Cool Play Lab Works
Originally published on WeChat account Cool Play Lab
Imagine waking up on a weekend, and a high-tech humanoid robot walks in, holding a steaming cup of coffee in front of you.
It sounds like a scene from a science fiction movie, right? But this January, a coffee-brewing robot called Figure 01 from Figure company made such a scene a reality.
After watching a 10-hour human demonstration video, the robot can autonomously open the coffee machine lid, put in the coffee bag, press the start button, and complete the entire coffee brewing process in one go.
Remarkably, when the coffee bag is not placed correctly, it even automatically corrects it.
In such “delicate tasks”, domestic robots also show impressive capabilities.
For example, last year, Zhiyuan Robot’s Expedition AI demonstrated multiple life-like application scenarios.
For instance, cracking an egg in the kitchen.
Pouring tea and water for the elderly at home is also a breeze.
Even delicate operations like preparing samples in the laboratory can be done with ease.
Similarly, at the 2023 World Robot Conference, a domestic humanoid robot called MagicLab also skillfully demonstrated coffee latte art:
Its smooth and natural movements are truly moving.
In addition to tasks requiring dexterity and precision, humanoid robots are also starting to shine in scenarios that require “strength”.
For example, Boston Dynamics’ robots are currently industry leaders in transporting objects.
Look at this agile movement, running up stairs with heavy objects and crossing a narrow bridge without hesitation.
After reaching the destination, it cleverly jumps back, using the inertia from turning around to throw the package high.
All of this makes people exclaim, “Wow! Is science fiction really coming true?”
So, why have humanoid robots emerged with various “surprises” in 2023? Behind the glamorous technological wonders, are humanoid robots really stepping from science fiction into reality?
01
A Year of Transformation
If we were to discuss the biggest variable for humanoid robots in 2023, it would probably be the application of AI large models.
From a physical dimension, humanoid robots consist of three modules: “limbs”, “cerebellum”, and “brain”, and the emergence of large models is akin to giving humanoid robots a new “brain”.
This is also the key to enabling robots to possess autonomous perception and decision-making abilities.
In other words, in the future, if humans want robots to learn a new skill, they only need to let them observe human examples and continuously learn from mistakes in practice, gradually improving their skills.
This was unimaginable before the emergence of large models.
Overall, the development of humanoid robots can be roughly divided into three stages: “humanoid”, “human-like”, and “real human”. They are currently transitioning from the “humanoid” stage to the “human-like” stage.
In the humanoid stage, humanoid robots typically rely on pre-programmed instructions and limited automation technology to execute tasks.
Imagine if you want a robot to help you cook in the kitchen, every action of the robot, such as chopping vegetables, stirring, and cooking, needs to be programmed in detail in advance.
These programming instructions include specific information on how to move the robotic arm, control force, speed, and various parameters.
Situations outside the program (like today’s meat being harder to cut) may leave the robot unable to respond effectively. Each time a new situation arises, you would have to reprogram and tell the robot what to do, which is quite inconvenient…
How does the large model solve this problem?
Specifically, the current methods for training robots with large models mainly include:multimodal learning and end-to-end learning.
Multimodal learning utilizes visual language models (VLM) for scene understanding, inputting descriptions into large language models (LLM) to obtain natural language instructions, enabling robots to perform various actions.
A prominent representative in this area is DeepMind’s RT-2 model (Robotics Transformers).
With this model, robots can learn from various videos on the internet and robot data, transforming the acquired knowledge into universal commands for robot control.
Specifically, this system is like equipping robots with a sophisticated brain and eyes. First, its eyes (VLM) can understand the surrounding environment, such as seeing a cup on the table and knowing where it is and what it looks like.
Then, its brain (LLM) will, based on the information seen by the eyes, convert image features into language descriptions using an encoder-decoder structure, describing the scene in human language, such as “pick up the cup”.
Finally, this brain will also tell the robot how to move, such as how to reach out and how much force to use, ensuring that the robot can correctly pick up the cup.
However, when specifically controlling the robot, RT-2 does not directly obtain or adjust the specific parameters of the robot hardware, such as the voltage or current of the motors. Adjustments to these hardware parameters are typically handled by the robot’s control system, while the RT-2 model provides high-level action commands.
In practical applications, these commands will be parsed by the robot’s control system and converted into specific hardware control signals to drive the robot to perform tasks.
End-to-end learning attempts to build a complete model directly from input (such as sensor data, images, videos, etc.) to output (such as actions, decisions, etc.). For instance, the previously mentioned coffee-brewing robot Figure 01 is a clear example of end-to-end learning.
In this process, the model learns all the steps by watching videos of humans brewing coffee. To achieve this, the key step is imitation learning.
Generally speaking, imitation learning can be roughly divided into the following steps:
First, the system records videos through cameras or directly collects sensor data from human operators.
For scenarios where expert action data is difficult or impossible to obtain directly, such as autonomous driving and medical surgery, imitation learning can even skip the next step and learn the corresponding actions just by watching videos.
During the specific learning process, the robot will extract key features from the observation data, which can include images, sounds, as well as the position of objects, motion trajectories, and environmental layouts.
Then, the system analyzes the data while beginning the model training, mapping input data to output behaviors.
Here, the mapping from data to actions can be seen as a physical causal relationship. For example, when the coffee machine is in a specific state (like the water temperature reaching a certain level), a specific action (like starting to pump water) should be executed.
Once the model is trained, the robot can begin to imitate the observed “causality” and continuously adjust and optimize based on feedback during the actual execution of tasks.
It can be said that it is precisely because of the support of large models that humanoid robots have truly begun to transform from mere “machines” to entities that can learn and adapt like “humans”.
An intriguing possibility is: since the data needed to train the robot’s “brain” (large model), such as video materials and examples, can generally be collected through public channels, can people in the future also use customized large models to DIY their personalized robots at home?
02
Household Servant or Factory Expert?
In the 2022 science fiction film “Finch”, the character portrayed by Tom Hanks is a survivor in a post-apocalyptic world. Knowing he has a terminal illness, he intends to use the data and materials at hand to train a robot that knows how to take care of his dog, so that after he passes away, his beloved dog can be looked after.
This is essentially a similar idea to training robots through large models.
However, to truly realize a future where everyone can DIY their robots, there are several key hurdles that humanity must cross, one of which is the cost of various “joint components”.
The total BOM (Bill of Materials) cost for Tesla’s humanoid robot Optimus is approximately $41,381. Among them, the cost of joint components is around $23,563, accounting for 56.9% of the total cost. This indicates that the value proportion of joint components in the total cost of Optimus exceeds half.
The high proportion of joint components’ value is mainly due to the need for humanoid robots to have precise and flexible joints to simulate human movements. These joints typically include complex mechanical structures, drive systems (like motors), sensors, and control algorithms, all of which are crucial to the robot’s performance.
Such high costs make it unlikely that humanoid robots will become “household servants” for ordinary people in the initial stages of commercialization, but rather more suited for production scenarios such as factories and logistics.
For example, the humanoid robot startup Figure, based in California, announced a commercial agreement with BMW to deploy Figure 01 in BMW’s manufacturing plants in the US.
This is also the first commercial agreement signed by Figure since its establishment in 2022. The humanoid robots will be deployed at BMW’s only factory in the US, located in Spartanburg, South Carolina.
While industrial robots on automated assembly lines are conventional operations, most traditional industrial robots can only perform repetitive tasks at fixed stations according to pre-programmed instructions.
For instance, although KUKA robots in Europe, especially in Germany, have performed excellently on production lines for material handling, processing, stacking, spot welding, and arc welding, these operations have all been pre-programmed. If you want to switch to another process, you need to reprogram them.
This limits the flexibility and adaptability of such robots.
In contrast, humanoid robots enhanced by large models, like Figure 01, can continuously adapt and learn through the large models, switching to any position or job, thus possessing greater flexibility and being able to perform a wider variety of tasks, especially suitable for manufacturing products designed according to human activity characteristics.
In the initial stage, Figure 01 will only handle simple tasks such as moving boxes, picking and placing, and pallet loading, while also taking on other “hard” and “heavy” tasks that few workers are willing to do, such as operating in high-temperature or noisy environments.
If successful, the deployment of Figure 01 will increase, with plans to integrate it into BMW’s manufacturing processes in the next 12-24 months, including body workshops, sheet metal, and warehouse areas.
Considering the nature of processes like body workshops and sheet metal, robots like Figure 01 are more likely to play a role in assisting assembly: helping workers secure parts or operate in confined spaces.
However, in the long run, the significance of humanoid robots in manufacturing is not limited to performing “menial” tasks. From an industrial perspective, humanoid robots entering automotive production lines undoubtedly stand at the forefront of current industrial upgrade competition.
If there is a “key point” in the competition for industrial upgrades between China and the US, it would be represented by cutting-edge technology chips and the automotive industry representing the “muscle” of manufacturing.
This is because the automotive industry is a typical “composite” industry, encompassing not only automobile manufacturing itself but also involving parts manufacturing, sales, service, and other links, with a huge market. Its development often drives the growth of the entire supply chain, significantly impacting the economy.
Moreover, the automotive industry, especially the one represented by new energy vehicles, is also a technology-intensive industry, involving artificial intelligence, new energy, mechanical manufacturing, materials science, and other fields.
Humanoid robots share many similar supply chains with industrial robots and new energy vehicles, such as batteries, chips, sensors, and controllers, allowing for collaborative migration of supply chains.
It can be said that the rise and fall of the automotive industry, from one perspective, indicates a country’s future industrial capacity/manufacturing strength.
Unfortunately, at this critical juncture, the US is currently falling behind, and humanoid robots may be the only variable capable of breaking the deadlock.
03
Manufacturing Backflow?
On the surface, the most superficial “role” of humanoid robots in manufacturing is to replace those workers who are “never satisfied” with their wages.
On September 15 last year, the United Auto Workers (UAW) contract with General Motors, Ford, and Stellantis expired, and the parties failed to reach an agreement on a new contract, leading UAW to initiate a strike involving workers from the three major automakers.
In addition to basic demands such as wage increases and shorter working hours, the core demand of the strike was actually to eliminate the unemployment risk brought about by the transition to electric vehicles—because producing electric vehicles requires less manpower than traditional cars.
At this point, a sharp contradiction emerges: the US wants to revive its physical manufacturing industry, bring manufacturing back, and upgrade its industry, but cannot withstand the demands of domestic workers who always want higher wages. If all the profits are given to workers as wages, where will there be money for industrial upgrades?
So why are American workers so eager to protest?
The underlying reason is that inflation in the US has been too high in recent years, and the wage increases are always submerged by rising prices, so workers can only keep demanding more pay.
Data shows that the inflation rate in the US reached 6.8% in 2023, the highest in nearly 40 years, with milk prices rising 29% compared to 2019.
However, behind the rising inflation rate in the US lies an unresolvable structural contradiction between reviving manufacturing and financial harvesting.
Specifically, to cope with the economic recession since the pandemic and to revive manufacturing, the US needs to implement a loose monetary policy, such as lowering interest rates to reduce production costs and stimulate investment.
However, to maintain the global attractiveness of the dollar, the US has to adopt the opposite policy—raising interest rates to attract foreign investors to hold dollar assets.
Because raising interest rates is like banks increasing the interest on deposits, people are more willing to save money in the US, and everyone wants to buy dollars because they expect the dollar to appreciate, which means they can earn more by selling dollars later. This way, global funds flow to the US because everyone thinks investing in the US is more profitable.
Thus, a situation arises where one cannot have both fish and bears’ paws.
At this time, the ever-increasing wages of workers also play a role in exacerbating this contradiction.
While raising interest rates can lead to a stronger dollar, higher borrowing costs increase the cost of manufacturing goods, which in turn raises prices. When workers see prices rising, they demand higher wages, but this leads to even higher costs for factories, making them less willing to produce in the US because it is not cost-effective.
It can be said that the “unstable existence” of human workers has become the biggest obstacle for American capitalists who want to “harvest global wealth through interest rate hikes while also wanting to bring back manufacturing”.
The potential to break this cycle is humanoid robots.
Currently, a humanoid robot costs about $20,000 to $30,000, while the annual salary of a worker in the US automotive industry is around $75,000.
With the intervention of humanoid robots, the bargaining power of unions may be weakened, allowing the US to continue its manufacturing “revival” while maintaining a strong currency.
Furthermore, when the US truly manages to bring manufacturing back (even if only partially) and bridges the gap between “creation” and “manufacturing”, it will gain an irreplaceable competitive advantage in the global supply chain due to the widespread adoption of new production technologies and the unique research and development ecosystem in the US.
04
Surrounded by Machines
Specifically, the current US has certain advantages in the research and development environment and ecosystem, such as:
—— A mature venture capital market and a culture that encourages entrepreneurship; —— A strict intellectual property legal system; —— An innovation ecosystem composed of top universities, research institutions, venture capital, and entrepreneurial culture; —— An open immigration policy and an attraction for global talent.
However, for a long time, the world market order has been dominated by the US, so it has been reluctant to compete for the middle and low-end markets by quantity and price, instead sitting back and reaping the rewards with a few high-value core technologies.
Today, as the world pattern gradually changes, the US finds that its previous international division of labor system is no longer effective, and countries like China are no longer satisfied with being low-end processing factories, but are challenging its high-end manufacturing.
This means that the US will also face homogenous competition.
In homogenous competition, if both parties’ products are similar, the key to winning is scale and price. This has been verified in history.
During the Napoleonic Wars over 200 years ago, Napoleon implemented a comprehensive “continental blockade” against Britain in an effort to economically cripple it, prohibiting all European countries from trading with it, attempting to starve Britain.
However, the result was that Britain survived.
The reason is simple: at that time, Britain had undergone the industrial revolution, significantly increasing productivity, and the prices of produced goods were extremely low. Although there were prohibitions, other countries could not resist the temptation of “low-priced goods”!
Thus, under the drive of economic laws, smuggling across European countries gradually flourished, rendering the continental blockade policy ineffective.
Looking back at the current competition between China and the US, is the pattern not very similar?
With this historical lesson, both China and the US understand the importance of price and scale. In future industrial competition, both sides will need to compete not only on the price of the products themselves but also on the price of the tools used to manufacture these products.
In the era of intelligence, such “tools” are humanoid robots that enter factories and workshops to collaborate with or even completely replace humans.
With the autonomous learning and adaptability provided by large models, along with increasingly dexterous limbs, humanoid robots may achieve the capability of “one machine serving multiple roles” in industrial manufacturing, switching between different jobs at will. This undoubtedly saves a lot of manpower and time in production.
As mentioned earlier, under the challenge from China, the high-end manufacturing sector in the US faces the risk of being overturned. How can the world return to the previous division of labor system dominated by the US?
A radical solution is to cut off the roots of Chinese manufacturing (which is known for its cost-effectiveness) using automated technologies like humanoid robots! By significantly reducing product prices while exporting large quantities of affordable US-made robots worldwide, the US can continue to create dependencies in countries with incomplete industrial chains.
A similar approach was already practiced in the 1920s when Ford revolutionized the automotive industry by introducing assembly line production, making the US the leading industrial power.
It can be said that humanoid robots are the future’s Ford production line. In a world surrounded by machines, whoever possesses lower-cost, larger-scale robots will have the upper hand in industrial competition.
05
China’s Response
In summary, in the future competition over humanoid robots, whoever can first lower the prices will gain a competitive advantage.
Fortunately, China has already shown considerable achievements in this area. Currently, the cost of humanoid robots is gradually decreasing from the previous 600,000 to 700,000 yuan to around 200,000 yuan or even lower, which is nearly one-third of the previous cost.
This significant cost reduction is largely due to the domestication of certain key components.
Currently, a significant portion of the cost of humanoid robots comes from their complex and numerous joint components. For instance, Tesla’s Optimus humanoid robot contains 28 joints from top to bottom, including 6 shoulder joints, 4 elbow joints, 4 wrist joints, and 2 waist joints…
Each of these joints costs several thousand to tens of thousands of yuan.
Additionally, to drive these joints to move freely, robots also require key components like motors, which are crucial for the rotation or movement of the robot’s joints, similar to the muscle and tendon system in humans.
In terms of reducing costs for these components, domestic humanoid robots mainly have two approaches: one is through large-scale production, which reduces the production cost of each robot. The second is through self-developed joints, drivers, reducers, and other key components, thereby lowering the costs of overseas procurement.
In terms of self-development, the prominent domestic companies include Zhiyuan Robot’s Expedition A1, Kepler’s Pioneer series, and Fourier Intelligence’s GR-1 general humanoid robot.
For example, the Expedition A1 features a dexterous hand with 12 active degrees of freedom and 5 passive degrees of freedom, adopting internal drive. It also comes with visual fingertip sensors that can distinguish the color, shape, and material of objects, with the overall cost of the dexterous hand being less than 10,000 yuan.
Here, degrees of freedom refer to the robot’s ability to move, which is the number of motion variables that the robot can independently control.
The higher the degrees of freedom, the more flexible the robot’s movements, allowing it to adapt to various complex environments and tasks.
Currently, the Expedition A1 has over 49 degrees of freedom, with 12 active degrees of freedom in the hands.
In addition to the joints, the collaborative operation algorithm also determines the strength of the robot’s capabilities.
The role of the collaborative operation algorithm in humanoid robots can be likened to the brain and nervous system of humans.
Just as the brain is responsible for thinking, decision-making, and coordinating the movements of different parts of the body, the collaborative operation algorithm processes information from sensors, formulates action plans, and completes complex tasks.
Specifically, the “eyes” of humanoid robots represent the visual processing system, the “brain” represents the decision-making system, and the “hands” represent the execution system.
How to use collaborative algorithms to make different control systems operate more smoothly and with higher data processing efficiency is the key to determining what tasks humanoid robots can perform.
In this regard, domestic company Kepler has developed the Nebula System, allowing robots to perceive their surrounding environment in real-time to solve the problems of seeing and sensing.
Specifically, it achieves this by integrating 3D visual cameras, microphone arrays, and numerous pressure and posture sensors into the robot, enabling it to clearly understand its state for executing the next operation.
Currently, through a series of self-developed technologies, domestic humanoid robots have reached a cost level comparable to that of Tesla and other star companies.
The robot startup Kepler has mentioned plans to control the price of humanoid robots at around $20,000 to $30,000 (equivalent to 140,000 to 210,000 yuan). Zhiyuan Robot estimates that the manufacturing cost of its first humanoid robot, Expedition A1, will be controlled below 200,000 yuan. Tesla’s “Optimus” humanoid robot is expected to cost around $20,000 (about 140,000 yuan).
It can be seen that domestic humanoid robots are closely competing with US companies, and there will undoubtedly be fierce clashes during the future large-scale production phase.
Overall, in the competition for humanoid robots, the US’s advantages mainly lie in its investment in cutting-edge technologies.
Startup companies focused on cutting-edge technologies provide rich funding sources and investment opportunities. For instance, the US startup Figure secured $70 million in Series A financing just one year after its establishment,
and similarly, a robotics company named 1X Technologies completed Series A2 financing in March 2023, led by the OpenAI Startup Fund, amounting to $23.5 million. In January of this year, 1X Technologies completed $100 million in Series B financing.
This kind of venture capital in the US, amounting to thousands or even hundreds of millions of dollars, is quite rare in China.
In contrast, China has greater advantages in supporting industrial policies. The core logic is essentially similar to the policies for the new energy vehicle industry, focusing on addressing the already clarified technological challenges to accelerate technology maturity and large-scale industrial development.
For example, the “Guiding Opinions on the Innovative Development of Humanoid Robots” encourages and supports research and development in key technology areas for humanoid robots, such as the “brain” (artificial intelligence), “cerebellum” (motion control), and “limbs” (mechanical arms, hands, and legs).
Therefore, overall, in the field of humanoid robots, the trend of “the US focusing on high-end technologies” and “China reducing costs” will likely persist for a long time.
However, aside from the aforementioned factors, there is an unavoidable issue that will also affect the success or failure of this industrial competition: how to balance the increasing application of robots across various industries, as many traditional jobs may be replaced by automation technologies. This will inevitably touch upon fundamental changes in social structure, the job market, and human lifestyles.
Finding a balance between robots and human labor, ensuring a smooth transition for society while advancing technological progress, becomes a “soft factor” that cannot be ignored, alongside technology, capital, and policy.
After all, machines are merely a means, while talent is the ultimate goal.
Cool Play Lab Edited
For reprints, please leave a message in the background
