Since Tesla released the humanoid robot Optimus in October 2022, the industry process has accelerated significantly. This is the beginning of the robot series, through the review of the 50-year development history of the robot industry, the positioning of the historical process of humanoid robots, and the discovery of the law of industrial development and the key variables of the current stage.
summary
The domestic substitution of industrial robots is in the ascendant, and the commercialization of service robots continues to be explored.
The development of operation control technology is mature, and domestic substitution is in the ascendant. In 2022, the Chinese market will occupy half of the world's market, and the pattern will be dominated by the four major foreign investments, and domestic substitution will be the general trend. The three core components of controller/servo system/reducer in the industrial robot industry chain account for 70% of the cost, and the localization rate has reached 40.6%/27.2%/33.6% in 2021.
The requirements for perception and control have been improved, and domestic manufacturers have led some tracks. Some scenarios with simple scenarios & low technical difficulty & high frequency are progressing rapidly, and domestic manufacturers have good global competitiveness, mainly due to supply chain cost reduction, technological comparative advantages and economies of scale. The cost of sensing systems in the service robot industry chain accounts for 40%+, and the domestic industrial foundation is good.
Humanoid robots are the ultimate "general-purpose" robots, and the hardware solutions have not yet been gathered, and the software capabilities need to be improved urgently. We believe that humanoid robots with humanoid forms and artificial intelligence may be the ultimate answer in order to achieve generalization in the robotics industry. Compared with traditional robots, the hardware of humanoid robots is upgraded to high availability and low cost, and the current solutions are in full bloom and have not yet converged, and there are differences in the selection of actuator components, the type and number of sensor applications, and the software is upgraded to multiple perception & intelligent decision-making, and the AI large model is expected to drive the intelligent and rapid iteration of humanoid robots. Recently, the industrialization process has continued to accelerate, and the design optimization and cost reduction of core hardware represented by planetary roller screws, and the iterative progress of AI+ robot large models are the key variables.
risk
The development of key technologies is less than expected, the commercialization progress is less than expected, the mass production progress of leading manufacturers is less than expected, and the cost reduction of the industrial chain is less than expected.
body
Robot status: domestic substitution of industrial robots is in the ascendant, and the commercialization of service robots continues to be explored
Robots: Multifunctional programming machines, which have a long history and are diverse
Robots, as a kind of high-level comprehensive technology in automation, have experienced long-term technology accumulation, from the advent of the first recognized robot in 1959 to modern robots, according to the technical form can be divided into three generations: in 2017, the emergence of Transformer architecture laid the foundation for the mainstream algorithm of large models, image recognition, natural speech processing and other AI technologies accelerated breakthroughs, robots to a higher level of human-computer interaction, multiple perception and autonomous decision-making development, general intelligent humanoid robots ushered in the dawn.
Figure 1: Three stages of development of robotics
Source: HG, Intelligent Robot System, CICC Research Department
Industrial robots: the industry is mature, and domestic substitution is in the ascendant
In addition to technology-based, the development of the robot industry also requires the close integration of manufacturing and application. Looking back on the 70-year history of industrial robots, thanks to the fact that only simple and repetitive work needs to be completed in a closed environment, industrial robot technology quickly reached availability, labor shortage + strong demand brought by the development of downstream automobile industry, and the industry quickly ushered in the promotion and application under the support of national policies.
Figure 2: Global industrial robot installed capacity and historical review
Note: Except for Japan, the sales data of industrial robots in various countries before 1993 are temporarily missing.
Source: IFR, CICC Research
Up to now, there are various types of industrial robots, which have been widely used in handling, welding and other processes in automobiles, electronics and other industries. The core application scenarios are simple and repetitive labor links such as handling (55% of the volume, 2021 data, the same below), welding (25%), etc., which are mainly used in industries such as automobiles (30% of the volume, 2022 data, the same below), electronics (35%-40%), and metal machinery (15%).
China's market demand accounts for half of the world's total, and the process of domestic substitution is accelerating. In 2022, the global sales of industrial robots will be 550,000 units, a year-on-year growth rate of 5.5%, and in 2022, China's industrial robot sales will be 290,000 units, a year-on-year increase of +5.5%, a new high. China is the largest and fastest growing market for industrial robots, the global robot industry chain is accelerating the transfer to the mainland, and domestic robots led by Estun have the advantages of price, channels, and response speed, starting from the less difficult collaborative robots and DELTA robots.
Industrial robots are divided into three parts: control, machinery, and sensing, of which the three core components of controller, servo system, and reducer account for 70% of the cost. From the point of view of value, reducer, servo system, and controller occupy 35%\25%\10% of the value respectively, and it is also the link with the highest barriers in the industrial chain, which directly determines the accuracy, stability, load capacity and other important performance indicators of the industrial robot system.
The industrial robot industry chain is mature, and domestic substitution has become the main theme. Thanks to the advanced industrial manufacturing capabilities of Japan and Europe and other countries, industrial robots are the first to usher in mature applications, and the four major foreign investors build moats through self-developed and self-made core components, and the three core components of reducer, servo system and controller are mainly dominated by foreign suppliers. With the continuous high growth of China's industrial robot market, the rise of local manufacturers, and driven by the demands of supply chain security, cost control, differentiation/customization, etc., the localization of the industrial chain has become the main theme. According to Tiger Sniff statistics, as of 2021, the localization rate of controllers/servo motors/reducers has reached 40.6%/27.2%/33.6%, a year-on-year increase of +4.1/2.7/2.4ppt, and domestic substitution is in the ascendant.
Figure 3: China's industrial robot system structure, key components and localization rate
Source: "Industrial Robot Technology Basics" (Sun Shudong, 2006), iResearch, Tiger Sniff, OFweek Robotics Network, CICC Research Department
We believe that industrial robots mainly take controllers, servo motors, reducers and other hardware as core components, and after years of development, the technology has matured and can be migrated to solve the operation and control problems of humanoid robots. These hardware are mainly monopolized by foreign capital, and domestic substitution has become the main investment opportunity.
Service robots: Perception and control requirements are improving, and domestic manufacturers are leading some tracks
During the epidemic, the service robot market continued to expand, and Chinese companies led the way in some tracks. According to IFR data, the scale of the global/Chinese service robot industry in 2021 will be 17.2/4.9 billion US dollars, a year-on-year increase of +24.9%/+33.4%, and the CAGR from 2017 to 2021 will be +27.1%/+42.2%, and the Chinese market accounts for about 30% of the global market. In the fields of unmanned delivery, sweeping robots, and medical sanitization, China has given birth to a number of global leaders in subdivisions.
In general, the scene is more closed, the technical difficulty is low, and the use frequency is high. Before 2017, the technical barriers of the service robot industry are not high, and it is in the early stage of industrialization, and the products are mainly simple superposition of sensors in closed scenarios, including indoor distribution/welcome exhibition/household cleaning/warehousing and logistics, etc., with many market participants and fierce competition; We expect that after 2030, the core technology will be advanced to high-end artificial intelligence and high-precision control units, and leisure and entertainment/education science/elderly care/housework in open scenarios will be popularized. Basic scenarios that are frequently used are technologically evolving.
Figure 4: Service Robots by Use Case
Source: Yinke Financial Research Institute, CICC Research Department
The requirements of service robots for perception systems and software algorithms are upgraded, and the requirements for operation control are downgraded. Compared with industrial robots, the changes of current service robots include: 1) mobile robots are the mainstay, mobile components are added at the hardware level, mainly in the form of mobile chassis, and real-time perception and positioning and navigation systems need to be developed at the software level. 2) The number of Toc scenarios has increased, and the requirements for human-computer interaction functions have been significantly improved. The core cost of dedicated service robots comes from sensors (mainly lidar, cameras, and ultrasonic radars) and algorithm platforms, accounting for 40%~50%, with perception and obstacle avoidance, navigation and positioning, and human-computer interaction as the core modules. We believe that the key to the cost reduction of dedicated service robots in the future lies in the price decline and localization of core sensors.
Figure 5: Component composition of a service robot
Source: Infineon, CICC Research
Figure 6: Distribution of the value of service robots in 2022
Note: The inner ring is part of the qualification environment perception and operation control part, and the outer ring is all procurement
Source: Yiou Think Tank, CICC Research Department
We believe that compared with industrial robots, the incremental technology of service robots lies in the intelligent upgrading of human-computer interaction systems, sensing modules and decision-making links, which has laid a certain industrial foundation for the perception and decision-making links of humanoid robots, but the overall technical route is still iterative and has not yet converged, and the development of the industry is still in the early stage, so it is necessary to continue to pay attention to the marginal changes in technology and industrial trends.
Humanoid robots: truly "universal" machines, the industry is waiting to be released
The deep integration of humanoid hardware + high intelligence refers to the real "universal" robot
Robots are in the stage of specialization development, and starting from the endgame thinking, they will surely give birth to general-purpose robots and bring subversive development to the industry. We believe that if the robot industry is compared to the communication electronics industry, the robot industry is still in the stage of BB machines, business and other special planes, and the smart phone represented by the iPhone is already a general terminal for personal communication and entertainment. The entire robot industry may usher in potential market expansion, product software and hardware iteration acceleration, and industry pattern reshuffle.
General robots need the organic combination of "intelligence" and "body", and intelligent humanoid robots may be the final answer.
► "Organism": From first principles, the human form may be the final form. The human world is designed to facilitate human activities, and we believe that machines that imitate humans to the greatest extent may become the most widely adapted type of robot to the production and life of the human world.
► "Intelligence": With the blessing of AI, self-learning and decision-making are possible. Based on general-purpose algorithms, humanoid intelligent robots have the intelligent attributes of autonomous learning, planning, decision-making, and execution through generative AI and GPT large language models, and can be applied to multiple scenarios.
Robot bodies and artificial intelligence form a flywheel effect, embracing the era of "embodied intelligence". In the future, humanoid robots are expected to become the best carrier of embodied intelligence, providing massive multi-dimensional training data for AI, and AI can improve the reliability and accuracy of models based on data training, which can further improve the working performance and scenario applicability of robots, and further promote the wide application of robots in all walks of life. We believe that in the long run, humanoid robots and AI promote each other to form a flywheel effect, which is expected to accelerate the arrival of the era of embodied intelligence.
Figure 7: Flywheel effect between robot bodies and artificial intelligence
Source: Research on Morphology-based Embodied Intelligence: Historical Review and Frontier Progress (Liu Huaping, 2023) CICC Research Department
Commercialization review: Half a century of ups and downs, the dawn of commercialization has emerged
Figure 8: The iteration process, technical solutions, application scenarios and commercialization progress of the main products of humanoid robots
Source: Companies' official websites, CICC Research Department
The birth of humanoid robots can be traced back to the 70s of the last century, and movement, perception and interaction were the early development priorities, resulting in high costs and difficulties in commercialization. Restricted by the technical capabilities related to operation control, increasing the number of joints often means high costs, and robots may not be able to carry out actual work due to the low degree of freedom due to the difficulty of realizing the functions of handling, picking, and going up and down stairs, or face the inability to mass produce due to high cost. Honda's ASIMO humanoid robot was introduced in 2000.
Another way to position the entertainment home scene and dilute the ability to exercise, but the commercial landing space is still limited. In order to seek commercialization, some manufacturers position the application of their products in entertainment scenarios such as scenic spots, speeches, press conferences, and home companionship, emphasizing interactive functions and diluting operation and control capabilities. Although this solution significantly reduces costs and makes commercial use possible, the application scenario space is limited, and once it is found that the demand is not enough to support it, it is difficult to extend the life of the product. Pepper was launched by SoftBank Group in 2014.
Hardware planning matches application functions, and the pace of implementation in logistics and inspection scenarios is fast. The products that are the first to be sold to the market and have the leading commercialization progress have commonalities in the application scenarios, and generally focus on logistics warehousing and security inspection. In terms of hardware functions, in order to realize the functions of carrying boxes, going up and down stairs, and distributing in the above-mentioned practical applications, it is necessary to configure a reasonable number of joint motors in the corresponding arms, legs, and dexterous hands to achieve the degree of freedom matched with the core functions. Digit launched by Agility Robotics in 2019.
With the upgrading of hardware industrial capabilities and the breakthrough of AI technology represented by large models, car companies are betting on humanoid robots, startups have been born, and the commercialization of humanoid robots has a tendency to accelerate. Tesla's Optimus has been rapidly iterating, from the concept of Tesla Bot in August 2021 to the video of Optimus moving smoothly and pinching eggs with two fingers with hand tactile sensors released in December 2023, it took almost 2 years from the birth of the concept to the start of industrialization. In addition to the entry of technology companies such as Tesla, Xiaopeng, and Xiaomi, they have gradually accelerated from the laboratory to the commercialization of humanoid robots after weighing the operation control plan and mass production costs, and gradually launching market-oriented products.
Industrial chain foundation: high degree of freedom and intelligent characteristics, bringing mechanical solution optimization and intelligent new increments
The organic combination of hardware integration and software algorithms builds the "humanoid" characteristics of humanoid robots. Humanoid robots involve three modules: perception, AI decision-making, and operation control and execution. The decision-making module includes chips and algorithms, involving the migration and upgrading of human-computer interaction, task understanding, trajectory planning, motion control and other technologies, and the perception and actuation system is mainly composed of various hardware, including rotary joints, linear joints, dexterous hands, torque sensors, vision sensors, cameras, IMUs, etc. Taking Tesla Optimus as an example, actuators and sensors are the core components, accounting for more than 70% of humanoid robots, and hardware is still the focus of the current robot industry.
Figure 9: Breakdown of the value of humanoid robot hardware (taking Tesla Optimus as an example)
Source: Tesla AI Day, CICC Research
On the basis of the industrial and service robot industry, humanoid robots put forward higher requirements for energy density, lightweight, precision and durability of parts and components such as drive, transmission and sensing, and key components are expected to be applied on a large scale, and related targets enjoy market expansion dividends;
Joint design has evolved from rigid actuators to elastic actuators and collimated actuators. Rigid joints use conventional brushless motor + high transmission ratio reducer (generally harmonic reducer) + torque sensor + double encoder, and collimated drive generally adopts high torque density motor (generally frameless torque motor) + low transmission ratio reducer (generally planetary reducer) + single encoder. Tesla's Optimus actuator uses a rigid scheme, and robots such as Zhiyuan and Xiaomi use a collimated drive scheme.
► Motor: Continuing the motor drive scheme, high energy density motor becomes the first choice. In addition to Boston Dynamics, which pursues performance, most humanoid robot manufacturers continue to mature electric drive solutions. In the motor sector, frameless torque motors (torso joints) and coreless motors (dexterous hands) with high energy density have basically become the mainstream development direction.
► Deceleration: The linear joint with the planetary roller screw as the core has begun to be used in a large area, and there is a possibility of change in the selection of the rotary joint reducer. Previously, LOLA of the Technical University of Munich and RH5 of the German Artificial Intelligence Research Center have used a combination of rotation + linear actuators, but Tesla Optimus is the first humanoid robot to use linear joints in a large area. Linear actuators are more compact and durable than rotary actuators, which can achieve low energy consumption and high loads, and are more suitable for leg needs.
• Rotary joint: generally motor + reducer, but for the sake of self-weight, inertia and space, RV reducer is less used, and harmonic reducer/precision planetary reducer is the main scheme. Tesla Optimus rotary joints (rotary actuators) are divided into three categories according to different output torques (20Nm, 110Nm, 180Nm), but they are all composed of customized permanent magnet motors (frameless torque motors) + angular contact ball bearings + harmonic reducers + crossed roller bearings + torque sensors + position sensors (double braided) + drivers + mechanical clutches;
• Linear joints: can be achieved with planetary roller screws/T-shaped screws. Tesla Optimus linear joints (linear actuators or linear drives) are also divided into three categories according to different output torques (500N, 3900N, 8000N), but they are all composed of permanent magnet motor (frameless torque motor) + ball bearing + four-point angular contact bearing + reverse planetary roller screw / T-shaped screw + force sensor + position sensor (single knitting) + driver + joint CNC parts.
Exhibit 10: Humanoid robot joint solutions (as of March 2024)
Note: Rigid/elastic/collimated drive bid* is expressed as an inference based on joint composition
Source: Companies' official websites, CICC Research Department
Dexterous hands of incremental components: the early development plan is diversified, and the technical route is not converged. Similar to the importance of both hands to human beings, dexterous hands have become an important indicator to measure the overall strength of bipedal robots, and micro motors, precision transmissions, and sensitive sensors have been intensively used, which are the core incremental components of humanoid robots. There are three main types of dexterous hand drives: (1) pure motor drive, each knuckle is equipped with a micro-motor to rotate, (2) motor + connecting rod combination drive, (3) motor + rope combination drive. At present, most of the drive methods are a combination of motor and connecting rod. For example, the Tesla Optimus coreless joint is composed of a coreless motor + multi-stage planetary reducer + worm gear + position sensor (double braided) + driver + metal tendon rope.
Exhibit 11: List of Dexterous Hand Solutions by Major Robot Manufacturers (as of March 2024)
Source: Companies' official websites, CICC Research Department
The working scene is complex + the terminal operation is diverse, and the amount and type of sensors are higher than those of industrial robots and service robots. Compared with industrial robots, humanoid robots need to make real-time judgments on scenes and paths, and have higher requirements for external sensors such as vision and touch, and compared with service robots, humanoid robots have higher requirements for end-effector operation, and the demand for compliance control and posture control also gives rise to the demand for higher-level internal sensors.
► IMU: Driven by the demand for attitude balance, IMU is expected to usher in applications on humanoid robots. IMUs are not common in industrial robots and service robots. The biggest challenge of bipedal robot walking is dynamic balance, and the IMU assists the "brain" in making decisions by sensing and feeding back the posture and acceleration of the body space. In addition, small-range ground speed movement scene positioning and limb acceleration feedback are also one of the important functions.
► Force sensor: Driven by the demand for hand supple control and steady walking on both feet, six-dimensional force transducer may be widely used. Torque sensors are often used in the internal control of robot manipulators, and more precise six-dimensional force sensors are more used in some industrial robot end-effectors and collaborative robots. Humanoid robots may be equipped with six-dimensional force sensors on their hands and feet due to their higher degrees of freedom and more precise requirements for walking and dexterous hand control.
► Tactile sensors: driven by the demand for precise operation of dexterous hands, electronic skin may usher in applications. The research and exploration of electronic skin began at the beginning of the 21st century, but it is not common in the application of industrial robots, and the industry as a whole is in the early stage of development.
► Visual sensing: Service robots have laid a good foundation, and Tesla's pure vision solution is outstanding. Thanks to the existence of mobile robots such as food delivery robots and sweeping robots among service robots, the visual SLAM of 3D active vision has a certain development foundation, and most humanoid robot manufacturers follow the scheme. Tesla's Optimus parallelizes the vehicle-end vision scheme and uniquely selects a pure visual passive vision scheme.
Exhibit 12: Statistics of mainstream robot sensor solutions (as of March 2024)
Source: Companies' official websites, CICC Research Department
According to the classification in "Robotics: Modelling, Planning and Control", the control of robots can be divided into four levels: task level, action level, initial level, and servo level.
Figure 13: The humanoid robot control process can be divided into four levels
资料来源：《Robotics：Modelling, Planning and Control》（Bruno Siciliano，2009），中金公司研究部
Task-level + action-level: The large model has stronger generalization ability, which can improve the versatility of humanoid robots. In traditional robots, task-level and action-level control are completed by professional algorithm engineers, while with the support of large models, task-level and action-level can be independently understood and planned by robots, and only need to send task instructions. The significance of task-level interaction is: 1) to increase the speed of development, thereby increasing the versatility of humanoid robots. 2) Lowering the threshold for use, users from engineers to ordinary users, may be expected to drive the industry to the starting point of rapid development.
Initial level + servo level: In the short term, it is difficult to overcome the real-time problem, and in the long term, there is a possibility of end-to-end large models. The advantages of large models lie in common sense understanding, logical reasoning, generalization ability, language communication, etc., but they are not suitable for low-level precise motion control, and have weaknesses in real-time, trajectory smoothness, accuracy, and observability. We believe that the current mainstream robot large model is biased towards task understanding and splitting, such as SayCan, PaLM-E, imageBind, etc., combined with the actual results in theoretical feasibility, in the short and medium term, "high-level large model + bottom-level robot algorithm" will become the most promising direction, but the model represented by RT-2 is still exploring the possibility of end-to-end large model landing, and if realized, it is expected to greatly improve the versatility of robots.
Figure 14: Comparison of different types of robots
Source: Elephant Robotics, CICC Research
Risk Warning
The development of key technologies has not been as expected. The key technologies related to humanoid robot software and hardware are still under development, and technological breakthroughs are needed to make the performance and cost of the robot meet the needs. If the progress of software and hardware technology is not as expected, it may lead to poor performance, limited functions or safety problems of the robot, which in turn affects the industrialization process.
Commercialization is not progressing as expected. The promotion of humanoid robots is affected by technology, user acceptance, scenario demand and other aspects, and the commercialization progress is uncertain, which may have an adverse impact on the industrial chain.
The mass production progress of the leading manufacturers is not as expected. The industry is currently driven by the supply side, and the mass production progress of humanoid robots from leading manufacturers such as Tesla has a leading role in the industry.
The cost reduction of the industrial chain is less than the expected risk. At present, the cost of humanoid robot solutions is high, and each core component has a large room for cost reduction, and the progress of cost reduction is not as expected will affect the downstream large-scale application.
Article source:
This article is excerpted from: "Tesla Humanoid Robot Tracking (1): Where Evolution Goes?" published on March 21, 2024
Deng Xue Analyst Automotive & Mobility SAC License No.: S0080521010008 SFC CE Ref: BJV008
Chen Xianfan Analyst Machinery & Military SAC License No.: S0080521050004 SFC CE Ref: BRO897
Jing Lan Wei Analyst Automotive & Mobility Equipment SAC License No.: S0080522080010
Cui Lidan Contact Person Automotive & Mobility Equipment SAC License No.: S0080123070118
Ren Danlin Analyst Automotive & Mobility SAC License No.: S0080518060001 SFC CE Ref: BNF068
Yuan Mu Analyst Automotive & Mobility Equipment SAC License No.: S0080523110002
Guo Weixiu Analyst Machinery & Military SAC License No.: S0080521120004 SFC CE Ref: BSI157
Wang Zilin Analyst Machinery and Military SAC License No.: S0080523080005
Legal Notices