The rapid growth of Tesla's humanoid robot, the focus of NVIDIA's founder semiconductor conference, and the implementation plan proposed by Shanghai Jiaotong University are making "embodied intelligence" the focus of market funds.
What is "embodied intelligence"?
Embodied intelligence is actually a fundamental problem of intelligent science, which refers to intelligence with physical experience.
From a cognitive point of view, humans are first-person intelligence, but feeding machines a lot of data for learning, belongs to third-person intelligence, such as giving machines a lot of boxes, and labeling this is a box, and then the machine will know that this mode is a box. But actually, how do humans know that this is a box? You know it through experience.
An experiment in 1963 could show the difference between the two to some extent, showing two cats in the picture, one cat is tied up and can only see the world, and the other cat connected to it can take the initiative to walk. The passive cat is a bystander intelligence, while the active cat is embodied intelligence, and at the end of the experiment, the active experience cat learned to walk normally, but the bystander cat did not gain the ability to walk.
Source: "Heart of the Machine" public account
Artificial intelligence belongs to the sum of many concepts, but some of these concepts are difficult to measure and verify, such as allowing machines to understand what is society and what is responsibility, although it can output a representation, but it is difficult to test whether the machine really understands these concepts, so it can make a closed loop on some verifiable and measurable concepts, and embodied intelligence is just such a closed loop, which is a good starting point for general intelligence.
Recently, NVIDIA founder Jensen Huang also said at the ITFWorld 2023 semiconductor conference that the next wave of AI will be "embodied intelligence", which describes "embodied AI" as intelligent systems that can understand, reason and interact with the physical world, including robotics, self-driving cars, and even chatbots, which will be smarter because they can understand the physical world.
How far away is Embodied Intelligence?
As early as 1950, Turing first proposed the concept of embodied intelligence. In the decades that followed, embodied intelligence was an important concept, but not much progress was made because the technology at that time was not enough to support its development.
Today, multidisciplinary technology has changed that. At present, various large models are blooming, and the mature technologies such as computer vision, computer graphics, natural language, and cognitive science will promote the rapid development of embodied intelligence.
The rapid development of AI large models is expected to break through the limitations and make robots "smart".
The large model of the robot includes LLM (large language model), VLM (vision-language model), VNM (visual navigation model). The "brain" AI domain of the robot is not limited to the language large model used by ChatGPT, Google mentioned in the LM-Nav study that LLM + VLM + VNM three models are combined with each other, from natural language (redundant colloquial description) to text (string of landmarks) to images (find objects in images according to text), which can finally generate the path planning of the robot. Based on this pattern of behavior, the robot can interact with humans and machines while achieving a certain degree of "adaptability".
Not long ago, Professor Lu Cewu of Shanghai Jiao Tong University delivered a keynote speech "Embodied Intelligence" at the Heart of Machine AI Technology Annual Conference, proposing the PIE scheme, believing that embodied intelligence includes 3 modules: Embodied Perception, Embodied Imagination and Execution, which is expected to accelerate the implementation of embodied intelligence.
At present, AI+ robots may be the current landing point of "embodied intelligence".
Because embodied intelligence has higher work efficiency than non-intelligent ordinary humanoid robots, its understanding, interaction, planning ability, etc., after the robot enters thousands of industries, it has strong practical landing. At the same time, its ability to be controlled with natural language is a necessary condition for assisting ordinary workers on a large scale in the future.
Therefore, in the future, we can pay attention to the types of hardware robots and application scenarios that can be transformed by large models, such as dialogue-based service robots, industrial robots, and humanoid robots in complex scenarios.
Many large manufacturers have laid out in the field of embodied intelligence, and Google released the largest generalist model in history, PaLM-E; Microsoft explores how to extend ChatGPT to robotics; Alibaba-Qianwen model is experimenting with access to industrial robots.
Among them, the Tesla humanoid robot Optimus is eye-catching.
Since the debut of Tesla humanoid robot Optimus in October last year can not walk autonomously and needs human support, on May 17, Tesla shareholders' meeting showed that Optimus can flexibly walk and grasp objects in the workshop, has environmental exploration and memory, motor torque control capabilities, AI training based on human tracking motion and object manipulation capabilities, and has opened up the FSD underlying module to achieve a certain degree of algorithm multiplexing.
TeslaBot recognizes and memorizes its surroundings while walkingSource: Computer Vision Alliance
The FSD algorithm refers to the algorithm used in its FullSelf-Driving system, which is used to realize the autonomous navigation and autonomous driving functions of the vehicle, so that the vehicle can sense, make decisions and control in various traffic environments. It mainly relies on neural networks and computer vision technology, and the core is neural network models: by processing and analyzing the data obtained by real-time sensors (such as cameras, lidars, etc.), and extracting information about roads, vehicles, pedestrians and obstacles, the environmental perception and object recognition of vehicles can be realized.
Coupled with OpenAI's previous investment in Norwegian humanoid robot company 1X, and Sanhua Intelligent Control and Green Harmonics jointly established a harmonic reducer company in Mexico, AI+ robots have made people see signs of industrial outbreak.
How big is the market?
In the short term, due to the immaturity of technology, it is difficult for humanoid robots to have clear application scenarios on the B side, and the price that is not mass-produced may be difficult for C-end users to accept, so the market is concentrated in specific consumer groups.
First of all, the robots released by ASIMO, Atlas, Tesla, Xiaomi, and UBTECH focus on their motor ability, and their ability to perform production tasks with hand and eye coordination is not much described, which means that it is difficult to enter the factory in the short term to replenish labor on a large scale. From a technical point of view, the current humanoid robot can only be based on fixed rules of movement, even if it is put into productive work, it can only be limited to limited actions and scenes, which is contrary to the expectation of "cross-scene flexible work" of humanoid robots, and needs to further mature the control algorithm.
Secondly, the current humanoid robot service ability is mainly reflected in the explanation and guidance, performance, and can not complete the housework well, in the home scene its function is more similar to the smart speaker, coupled with the higher price, C-end users may not accept a large number in the short term.
Although the practical functions are not rich enough, the early release may still attract technology enthusiasts and high-end consumers with sufficient disposable income to buy, at this time the humanoid robot meets the user's scientific research, early adopters, and show-off needs.
The price of TeslaBot in the early stage may be set at about 500,000 yuan, and the corresponding consumer groups have a high degree of overlap with the current buyers of luxury cars and ultra-luxury cars. However, considering that humanoid robots are less practical after purchase and difficult to carry out for display, the penetration rate among high-income people may be significantly lower than that of cars.
The agency expects that from 2025 to 2027, the penetration rate of TeslaBot among luxury car buyers will be 1%, 2% and 3%, respectively, and the penetration rate among ultra-luxury car buyers will be 6%, 7% and 8%, respectively, bringing market sizes of 520.5, 102.20 and 152.35 billion yuan, respectively
Source: Soochow Securities Research Institute
Later, with the improvement of technology, it will gradually help robots fill the manufacturing labor gap.
At this stage, the robot's motion control ability and endurance have been improved, and it can give full play to its advantages and undertake cross-scenario work in the manufacturing industry. According to the Manufacturing Talent Development Planning Guide, by 2025, there will be a talent gap of 30 million in the mainland manufacturing industry, accounting for the main share of the global gap.
Suppose the price of humanoid robots is 370,000 yuan, and the new penetration rate of humanoid robots to labor in the manufacturing field from 2026 to 2030 is the same. It is estimated that the cumulative replacement rate of humanoid robots for labor is 9% and 11%, respectively, and the new penetration rate will be 1.8% and 2.2% in 2030, respectively, creating a market size of 199.8 billion yuan and 244.2 billion yuan, respectively.
With the improvement of comprehensive services and emotional interaction capabilities, the penetration rate of humanoid robots may begin to increase in home scenes.
At this time, humanoid robots may be able to complete a variety of more complex housework, practical ability can be improved, coupled with the factor of price decline, at this time, not limited to high-end consumers, more families are willing to add humanoid robots at home. Under the premise that the price of robots at this stage is 250,000 yuan, the agency predicts that in the three scenarios of pessimistic, neutral and optimistic, the market size created by family scenes may reach 14,700, 18,800 and 2.3 trillion yuan, respectively, and the total market size of industrial and commercial service scenarios may reach 19,000, 25,500 and 3.16 trillion yuan, respectively.
Later, benefiting from the development of AI technology, human-computer interaction has been further improved, and it can assume the functions of accompanying and taking care of people, and may further increase the penetration rate in families with children and the elderly.
Source: Soochow Securities Research Institute
What are the enterprises related to the industrial chain?
Referring to industrial robots, AI+humanoid robots are essentially a combination of "hardware + software", although Tesla and other companies will have a certain lead, but hardware is usually also purchased.
The robot industry chain consists of four links: parts manufacturers, robot ontology manufacturers, system integrators and end users, and ontology manufacturers are in the core position. The software part of industrial robots involves the control of robots and the understanding of downstream processes, which needs to be replicable while meeting the needs of different customers, which is the core competitiveness of ontology manufacturers. The production of hardware requires economies of scale, usually by means of outsourcing.
At present, it seems that the robot industry chain related enterprises have Sanhua Intelligent Control related to actuator assembly; Rotary actuator-harmonic reducer related to Green Harmonic, Fengli Intelligence, Hanyu Group, Guomao Co., Ltd., RV reducer related double ring transmission, Qinchuan machine tool, Zhongda De; Linear actuator-torque motor related Buke shares, ball screw related Dingzhi Technology, Qinchuan machine tool; knuckle micro motor-coreless cup motor related to Mingzhi Electric, Dingzhi Technology, Jiangsu Leily; Environmental exploration - machine vision related Opt, Lingyunguang.
Source: Soochow Securities Research Institute drawing
Among them, Sanhua Intelligent Control is the world's leading manufacturer of refrigeration and air conditioning control components and components, and is also the core supplier of Tesla's automotive thermal management system, and has laid out the robot industry. In April 2023, it plans to establish a joint venture with Green Harmonics in Sanhua Mexico Industrial Park, the main business is the research and development, manufacturing and sales of harmonic reducer related products.
Green Harmonics is engaged in the research and development, design and production of precision transmission devices, focusing on harmonic reducers, mechatronics products, industrial automation and other products. Harmonic reducer is one of the core components of robots, the company has broken the monopoly of international brands in the field of harmonic reducers for robots, and achieved mass export.
Mingzhi Electric's main business is to control the motor and its drive system, the control motor is the core industrial equipment, the company has broken the Japanese monopoly, is the only domestic enterprise that changes the global competition pattern of HB (hybrid) stepper motor within ten years. Its subsidiaries Anpu Mingzhi, Switzerland Tmotion, Mingzhi Paisil Boisil deeply layout the mobile robot industry.
At the same time, there are also Hongsoft Technology, which focuses on the field of computer vision, provides algorithm authorization and system solutions for the industry, and provides visual algorithm product lines for intelligent driving of intelligent terminals on a global scale, Hikvision, a leading intelligent Internet of Things with machine vision, artificial intelligence and navigation control as the core, and Dahua Co., Ltd., a leading player in video Internet of Things, are all machine vision-related beneficiary enterprises.
In addition, there are also large model-related beneficiary enterprises such as Chuangda, SenseTime, Cloudwalk Technology, and iFLYTEK.