Text | Qingbo
Editor|Tang Xiaoyuan
The text totals 8627 words
Estimated reading time 6 minutes
There is a counterintuitive phenomenon in the humanoid robot industry:
At present, the players of humanoid robots do not include the four well-known industrial robot families (Fanuc, ABB, Yaskawa, and Kuka), nor the top domestic robot giants (Eston, EFORT, Guangshu, Sinsong, Huichuan, Xinshida, Qianjiang).
At present, the players in the leading position of the humanoid robot body are the big guys in the automotive industry, such as Tesla, Honda, and Toyota. Or technology companies such as Xiaomi, Google, Amazon.
Traditional industrial robot giants such as KUKA have rarely entered the market.
But on the other hand, when the humanoid robot concept stock ushered in a blowout, the first to detonate was the old three industrial robots - controllers, servo motors, and precision reducers. For example, Tuopu Group of production line joints; Sanhua intelligent control of rotating joints; Naruzhi Electric of Hollow Cup Motor; Jiangsu Rayleigh of sliding lead screws, planetary roller screws and hollow cup motors; servo motor Huichuan Technology; The green harmonics of the harmonic reducer are all good performance with the help of the humanoid robot concept.
To answer this counter-intuitive phenomenon, Starship Zhizo also has to analyze from the perspective of the industrial chain of humanoid robots that are both traditional hardcore (the old three types of industrial robots) and cool frontier (artificial intelligence) -
1, Why is humanoid robot an industry of tomorrow and how to understand its huge potential?
2, The moment of the humanoid robot iPhone is far from coming, what areas is it stuck in?
3, The current situation of each link of the humanoid robot industry chain
4, GPT technology in humanoid robot scene application and humanoid robot ultimate conjecture
First, humanoid robots and embodied intelligence
There is such a dialogue in Asimov's "Steel Cave" to the effect of:
If you're managing a farm, you have two options. One is to install an "electronic brain" on tractors, harvesters, upholsters, cars, milkers, and other machinery, making them intelligent machines. The second is to keep the harvesters, dumpers, cars, and milkers as they are, but use a robot with an "electronic brain" to operate them. What will smart people choose?
We believe that the above text image shows the meaning and advantages of humanoid robots:
Instead of redesigning all the tools, it is better to produce robots that can mimic the way humans look and behave. TA perfectly accesses the existing life of mankind, enters all the scenes where human beings already exist, and uses all the tools that humans are using.
In a word, all our lives and production do not need to make any superfluous and additional changes for this robot.
source：giphy
As an industry integrating mechanics, electricity, materials, computers, sensors, control technology and other disciplines, humanoid robots are an important symbol of national high-tech strength and development level. Developed countries around the world do not hesitate to invest heavily in development and research.
Japan, the United States, South Korea, the United Kingdom and other countries have invested heavily in the development of humanoid robots
At Tesla's shareholder meeting in May, Musk showed investors around the world the latest progress of Tesla's humanoid robots through a video:
Optimus can already complete complex tasks such as sorting items, such as picking up items, environmental discovery, and memory. Compared with its debut at the end of September last year, its motion control capabilities continue to evolve and AI capabilities improve. Musk believes that Tesla's future long-term value may be given by Optimus.
Musk's total wealth has surpassed the president of French luxury giant LVMH to regain the title of the world's richest man
The new robot ASIMO, newly developed by Japan's Honda, is 120 centimeters tall and weighs 43 kilograms, and it walks closer to a person.
Source:Honda
South Korea's KAIST representative product HUBO+ is 1.7m tall, weighs 80kg, and has 32 degrees of freedom.
source：KAIST
Atlas, a representative product developed by Boston Dynamics, is 1.8m tall, weighs 80kg, and has 28 joints, which can complete difficult movements such as jumping and turning in place for a week.
In terms of hardware structure, Atlas has lightweight structural skin and foot force control sensors, radar and depth cameras to form visual perception, 28 hydraulic joint drives to complete a series of agile actions, and the body is equipped with 3 NUC/industrial computers responsible for the calculation of the overall control system.
In terms of software, Boston Dynamics uses behavioral libraries, real-time perception, and model predictive control (MPC) technology to analyze data received by sensors, such as cameras and radar, and support decision-making and action planning.
Source:Boston Dynamics I have to say that Boston Dynamics is cool
In the field of scientific research, there are "pioneer" robots developed by the National University of Science and Technology, which can complete static and dynamic walking actions; The "HIT-III" robot launched by HIT can complete actions such as going up and down slopes; "THBIP-II" developed by Tsinghua University has 24 degrees of freedom, etc.
In the industrial field, the "Walker" robot launched by Shenzhen UBTECH can complete actions such as going up and down steps; Xiaomi's humanoid robot Cyber One (Tieda) announced in August 2022, the upgraded motion control algorithm dominates the robot's 13 joints and 21 degrees of freedom throughout the body to achieve bipedal movement posture balance; The motor performance is 10 times enhanced, and the peak power torque of the main motor of the hip joint can reach 300Nm, and the peak torque density is 96Nm/kg; The humanoid robot displayed by Zhejiang Laboratory can use a high-precision visual perception system to locate piano keys for high-precision finger movements.
Shenzhen UBTECH "Walker" robot
Drafting: Starship Chizo
Musk's judgment on the robot market is not "out of nowhere". IFR data, the global robot market size has reached a high of 51.3 billion yuan in 2022, with a CAGR of 14% from 2017 to 2022.
Among them, the market size of industrial robots is 19.5 billion US dollars, the market size of service robots is 21.7 billion US dollars, and the market size of special robots exceeds 10 billion US dollars. By 2024, the global robotics market is expected to exceed $65 billion.
According to Markets and markets, the global humanoid robot market size (only considering single machines) will increase from US$1.5 billion in 2022 to US$17.3 billion (about RMB103.8 billion) in 2027, and the capacity of the 100 billion market is just around the corner.
From the retail price point of view, the four-legged robot Spot Mini launched by Boston Dynamics in the United States was priced at about $74,500, and about 400 units were sold in a year after it was launched, and the humanoid robot Atlas was not listed. Faced with high prices, Musk has publicly stated that his humanoid robot will be sold for less than $20,000 and achieve mass production.
Make the simplest calculation: even if the future sales of Optimus are only 1/100 of Musk's 10 billion units, then its market size is 20,000 * 100 million = 2 trillion (US dollars), which is a spectacular number.
Source: Tesla
The current iPhone moment for humanoid robots is far from coming.
From a broad perspective, humanoid robots are not shackled by machines in essence, and even if they have the shape of a human body, they are still a kind of robot. However, after being equipped with five systems: perception system, drive system, terminal execution system, energy supply system, computing system and software, humanoid robots have made great progress in artificial intelligence compared to industrial robots that we are more familiar with.
Traditional industrial robots are more like a machine than a "human". Because the trajectory of its operation is programmed in advance, it can only be done in a fixed area and a specific place to do prescribed and repetitive actions, and does not have the ability to independently deal with sudden (non-programmatic) problems, and can only be classified as a non-intelligent robot.
The need to interact and react to the environment (including people) requires humanoid robots to have a considerable degree of artificial intelligence. To describe it in more standardized academic language, such an artificial intelligence is called embodied intelligence.
Embodied Intelligence refers to the ability of an agent to understand and transform the objective world through its own learning after interacting with the environment.
The hypothesis that embodied intelligence can be produced on the premise that intelligent behavior can be learned by agents with corresponding forms by adapting to the environment. In view of this, all living things on the earth can be said to be embodied intelligence.
Speaking at the 2023 ITF World Congress, Jensen Wong said: "Together, AI and accelerated computing are transforming the technology industry. The next wave of AI will be a new type of AI known as embodied AI, an intelligent system that can understand, reason, and interact with the physical world, known as embodied intelligence. ”
Embodied intelligence is an important way to improve the cognitive capabilities of current "weak artificial intelligence". It is also a possible path to produce super artificial intelligence.
Embodied intelligence can be understood as a high-order form of artificial intelligence development, and it may also become the ultimate form of AI. To put it more thoroughly, the future of humanoid robots is truly intelligent robots that empower embodied intelligence.
How would such an intelligent humanoid robot receive instructions to complete its task?
First of all, the robot's perception system transforms the robot's various "internal state information" and "environmental information" from "signals" to "data" and "information" that can be understood and applied by the robot itself or between robots. That is, let the robot understand its surroundings.
Scientists (manufacturers) install various sensors on the robot, including light, sound, temperature, distance, pressure, positioning, contact, etc., so that the robot can approach humans through the "five senses" and collect information that perceives the outside world.
The rapid development of various sensor application technologies has ensured the development of humanoid robots.
The continuous improvement of sensor accuracy and reliability is one of the most pressing problems to be solved. For example, after entering water, dust or bumps, the calibration of the camera is easy to fail, and hardware problems such as pixel necrosis after long-term use will directly lead to the loss of accuracy of the machine recognition rate.
Second, the robot's drive system and end-execution system need to work together to perform their own duties, just as humans rely on 78 joints on various parts of the body to act.
The technical threshold and cost of each joint used in the robot are very high, which not only requires small size, high precision, light weight, but also requires anti-fall and crash-resistant. When the robot moves rapidly, the driving force output power is very high, to ensure that it will not burn out due to heat problems, and at the same time has a cushioning ability to protect the "robot joint" from impact.
The "hand" in the end-execution system is a very difficult part of the technical difficulty. In the actual operation process of grasping, lifting, holding, pinching and other actions, the coordination of robot finger softness and gripping strength is very important: it is easy to fall off if you grasp something light, and it is directly broken if you grasp it heavily.
For example, Tesla's Optimus shows a robotic hand that is very similar to a human hand:
Optimus has 11 fine degrees of freedom, combined with control software, to perform operations as complex as a human hand, carrying a load of about 9 kg. In the latest demo video, we can also see certain advantages of Tesla robots in hardware, including the ability to control the grip of many items, and will not break eggs.
Third, the robot's drive system and energy supply system are used to ensure the completion of its actions. According to the description of Boston Dynamics' official website: Atlas robots, although all kinds of showy moves are cool, but must be equipped with a very powerful hydraulic drive. In fact, Atlas is equipped with 28 hydraulic drives to allow the robot to complete a variety of explosive actions, and this price is also an old problem of Boston Dynamics.
Manufacturing costs remain high, making it difficult to get out of the lab to complete commercialization.
Musk chose a stable and more cost-effective motor drive solution on Optimus for cost considerations. It is precisely because of this that the Optimus target price is set at $20,000 to meet the huge potential market demand.
Finally, the robot's computing system and software are used to analyze and think and issue corresponding instructions.
First, understand the needs and environment. The robot uses sensors to understand its surroundings and figure out what to do.
The difficulty of this layer lies in the recognition and understanding of the environment such as vision, including the recognition of unknown objects and the unknown posture of objects.
Second, dismantling tasks and path planning.
The difficulty of this layer is the instability of the AI output. Because when artificial intelligence disassembles tasks, the solution may be different each time, which may lead to inconsistencies in task disassembly and produce unexpected results.
Third, the command drives the hardware to perform the task. It is necessary to convert the motion planning into mechanical instructions sent to the robot's drive system, determine the energy, momentum, speed, etc. are appropriate, and start the task.
The difficulty of this layer is that the current artificial intelligence can not do 100% accuracy, and the speed is slow, time-consuming and labor-intensive, but the drive of hardware execution tasks requires precise control, and it is necessary to use 100% accuracy as the basis, so the industry will still use traditional cybernetics to give orders and perform tasks.
It can be seen that it is not easy to organically integrate the four major hardware and a large software system on the humanoid robot, work together, and cooperate with each other, while also allowing the right part to have the right power, speed and accuracy to complete the required task.
Again, take Tesla's Optimus. In August 2021, Musk publicly demonstrated Optimus' idea for the first time at Tesla's annual AI Open Day. Only a year later, the "Optimus Prime" prototype was released on Tesla AI Day on September 30, in terms of hardware, "Optimus Prime" is 172CM tall and weighs 73KG overall; The walking power is 500W, the sitting power is 100W, and the overall parameters are slightly different from the 2021 concept machine (concept machine parameters: height 172CM, weight 57KG, load 20KG, walking speed up to 8 kilometers per hour).
In terms of motor drive, Optimus has a 2.3KWH, 52V voltage battery pack, and an integrated unit with built-in electronic and electrical components, which supports humanoid robots to work all day; Choose 28 custom articulated drives, reuse automotive powertrain design experience, design 6 types of joint drives, including 3 different specifications of servos (using harmonic reducer) and 3 different specifications of linear actuators (using permanent magnet motor, can lift 1.5 tons of grand piano), to find the best combination of cost and efficiency.
At present, Tesla has applied some of the car's technology to the Optimus, such as battery packs, cooling systems, etc., and also uses similar technology to car testing to carry out robot motion and external collision simulation.
Tesla's fully self-driving system, FSD, is also directly applied to Optimus, where the robot uses the same visual perception as a car, using a camera to input data and using a neural network to perform calculations.
In the future, there may be the following surprises for Tesla humanoid robots:
Tesla has strong research and development capabilities in new materials and electronic technologies, and may use Tesla's technology accumulation in rockets; It brings continuous explosive topics to Tesla.
Second, the analysis of each link of the industrial chain of humanoid robots
Let's start with the conclusion. The competition between humanoid robot products in the future is, in the final analysis, a competition between various ontology manufacturers on the artificial intelligence (AI) track.
The first is the level of artificial intelligence: whether it can understand the information and instructions issued by the external environment and internal centers like people, understand, judge and respond correctly.
The second comparison is the ability to work together between various software and hardware: even if each component is the best in its class, the assembly is not good, the matching is improper, and the best practical effect is also not played.
Once again, it is the cost competition after mass production: humanoid robots are a commodity after all, and if their final cost is not acceptable to the market, then no matter how good it is, it is at best a laboratory product with no commercial value.
Humanoid robots, as an important branch of the new rise in the field of robotics, its industrial chain is also composed of three parts: upstream, middle and downstream.
The upstream is the core software and hardware, including servo motors, reducers, controllers, sensors, etc.; Software mainly includes machine vision, human-computer interaction, machine learning, system control, etc.
Midstream is a manufacturer of humanoid robots, that is, manufacturers that can combine upstream core software and hardware to produce humanoid robots that can walk, squat, run, jump, pick up things, and grasp objects. Domestic mainly include UBTECH, Beijing Iron and Steel Technology, Xiaomi, etc., while foreign ones include Boston Dynamics, American Agile Robot, Tesla, Toyota of Japan, Honda, etc.;
At present, there are no particularly mature commercial applications in the downstream
From the current industrial chain of humanoid robots, although Musk's Optimus has already injected a super booster into people looking forward to its official listing, in the real market, its upstream industry chain is still the route of conventional industrial robots.
In addition to the sensor, the other three main hardware parts are the three core components of industrial robots (controller, servo motor and reducer). From this, it can be found that even if the current humanoid robot is hot, the limelight, and the "human" flavor, the final use is still the basic architecture of the industrial robot.
One of the core components of the robot that can work flexibly lies in the precise control of motion, and the three core components with the highest technical difficulty are the controller, servo motor and precision reducer, which account for about 15%, 20% and 35% of the cost respectively.
Take the first humanoid machine released by Tesla in September 2022, the Human, Optimus, as an example:
It has 28 joint actuators, enabling the whole body to produce more than 200 movements at different angles.
With five fingers and 6 actuators, the hand has human dexterity.
To achieve these functions, the performance of the core components of the above three industrial chains must meet quite high requirements.
From the cost analysis, the cost of humanoid robots can be roughly divided into three parts: powertrain system (accounting for 60% of the total cost), intelligent induction system (accounting for 20%), structural parts and others (accounting for 20%).
Among them, the powertrain system includes the battery system and the electric drive system (the powertrain definition here is similar to the "three electric systems" on electric vehicles), which are expected to account for 10% and 50% of the total cost respectively.
From the perspective of comprehensive technology and cost, the importance of core components is particularly prominent.
On the one hand, the essence of humanoid robot technology is high-dimensional perception and movement in 3D space, and high-performance core components are the basis for perception and movement.
On the other hand, the three most important core components, controllers, AC servo motors and precision reducers, still account for a fairly high proportion of the cost.
The controller is the brain of the industrial robot and plays a decisive role in the robot's performance. The industrial robot controller mainly controls the movement position, attitude and trajectory of the robot in the working space, the operation sequence and the time of action;
The servo motor is used as an actuator in the automatic control system to convert the received electrical signal into angular displacement or angular velocity output on the motor shaft. Each joint movement of the robot needs to be driven by a servo motor to achieve multi-degree of freedom movement;
The precision reducer is an intermediate mechanism connecting the power source and the actuator, which has the role of matching speed and transmitting torque, and is also the component with the highest barrier in robot production, mainly including harmonic reducer and RV reducer, but its working principle and application scenarios are quite different.
Six-axis industrial robot source: AOBO robot
In addition to the above three core components, the role of sensors in humanoid robots is also extremely important.
To have the ability to recognize the external environment and realize functions such as navigation, obstacle avoidance, and interaction, humanoid robots need to use sensors to identify objects and measure distances.
Sensors such as cameras and lidar are mainly used to identify the external environment, and their sensor solutions and demand scenarios are similar to those of autonomous driving, so there are new opportunities for autonomous driving sensors with high value.
Autonomous driving sensor manufacturers and machine vision manufacturers are all participants in humanoid robot sensors.
The way to classify the three core components is usually more suitable for relatively rough industrial robots, and to better analyze humanoid robots, the division of the five systems is actually more scientific.
The perception system, that is, sensors, such as cameras, microphones, lidars, distance sensors, etc., is responsible for collecting and transmitting information, which is equivalent to human eyes, nose, and ears, that is, people's "five senses";
Drive systems, such as motors that drive wheels, hydraulic power systems or pneumatic systems on robotic arms, are equivalent to human "muscles";
The end execution system, which can be either a manipulator, or a screwdriver or spray gun on a robotic arm, is used to physically interact with the external environment, equivalent to human "hands" or "limbs";
The energy supply system, such as a power supply or battery, is the energy supply center of the robot;
The computing system and software integrate all the above systems to complete the specified task, which is the "brain" of the robot.
The reason why it is necessary to spend a considerable amount of space to introduce the core components, software and system parts of the upstream industry chain is because they are the key to humanoid robots that can run and squat lightly, grasp and lift flexibly, identify objects and people intelligently, and obey the words.
Without the careful cooperation of these parts, the most realistic humanoid robot can only be regarded as an embroidered pillow, which is not useful.
source：pexels
Compared with a wide variety of upstream industry chains, the midstream industry chain of humanoid robots is simple and clear: those ontology manufacturers that can combine these core components and software to enable robots to exert maximum humanoid efficiency.
As mentioned at the beginning of this article, the players who are currently in the leading position on the humanoid robot ontology are either automotive industry bigwigs like Tesla or technology companies like Xiaomi. On the contrary, industrial robot bosses rarely enter the game.
In the author's judgment, although humanoid robots are born from industrial robots, because the former's requirements in the field of AI are much higher than the latter, it is a latecomer advantage for high-tech companies (especially those with higher achievements in the field of AI).
Of course, it is too one-sided to classify Tesla, Toyota, and Honda as automobile companies, which are high-tech companies with strong scientific research and development capabilities (especially Tesla's super strength in fully autonomous driving technology FSD, which has given humanoid robots the same visual perception as cars: input data with cameras and calculations through neural networks).
Third, ChatGPT materialization and the ultimate conjecture of bots
In a recent humanoid robotics competition, EVE, produced by OpenAI-backed physical robotics company 1x, beat Tesla's Optimus robot.
Some of the software functions of EVE bots are supported by ChatGPT, which means that ChatGPT is materialized and applied in real-world scenarios.
The application of GPT technology in humanoid robots has opened up new imagination space and aroused the high attention of the domestic capital market to the robot industry.
From the perspective of industry insiders, the application of GPT large model technology to humanoid robots can provide "common sense" for robots to understand and reason.
The middle model for robot operation can be arranged at the edge end, which can process the sensing information of the robot relatively quickly, and combine the motion planning and control small model deployed on the robot end, so as to build an intelligent robot system integrated with "cloud and edge" and complete the intelligent closed loop of humanoid robots.
This once again consolidates our judgment that the competition between humanoid robot products is, in the final analysis, a competition between various ontology manufacturers on the artificial intelligence (AI) track.
First fight AI research and development capabilities, then compare the collaborative work capabilities between software and hardware, and finally everyone fights for costs together.
Starship Zhizo believes that humanoid robots will have three development trends in the near future:
First, the cost is getting lower and lower, and it is the general trend to enter thousands of households.
In view of the current high retail price of humanoid robots (not mass-produced state), even if the prospects of its downstream industrial chain are long-term optimistic by the capital market, at least at present, there are only a few scenarios that can achieve commercialization.
Boston Dynamics' Atlas is worth $2 million, Honda Asimo's is worth $2.5 million, and Xiaomi Cyber One costs as much as 600,000-700,000 yuan. It basically eliminates the luxury of ordinary people.
However, if Optimus is mass-produced in 3-5 years, the price is $20,000 a piece, and it replaces humans to complete trivial and complex work (not simple and repetitive type), such a downstream industry chain scenario will undoubtedly be quite extensive and of great commercial value.
As Musk described, after mass production reaches the level of Tesla vehicles (several million units) and various more economical and practical alternatives are put into the original expensive and vulnerable parts, the price of $20,000 per unit will only be the initial goal. After all, Karl Friedrich Benz's invention of the world's first internal combustion engine car wasn't enough to get it into millions of homes, the mass-produced Ford Model T was.
Second, intelligence is getting higher and higher, and replacing complex and cumbersome work is the will of the people.
Unlike industrial robots, which are mainly used to replace simple, repetitive and dangerous labor positioning in factories, humanoid robots are more suitable for replacing relatively more intelligent, cumbersome, and personal jobs that have direct contact with humans, such as shopping guides, reception, care, and feeding pets. (For more robots entering downstream applications, please click to read the Starship Zhizuo article: "In the trillion-level robot track: why is it the most volatile?") 》）
With the improvement of the level of artificial intelligence, more and more AI technologies are applied to humanoid robots, and robots are becoming more and more "humanized", and the closeness and compatibility with humans will become higher and higher.
Third, the surpassing of human intelligence by artificial intelligence will be a high probability event in the future.
Both the explosive ChatGPT and the AlphaGo, which defeated Lee Sedol, are weak artificial intelligence.
Strong artificial intelligence has not yet appeared. But on the whole, artificial intelligence has the following three advantages in the confrontation with the human brain:
The first is energy consumption, artificial intelligence thinks and answers questions much less energy than the human brain. The brain is the largest energy consumption organ in the human body, why human beings always hate learning and thinking, this is determined by survival instinct, in order to avoid excessive consumption of body energy caused by overthinking.
The second is storage, which people forget. But AI stores information as long as it is entered. Strong AI can also connect more knowledge and information through learning.
The last and most fatal advantage is the time efficiency of artificial intelligence. There are two aspects of understanding of efficiency here, one is learning efficiency, compared to human needs entertainment, socialization, sleep, etc., AI 24 hours of sleepless learning and evolution, yesterday or baby tomorrow is an adult is the strongest brain. The second is the efficiency of solving problems, artificial intelligence to deal with problems and work around the clock, in the future artificial intelligence will be more proficient than humans in the use of various tools, may be your lifetime to master the craft of operating precision machine tools, AI will learn overnight.
Finally, let's go back to Asimov's robot trilogy.
In "Steel Cave", the middle-aged plainclothes police officer Baile asked the robot Daniel why are humans so persistent in creating robots with hands, feet and heads? Why do they always have to act more or less like individuals?
Daniel's answer was simple. Everything invented for humans so far can be used directly by humanoid robots.
So, yes, everything is based on economic considerations.
Starship Zhizuo, the same name on the whole platform. Welcome to follow
This article is the original content of Starship Zhizo