "After reinforcement learning training, the robot can not only walk stairs, but also walk snow and grass, so as to achieve better generalization. ”
However, it should be noted that the simulation environment cannot be exactly the same as the real world, and the real-world interaction environment and interactive objects will be more complex than the simulation environment. As a result, the migration of simulation training results to the real world will be biased, which is also a challenge for the entire industry today.
Titanium Media APP has exclusively learned that Xingdong Era has open-sourced the Humanoid-Gym training framework. After the open source of Humanoid-Gym, users can use the framework to verify robot training in Mujoco, a more high-precision simulation environment, through the sim-to-sim conversion interface, thereby improving the efficiency and success rate of sim-to-real (simulation to reality) conversion.
In addition to the training of the brain and cerebellum, the last threshold for humanoid robots to move towards generalization is whether the ontology can fully undertake the action instructions transmitted by software algorithms.
"The hardware technology products of humanoid robots mainly focus on sensors, actuators and drives, energy management and new materials. ”
Li Junlan, research manager of IDC China, told Titanium Media APP that at present, although a variety of sensors have been applied to humanoid robots, there is still room for improvement in accuracy, response speed and integration. At the same time, the energy consumption of humanoid robots is large, and energy-efficient energy power management and energy storage technology are also an important challenge.
"The introduction of vision sensors may make robots no longer blind. But there are many more dimensions of perception that today's robots lack. Yan Qianhang said that there are many sensors such as tactile and force sense, but they are not currently popularized in the field of robots, mainly because these sensors have low integration, are very expensive, and are too large for humanoid robots.
It is precisely because of various constraints that the road to the general purpose of humanoid robots has become longer.
A more realistic present, a possible future
When the "general moment" of humanoid robots has not yet arrived, how to survive has become the most real reality of startups.
"Our company's commercialization strategy is four words – lay eggs along the way. ”
In terms of Jiji Power, the application scenarios of embodied intelligence (including humanoid robots, quadruped robots and other product forms) are very wide, and compared with humanoid robots, quadruped robots will have stronger commercial landing capabilities. The latter is the egg that needs to be laid by the power of the world - relying on the mature movement ability of the quadruped robot to realize the commercialization of the product.
At present, the products of Progressive Power include the humanoid robot CL1, the two-point foot robot P1 and the four-legged robot W1. Among them, the two-point foot robot P1 and the four-legged robot W1 are applied in industrial inspection, logistics and distribution, special operations and other fields.
Image source: Chase Power Official
Similarly, Unitree, which was established earlier, also derives its main revenue from quadruped robots. This is also the business direction of Unitree Technology at the beginning of its establishment, and there are many products such as Go2, B2, and Aliengo. According to public data, Unitree's quadruped robot products account for more than 60% of global shipments, ranking in the leading position in global sales over the years.
Xingdong Era said that the subdivision scenarios of automobiles and consumer electronics, such as factory inspection and the logistics part of the automobile assembly production line, are the commercialization direction that the company is currently exploring. At the same time, there is also the possibility of commercialization of services such as shopping mall greeting.
"Of course, for startups with humanoid robots, financing is a must. Xi Yue, co-founder of Xingdong Era, told Titanium Media APP that because the humanoid robot industry is still too early, the technical threshold is higher, and the R&D cycle is relatively longer, financing is definitely needed in the early stage to survive.
In fact, the moment when the humanoid robot industry appeared was also the past experienced by the domestic autonomous driving industry.
From 2017 to 2018, a large number of autonomous driving startups began to emerge in China, and attracted a large number of venture capital institutions. Similar to humanoid robots, the field of autonomous driving also requires long-term technology development, so it is highly dependent on investment institutions in the early stage. But soon, as the investment boom faded, the commercialization capabilities of autonomous driving companies began to be questioned. After that, a large number of autonomous driving teams were disbanded, laid off, and even went to court.
"From the perspective of technical thresholds, founding teams and industry influence, humanoid robots and autonomous driving are indeed very similar. However, the valuation of humanoid robot companies in this round is generally not as high as that of autonomous driving companies in the previous round. ”
An industry insider who has worked in both autonomous driving and humanoid robotics said that this is a good thing, and that people don't have to blindly pursue the company's valuation and ignore commercialization. "Some of the entrepreneurs of humanoid robots have already seen the problems and risks in the development of the last wave of autonomous driving, so they will have a higher degree of awareness of product commercialization. ”
In addition, the person also said that in the entrepreneurial wave of autonomous driving, everyone is accustomed to working alone. But in the humanoid robot industry, more emphasis is placed on collaboration. For example, at present, Beijing, Shanghai, and Shenzhen are all led by relevant government departments to set up humanoid robot innovation centers. The government departments will come forward to connect the upstream and downstream of the industrial chain, including those who do technology, robot joints, and commercialization. "Everyone forms an entity, and the upstream and downstream companies are shareholder units, which can open up the entire chain. ”
Universal humanoid robot mother platform "Tiangong" Image source: official
Taking Beijing as an example, on April 27, the Beijing Humanoid Robot Innovation Center released the world's first full-size humanoid robot "Tiangong" with pure electric drive and anthropomorphic running in the Beijing Economic and Technological Development Zone, which can run stably at a speed of 6 km/h. "Tiangong" is 163cm tall and weighs 43kg, and the robot is equipped with multiple visual perception sensors, equipped with a computing power of 550 trillion operations per second, a high-precision inertial measurement unit (IMU) and 3D vision sensors.
At the press conference, Xiong Youjun, general manager of the innovation center, said that in order to solve the general problems of the humanoid robot industry and promote the overall development of the industry, the Beijing Humanoid Robot Innovation Center is committed to the research and development of key common core technologies in the industry, and creates two general mother platforms of soft and hard. At present, the general humanoid robot mother platform "Tiangong" has been successfully developed.
According to the relevant person in charge of Beijing Economic and Technological Development Zone, as an important robot industry gathering place in Beijing, Beijing Yizhuang currently brings together 110 robot ecological enterprises to form a robot industry chain system covering core components, complete machines and applications. In the field of humanoid robots, there are not only Xiaomi, UBTECH and other humanoid robot head enterprises to land and develop, but also high-precision reducer, servo system and other humanoid robot parts and components.
At the level of software algorithms for machine learning, the success of Tesla's FSD (Full Self-Driving) has also seen a possible future for the humanoid robot industry.
In Tesla's latest FSD V12 version, FSD Beta was renamed to FSD (Supervised). According to Tesla's official statement, the latest version of FSD Supervised can drive a Tesla almost anywhere under the supervision of the owner.
Before FSD V12, Tesla's autonomous driving scheme has always relied on rule judgment, and any driving behavior is supported by code, and FSD V11 C++ code has more than 300,000 lines. In FSD V12, however, the solution that relied on hand-coded rules was completely abandoned in favor of an end-to-end neural network AI system, and the code industry was reduced to only 3,000 lines.
Tesla's end-to-end approach to FSD is essentially completely data-driven. By compressing the high-quality data of tens of millions or even hundreds of millions of human driving videos into a large model, Tesla FSD can think in an AI way - directly input sensor data when encountering scenes, and output steering, braking and acceleration signals, and there is no coding in the process.
According to information released by Tesla in October 2022, the Optimus humanoid robot uses the same fully self-driving (FSD) computer as Tesla vehicles, as well as Autopilot-related neural network technology.
This also means that humanoid robots can use the same training method as FSD to move towards the road of generalization. According to Wang Xingxing, at present, Unitree Technology's humanoid robot, from walking and running to dancing somersaults, has completely adopted a similar end-to-end solution, from visual perception to leg execution, a model can be realized, without any intermediate process and coding.
"It's only a matter of time before the hardware part of the humanoid robot matures. The most important thing is the AI basic model of the general humanoid robot. Wang Xingxing said that optimistic estimates that the breakthrough of the basic model may occur before the end of next year. However, it is also possible that it will not happen. "Sometimes technological breakthroughs depend on the luck of human beings around the world. Just like if there was no Einstein, there is a high probability that his theory would have been discovered, but it would have been a few years to decades later. (This article was first published in the Titanium Media APP, by |.) Rao Xiangyu Ed. | Zhong Yi)