Author | Xu Shan
Edit | Yun Peng
On January 17, 2024, David Holz, founder of AI research lab Midjourney, wrote in a post on social media X: "We have reason to expect that by 2040, there will be 1 billion humanoid robots on Earth. By 2060, there will be 100 billion humanoid robots in the world. This sentence is revealing that a new era of humanoid robots is coming.
▲David Holtz X social platform (source: X)
At the beginning of 2024, in addition to AI large models, the hottest technology track is humanoid robots. The humanoid robot on the market today is no longer a clumsy and fancy humanoid robot prototype in the early laboratory, but a humanoid robot finished product that has been continuously polished, and has begun to show its diversified capabilities and applications in logistics factories and quality inspection sites.
For example, Figure, an American humanoid robot start-up, arranged for its robot Figure 01 to "work" at the U.S. manufacturing plant of the automobile giant BMW as soon as 2024 began.
▲Figure 01（来源：X）
At the same time, Musk's Optimus Prime robot is not far behind, and at the beginning of the year, it showed off the housework ability of its own humanoid robot - it can fold clothes in 30s. But then Musk also responded calmly that until now, humanoid robots cannot fold clothes completely autonomously.
▲特斯拉Optimus Gen 2(来源:X)
We can see that humanoid robot start-ups at home and abroad are exploring the development of humanoid robots with infinite vitality. On the one hand, major robot startups have begun to accelerate the layout of humanoid robots, in the past two years, many rookie startups have launched their first humanoid robots, and the industrial track is becoming more and more lively.
On the other hand, the manufacturing cost of humanoid robots has begun to decline, and the manufacturing cost of humanoid robots is no longer in the millions, and the cost is gradually approaching 200,000 yuan or even lower, almost one-third of the previous one, which also brings more possibilities for the commercialization of humanoid robots.
At the time when the enterprise is running wildly, the national industrial policy has sent another east wind.
In the second half of 2023, the Ministry of Industry and Information Technology issued the "Guiding Opinions on the Innovation and Development of Humanoid Robots", mentioning the need to build the "brain" and "cerebellum" of humanoid robots, break through the key technologies of "limbs", and improve the technological innovation system.
▲Notice of "Guiding Opinions on the Innovation and Development of Humanoid Robots" (Source: Official website of the Ministry of Industry and Information Technology)
Local governments have also responded in a timely manner and accelerated the policy layout in the direction of humanoid robots. Beijing has set up a 10 billion yuan investment fund for the development of robot industry, Shanghai will lay out intelligent robot innovation infrastructure and build a collaborative innovation platform of "large model + humanoid robot", and Shenzhen will use the "Shenzhen Action Plan for Accelerating the High-quality Development and High-level Application of Artificial Intelligence (2023-2024)", indicating its recent high attention and support for the humanoid robot industry.
Humanoid robots will continue to become the hottest hard technology track in 2024, and infinite possibilities will emerge in industrial technological innovation.
In 2024, what technological changes will be ushered in in the humanoid robot track? How far is it from the era when everyone can buy robots? Will popular players be reshuffled in this year, some people will go ashore in commercialization, and some people will go out?
These questions will determine the direction of the humanoid robot track and the survival of players, after in-depth interviews and conversations with a number of industry insiders, we found that the decisive focus focuses on these three keys: the continuous reduction of manufacturing costs, the development of application scenarios, and what kind of computing power is used to drive the brain of humanoid robots.
Why are these three points so important, what are the skills of the players, and what are the chances of winning?
First, how to bring down the price? The cost of "joint parts" is the most critical
What kind of price can be called "low enough" in the current humanoid robot stage?
If this question had been a year ago, the answer might have been 700,000. Xiaomi's CEO and founder Lei Jun has publicly stated that the current cost of humanoid robots is about 600,000 or 700,000 yuan each, and mass production cannot be realized. Hangzhou robot startup Unitree Technology also mentioned on social media that the cost of the humanoid robot H1 is about 650,000 yuan. At the time, this was almost the lowest price.
But in 2024, everything seems to have taken a new turn. Players on the robot track almost unanimously gave a price range for the "low manufacturing cost" of humanoid robots: about 150,000 yuan to 200,000 yuan.
Of course, this is not the final price range, but a small goal that robot players have set for themselves.
Recently, the robot start-up Kepler Robotics mentioned that it plans to control the price of humanoid robots by about 20,000-30,000 US dollars (equivalent to 14-210,000 yuan). Zhiyuan Robot, a robot start-up, estimates that the manufacturing cost of its first humanoid robot Expedition A1 will be controlled within 200,000 yuan in the future, and the cost of the star Tesla "Optimus Prime" humanoid robot is expected to be about 20,000 US dollars (about 140,000 yuan).
▲The evolution of Tesla's robots (source: X)
From RMB 700,000 to RMB 200,000, if you want to save at least 500,000 yuan in costs, it is a real test of the "accounting ability" of major robot players.
So what is it that makes humanoid robots cost millions of dollars?
An industry person engaged in robot research analyzed that the most important thing about the million-dollar cost of humanoid robots is the high cost of hardware parts. For example, high-performance controllers, high-performance motor drives and other components, almost every one of them costs as much as tens of thousands of yuan. Among them, if some parts need to be imported from abroad, or even specially customized, their individual manufacturing costs will double again.
▲Tesla robot Optimus Prime specific joint display (source: Youtube)
You must know that whether it is a hand or a foot, each part of the humanoid robot involves more than a dozen to dozens of parts, and the hardware of the humanoid robot is expensive.
If you want to reduce the manufacturing cost of humanoid robots, of course, you must start with these expensive parts.
UBTECH CTO Xiong Youjun told Zhidong that in recent years, they have tried to replace the drives, reducers, dexterous hands, sensors, etc. in the robot power system on the premise of keeping the performance unchanged, and changed the supplier strategy, and also improved the design and processing methods of some structural parts, effectively reducing the hardware cost.
He feels that with the improvement of the product level of domestic suppliers and the entry of a large number of new players, the shipment of some key components has increased, and there will be room for cost reduction.
As for the overall cost estimate, Xiong Youjun and we have calculated that the cost of humanoid robots will be reduced to half of the current cost in 2025. "The cost of humanoid robots can be saved by 50% in the power system, 15% to 25% in the computing system, and 40% or more in the structure and appearance. ”
On the whole, the cost of humanoid robots is mainly reduced from two aspects, one is the increase in mass production of humanoid robots, resulting in a decrease in the production cost of a single robot. The second is to reduce the cost of overseas procurement through self-developed key components such as joints, drives, and reducers.
Second, from the ability to work: determined by two key systems
In addition to the cost problem, how to make humanoid robots play more roles, so that the market, investors, and customers can see their commercial value, is also one of the problems that enterprises urgently need to solve. This is closely related to the intelligent brain of humanoid robots - the system.
1. Gait algorithm system: adapt to more site conditions, and make a breakthrough from the terrain perception ability
The first is walking, although there has never been a shortage of short video demonstrations of humanoid robot parkour, stir-frying, paper-cutting and other big shows in the field of humanoid robots, we can also find that most of these humanoids are carried out on flat ground or operated on special venues, and it is difficult to adapt to more life scenes or factory environments.
Most humanoid robots are still unable to control their gait planning such as going upstairs and downstairs and maintaining balance without falling in a diverse environment.
At this stage, humanoid robots still have some deficiencies in reliability, moving speed, operation accuracy, endurance and human-machine collaboration, and can only work in some specific scenarios and processes. Most humanoid robots can only walk on flat ground, and have high requirements for surrounding light and human interference.
An industry expert related to humanoid robots mentioned that the terrain generalization ability and operation generalization ability of humanoid robots are the core problems that need to be solved. If you want a humanoid robot to be able to jump like a robot dog, it may take a long time.
For gait planning, the domestic humanoid robot start-up has recently made a breakthrough. The company provides humanoid robots with all-terrain mobility and universal operation capabilities, and through real-time terrain perception dynamics, humanoid robots can smoothly and smoothly ascend and descend stairs.
▲Step-by-Moment Power shows its ability to go up and down stairs (Source: Step-by-Moment Power)
2. Algorithm collaboration system: information input technology is diversified, and AI improves information processing efficiency
In addition to the gait algorithm of humanoid robots that needs to be broken through, the multi-algorithm collaboration ability of humanoid robots also needs to be improved. For example, the visual processing system represented by the "eyes" of humanoid robots, the decision-making system represented by the "brain", and the execution system of the "hands". How to make different control systems smoother and data operation more efficient will still be a new pain point for enterprises.
To solve this problem, it is inseparable from the latest innovative technologies and the hottest AI models. We have seen a lot of electronic skin technology, algorithm 3D visual recognition technology and compliant control technology, combined with the recently popular AI large model technology, which is improving the information ability of data at the information input end.
At the same time, these collaborative algorithms need to be optimized according to the application scenario. The relevant person in charge of UBTECH revealed that UBTECH Walker X has been continuously trained and tested in smart factory scenarios, based on the self-developed object 6D pose recognition technology and autonomous hand-eye coordination grasping technology, which has been able to ensure that the humanoid robot Walker X can accurately grasp objects.
If the hardware determines the upper limit of the humanoid robot's capabilities, what kind of humanoid robot you can build, then the algorithm will determine its lower limit, what your humanoid robot will be like to move, and what is the role. Judging from the existing analysis, in 2024, the gait control system and collaborative processing system of humanoid robots will promote the commercialization of the humanoid robot market from two levels: scenarios and capabilities.
3. Humanoid robots cannot be used with power: large models and computing power determine the standby time of the brain
As a "behemoth" in the consumer electronics industry, humanoid robots also cannot escape the two common problems of consumer electronics: battery life and heat dissipation.
As a humanoid robot that needs to work 24 hours a day, its heat dissipation will affect its working time, and its battery life will affect its commercial value.
In general, there are many reasons why a humanoid robot can overheat, the most common being an overheated processor. This also involves the problem of how to allocate the computing resources of humanoid robots.
If all the data processing is placed in the cloud, the data processing time will be long, and the humanoid robot will appear "stuck" and "stuck". If the data processing is completely placed at the edge, that is, the local processing is placed on the humanoid robot, when the humanoid robot is running at full capacity, the robot will also have various problems such as heat dissipation and low processing efficiency.
Now, this puzzle is in the hands of AI. The emergence of large AI models has enabled humanoid robots to have a more powerful tool chain, which can process multi-modal information more "intelligently" and have higher multi-threaded processing capabilities.
It is reported that UBTECH has reduced the training time of various tasks of humanoid robots through a special large model for humanoid robots, and also reduced the deployment cost of humanoid robots to a certain extent.
▲UBTECH robot in the factory
"At this stage, the combination of humanoid robots, large models and general AI can help AI break the boundary between virtual and reality, interact more with the physical environment, collect more data, thereby promoting the collection of embodied intelligence data and the improvement of algorithms, and accelerating the interaction and commercialization of embodied intelligence, large models and humanoid robots. Xiong Youjun said.
It can be seen that in addition to the problems created by the robot itself, empowering the capabilities of AI large models to humanoid robots and handling the operation of resources between different operating systems may become one of the important technological watersheds for enterprise players in 2024.
Conclusion: To solve the three major obstacles, humanoid robots rely on "intelligence" to make a living
It's not difficult to build a humanoid robot these days, but there are still many challenges in mass producing a humanoid robot that everyone can afford.
As the data of humanoid robots in diverse scenarios becomes more and more abundant, more and more enterprises hope that humanoid robots will go to the "general field". However, some users are eager to have humanoid robots, but they are defeated by the high cost of robots, or humanoid robots at this stage still lack the ability to understand instructions.
With the advent of the AI era, in 2024, humanoid robots may be expected to become the right-hand man of human beings, beginning to help users improve work efficiency and liberate some workers from tedious labor work.