The White Paper on Exploration and Development Prospects of Embodied Intelligence in China |
| 时间：2025-07-13 21:37:10 来源：chinaunicom 作者：chinaunicom |
The White Paper on Exploration and Development Prospects of Embodied Intelligence in China
Preamble Looking back on the evolution of human civilizations, we may find the three industrial revolutions, like break of dawn, eventually evolved into a wave of changes that have redefined the world landscape. In the 18th century, the improvements in the steam engine broke through the energy shackles. The 19th century witnessed electrical technology dramatically reshaped the production. The 20th-century information revolution caused cognitive restructuring. The revolution history reveals that despite doubts and wait-and-see attitudes always coming along at the early stage of technological revolutions, those sprouts of record-breaking innovation will eventually grow into towering trees, advancing the development of the times. As the wheel of history enters the critical point of the fourth industrial revolution, embodied intelligence, as a cutting-edge paradigm for the deep integration of artificial intelligence (AI) and the physical world, is becoming a key fulcrum for intelligent technology to break through the virtual boundaries. Embodied intelligence breaks the "disembodied cognition" limitation of traditional AI. Equipped with physical carriers, AI systems can perceive, understand, make decisions and take actions in real environments, taking a leap from pure information processing to physical world manipulation. In this way, it provides AI with a "body" for understanding and exploring the physical world and expands the intelligent application boundaries, achieving "unity of knowledge and action" in the real sense. This White Paper focuses on the forward-looking trends in the development of the embodied intelligence industry, and provides in-depth insights into the applications and future prospects of embodied intelligence in China. In the short term, we will focus on structured scenarios mainly in manufacturing, with breakthroughs achieved first in industrial applications; in the medium and long term, we will expand our horizons to wider commercial scenarios, including medical care, logistics and public services, to play a greater role in the business ecosystem; in the long term, we are looking forward to the future when embodied intelligence reaches thousands of households as the most intimate partner of mankind. The development of embodied intelligence is just in 2The White Paper on Exploration and Development Prospects of Embodied Intelligence in China its infancy and, just like every industrial revolution, requires several generations to run in relay. The wheel of the times is rolling forward and the prelude to intelligence has been sounded. We hope that China's exploration in embodied intelligence can provide experience for the BRICS countries to draw on in their future development, thus benefiting their people and promoting the building of a community with a shared future for mankind. 3The White Paper on Exploration and Development Prospects of Embodied Intelligence in China Members of the Writing Group Lead Organizations: China Unicom Research Institute, China Unicom Global Limited Members of the Expert Advisory Committee Jian Qin, General Manager of the China United Network Communications Group Co., Ltd. Meng Shusen, Chairman & General Manager of China Unicom Global Limited Li Hongwu, Director of the China Unicom Research Institute Chen Hui, Director of China Branch of BRICS Institute of Future Networks Wei Jinwu, Vice Director of China Unicom Research Institute Zhang Liren, Vice Director of China Branch of BRICS Institute of Future Networks Key Members of the Writing Group: Guo Yaolong, Yang Jinzhou, Xia Fan, Xue Jing, Luo Lizhuo, Yang Zheng, Zhou Yining, Han Yingying, Fu Yuyu, Yu Donghai, Lin Zhen, Li Jichao, Zhang Yawen, Ou Qiqi Participating Organizations: China Branch of BRICS Institute of Future Networks Shenzhen UBTECH Robotics Corp. Ltd. Shanghai AgiBot Innovation Technology Co., Ltd. AI² Robotics China Mobile Smart Home Operation Center 4The White Paper on Exploration and Development Prospects of Embodied Intelligence in China I. New Development Trend of Embodied Intelligence As a revolutionary breakthrough in the interaction between AI and the physical world, embodied intelligence reshapes the global technological competition landscape with disruptive innovation, accelerates the formation of new quality productive forces integrating digital and real economy, gives birth to new production relations featuring human-machine collaboration and reconstructs the underlying logic of the digital economy. It serves as a core engine driving the leap from the digital economy to an autonomous economy, and provides stronger resilience for the sustainable development of the digital economy. (I) Embodied Intelligence Opens up New Space for Digital Economy Development 1. What Is Embodied Intelligence Embodied intelligence is an intelligent system that realizes perception, cognition, decision-making, and action through the interaction between physical entities and the environment. It can achieve intelligent growth and adaptability through autonomous learning, and can perform tasks to affect the physical world. Embodied intelligence has four major elements: ontology, agent and data, as well as learning and evolutionary framework. First, "ontology" means that embodied intelligence requires humanoids and any other forms of physical entities. Physical entities affect cognitive processes, such as thinking and decision-making, reflecting that cognitive needs coexist with the body. Second, "agent" can actively perform such tasks as perception, understanding, decision-making, and control, and improve knowledge and skills through autonomous learning. Third, "data" comes from body sensors, including cameras, microphones and tactile sensors, but also from the external environment. Last, "learning and evolutionary framework" means that an agent interacts with the physical world and is enabled to perceive, understand and transform the physical world through multimodal large models, reinforcement learning 5The White Paper on Exploration and Development Prospects of Embodied Intelligence in China (RL), imitation learning and other methods. Figure 1 Four Elements of the Evolutionary Framework for Embodied Intelligence (Source: China Unicom Research Institute) Embodied intelligence is a deep integration of AI and robotics. AI serves as a "brain" for robots, while robots provide the "body" for AI. Over respective development, both help each other advance while evolving at the same pace. AI has gradually evolved from rationality to deep learning. After the emergence and evolution of large models, AI has demonstrated powerful capabilities in understanding, reasoning and creativity. However, it is also faced with development bottlenecks. Increasing the quantity of parameters and data for model training can hardly improve the model performance fundamentally. For breakthroughs, AI must be connected with the real physical world. Embodied intelligence is an important path for AI's journey towards artificial general intelligence (AGI). Embodied intelligence is not exactly the same as an agent. The two may converge at some point, but their focuses are different. An agent, emphasizing autonomy and goal orientation, can autonomously perceive the environment and take action to achieve specific goals. An agent includes software agents and virtual agents in the virtual world, such as the chatbot ChatGPT, the general AI agent Manus AI and virtual digital humans, as well as intelligent entities in the physical world, such as intelligent robots. Embodied intelligence exists in the form of various physical entities in the physical world, and has a variety of body forms according to different uses and scenarios, such as humanoid robots, quadruped 6The White Paper on Exploration and Development Prospects of Embodied Intelligence in China robots and L4 autonomous vehicles. Among them, humanoid robots, a typical embodiment, are widely regarded as the best form for their applications, with a focus more on the interaction between the physical form of the agent and the environment, and abilities to perceive, change, learn and adapt to the environment. 2. Development of Embodied Intelligence According to the mainstream view in the academic circle, embodied intelligence has gone through five major development stages so far, starting from a budding period as a theoretical concept, undergoing technology accumulation and verification by all parties in the global industry, and gradually moving towards industrialized application practice. From 1950 to 1980, when the concept of embodied intelligence was in its infancy, Turing proposed in his papers where AI possibly moved toward, laying the foundation for the concept of embodied intelligence. The prototype of modern robots is a pre-programmable robotic arm created by the Massachusetts Institute of Technology (MIT) in the United States in 1954. General Motors of the United States first put the industrial robot Unimate into production lines for welding in 1960. Stanford University and Waseda University have developed the first mobile robot and a humanoid robot that can talk to each other. From 1980 to 2000, it entered the stage of early exploration and theoretical development. The research of Rodney Brooks & Rolf Pfeffer et. al provided important theoretical support for embodied intelligence. Thanks to breakthroughs in computer hardware and sensor equipment, Japan's Epson company launched the home robot AIBO, and MIT launched the robot Kelmet with functions similar to human hearing, vision and proprioception. In 1991, the six-legged robot Genghis appeared and could autonomously walk. In 1999, Sony launched the robotic puppy AIBO. From 2000 to 2010, at the stage of interdisciplinary integration and technology accumulation, embodied intelligence research began to integrate interdisciplinary methods and technologies, such as mechanism science, machine learning and robotics, basically forming an independent discipline branch. In 2002, a Danish 7The White Paper on Exploration and Development Prospects of Embodied Intelligence in China company released the first robot vacuum Roomba. From 2011 to 2020, it witnessed technological breakthroughs. The rapid development of deep learning technology injected new impetus into the development of embodied intelligence. Numerous bionic and humanoid robots emerged to help robots adapt to the natural environment. Boston Dynamics launched the legged robot "BigDog", dragonfly robots and soft octopus robots. From 2022 to now, its industrial applications are gaining popularity. Large language models represented by ChatGPT have brought great potential for embodied intelligence in terms of intelligent perception, autonomous decision-making and even anthropomorphic interaction. Gradually moving towards industrial applicationsdrives the development from special robots to general robots. Various forms of ontology, such as collaborative robotic arms, mobile operating robots, bionic dexterous hands, and unmanned taxis, have demonstrated the trend of intelligence upgrades. With the support of capital, many innovation companies in embodied intelligence have emerged, such as UBTECH, Figure, and Optimus. In the next 10 years, embodied intelligence will gradually shift from scientific research and commercial scenarios to being applied in thousands of households, realizing "one brain with multiple functions" and "one brain with multiple machines" and expanding to more open scenarios. Figure 2 Development History and Outlook of Embodied Intelligence (Source: China Unicom Research Institute) 8The White Paper on Exploration and Development Prospects of Embodied Intelligence in China 3. Embodied Intelligence Promotes Iterative Upgrading of Industrial Digitalization As a revolutionary integration of AI and robotics, embodied intelligence reconstructs the productivity paradigm through multidisciplinary cross-innovation and becomes the core engine for cultivating new quality productive forces. Embodied intelligence is the "general hub" for the cross-integration of cutting-edge disciplines. The rise of embodied intelligence is an inevitable result of the respective development of AI and robotics technology to a certain degree, manifesting the deep integration of the two. To realize magnification effect, superposition effect and multiplication effect, AI should have the ability to interact with the physical world. What's more, robots will also develop toward more advanced humanoid interactions and intelligence in the future. Embodied intelligence gives AI a "body" that can interact with the physical world and through interactions, actively explore, understand and change the world, marking that we are about to enter a new era characterized by the "unity of knowledge and action". Intelligence, no longer limited to inanimate algorithms and data, will instead be closely intertwined and symbiotic with the real world. Its development and application will have a complex and far-reaching impact on society in every aspect. Embodied intelligence is a "new target" for AI to empower thousands of industries. As a core carrier of interaction between AI and the physical world, embodied intelligence is reconstructing industrial productivity through multimodal perception fusion and adaptive decision-making control technology. Embodied intelligence has widely expanded and deepened the AI application scenarios, promoting thousands of industries to move towards higher-quality development. Especially in main industrial scenarios, such as industrial manufacturing, people's livelihood services, agricultural planting, and extreme operations, the development of embodied intelligence is expected to lead the industry into a new stage of the flexible manpower line, achieving a higher AI penetration rate in the physical world. Embodied intelligence is the "training ground" for developing the AI industry ecosystem and talents. To achieve better development of embodied intelligence, it requires opening up the entire industrial chain from chips, algorithms and mechanical manufacturing at the upstream, to downstream scenario applications, and promoting 9The White Paper on Exploration and Development Prospects of Embodied Intelligence in China the construction of new infrastructure, such as low-latency, lossless networks and intelligent computing clusters, providing underlying support for the AI industry to take a leap. At the same time, embodied intelligence helps talent training cover the capability matrix of robot control, algorithms, and models, accelerating the exponential growth of versatile AI talents. (II) Broad Prospect in Global Embodied Intelligence Market �. Different Focuses in the Strategic Layout of Embodied Intelligence in Various Countries Major economies around the world attach great importance to the development of embodied intelligence. New policies are continuously introduced to support its development and raise the whole society's awareness and support for robotics. The United States is focusing on basic research in AI, while maintaining its leading position at the forefront of embodied intelligence. The National Robotics Initiative 3.0, Data, Analytics, and Artificial Intelligence Adoption Strategy released by the Department of Defense, has been released successively, elevating AI technology competition to the national security strategy. Since 2025, several American robotics companies have also urged the United States to launch a "national strategy" for developing robots and establish a federal office to promote the robotics industry development to prevent the United States from falling behind competitors such as China. The EU is actively advancing the development of embodied intelligence, with a focus on its safety and transparency. It successively launched the Public-Private Partnership in Robotics and European Approach to Artificial Intelligence. The EU has also built an ethical framework and industry norms for the development of embodied intelligence based on the Artificial Intelligence Act. In Asia, Japan is integrating robots into society and making them a key part of its social foundation. In 2022, the Japanese government increased its financial support to 105.7 billion yen for robot-enabled application scenarios in manufacturing, services and medical services. Against the backdrop of an aging society, robots replacing manpower is a 10The White Paper on Exploration and Development Prospects of Embodied Intelligence in China key support area, with the rollout of such policy systems as the Key Areas for Adoption of Robots in Social Infrastructure and International Standards for Safe Operation of Service Robots. South Korea has certain technology strengths in robot products or services, but is relatively lagging behind in robot core components or software technology. It adopted the Intelligent Robots Development and Distribution Promotion Act in 2008, and in 2023, released the strategy for robot industry development and the Plan for Innovation in Advanced Robot Regulation. 2. Rapid Development of Global Embodied Intelligence Market The global market for embodied artificial intelligence is currently in the nascent stages of a significant and rapid expansion, indicating considerable potential for future growth. According to projections from industry analysts, including Market.US, the market is anticipated to reach a valuation between $3.0 and $3.3 billion in 2025. It is projected to continue its upward trajectory at a compound annual growth rate (CAGR) of 15% to 18%, with a conservative estimate suggesting the market could expand to approximately $10.75 billion by 2034.The year 2025 is poised to be a landmark year, signaling the commercial debut of humanoid robots, which are expected to be the primary catalyst for the future growth of the embodied AI market. The market for humanoid robots is forecast to be valued at approximately $2.0 billion in 2025 and is projected to experience a remarkable CAGR of 40% to 50%. This substantial growth trajectory is expected to position the humanoid robot segment to gradually assume a dominant role within the industry. Figure 3 Global Market Size and Outlook of Embodied Intelligence (Source: Market US) 11The White Paper on Exploration and Development Prospects of Embodied Intelligence in China In terms of investment and financing, data from CB Insights shows that from 2020 to 2025, the total amount of financing raised by global embodied intelligence-related companies has exceeded US$30 billion, with a year-on-year increase of 68% in 2024. Here are recent milestones in investment and financing. SoftBank Group Corp. intends to invest up to one trillion U.S. dollars in American industrial intelligence projects. Figure 4 Overview of Foreign Company Financing 3. Accelerated Implementation of Embodied Intelligence by Technology Giants from Various Countries Overall, embodied intelligence products around the world have all already entered the product launch and application verification stage, but there is still a long way to go before commercialization and mass production at a large scale. American technology giants have announced the launch of multiple embodied intelligence robots. Tesla, Inc. released the humanoid robot Optimus Gen 2. In June 2024, Optimus Gen 2 had been used in Tesla's factory for battery sorting training. Boston Dynamics released the humanoid robot Atlas, which has also been used in car factories to perform such tasks as moving car pillars. Agility Robotics released the humanoid robot Digit. Digit can complete daily warehouse work such as handling goods, palletizing, and unloading, and has been applied in GXO and Amazon. The unicorn company Figure AI empowers humanoid robots through the OpenAI large model and launches the robot Figure 02, which can complete complex tasks such as folding clothes, cleaning tables, and packing shopping bags. It has also been tested in car factories of Bayerische Motoren Werke (BMW) to carry out parts assembly and other work. 12The White Paper on Exploration and Development Prospects of Embodied Intelligence in China Japan, Singapore, Spain and other countries have also launched embodied intelligence-related products successively. The tipping point for the breakthrough of embodied intelligence technology is coming soon. In the future, it will surely drive the comprehensive industrial upgrading and profoundly reshape the social form. Countries are going all out to participate in this fierce competition that reshapes the global competitive landscape to have an advantaged stance in the new global technology map. (III) China's Accelerated Breakthroughs in Embodied Intelligence �. China's Policies-led Innovation Action of Embodied Intelligence At the national level, embodied intelligence has been elevated as a national strategy, and relevant policies continuously rolled out, offering more favorable conditions and providing strong support for the development of embodied intelligence. In the Implementation Plan for "Robot plus" Application Action issued by China's Ministry of Industry and Information Technology and 17 other ministries in January2023, it is the first time that goals are set for robot development, including realizing more than 100 new robot application technologies and solutions, and promoting more than 200 typical robot application scenarios. Subsequently, the Guiding Opinions on Innovative Development of Humanoid Robots was issued. Focusing on the key frontier field of humanoid robots, it clarified the development directions such as strengthening the forward-looking development of basic standards, improving the innovation ability of humanoid robots and building an innovation system for humanoid robots. At the local level, more than 20 cities in China have made a plan to develop embodied intelligence, and many places have begun to build industry innovation centers for humanoid robots. Some first-tier cities have been the first to take actions based on their strengths in technology research and development, the industrial base and innovation ecosystem. Beijing issued the Beijing Embodied Intelligence Technology Innovation and Industry Cultivation Action Plan (2025-2027), established 13The White Paper on Exploration and Development Prospects of Embodied Intelligence in China the Beijing Humanoid Robot Innovation Center, and built a mother platform for hardware of general humanoid robots. Shanghai Zhangjiang High-tech Park has built the first heterogeneous humanoid robot training facility, and it is expected to accumulate 10 million high-quality embodied data sets by 2025. Guangzhou has established an industry alliance for intelligent equipment and embodied intelligence robots, focusing on core technologies such as AI chips and motion control, and promoting the realization of application scenarios in industrial manufacturing and domestic services. �. Capital Boosting the Development of China's Embodied Intelligence Industry China's embodied intelligence industry is experiencing leapfrog development. Overall, the market size of China's embodied intelligence is increasingly growing, nearly tripling in five years from 292.3 billion yuan in 2018 to 748.7 billion yuan in 2023, with a compound annual growth rate of up to 20.7%, and is expected to exceed one trillion yuan in 2026. According to the forecast by the China Business Industry Research Institute, the embodied intelligence market is mainly occupied by robots and autonomous vehicles (such as self-driving cars), accounting for 55.6% and 44.4% respectively in 2024. As the most important carrier of embodied intelligence, humanoid robots will become a robust driver of the growth of the embodied intelligence market. According to (China International Capital Corporation Limited) CICC's forecast, China's humanoid robot shipments are expected to reach 350,000 units in 2030, with a compound growth rate of 317% from 2024 to 2030; the market size is expected to reach 58.1 billion yuan in 2030, with a compound growth rate of 259%. 14The White Paper on Exploration and Development Prospects of Embodied Intelligence in China Figure 5 Forecast of Market Size of China's Embodied Intelligence Industry (unit: Hundred million yuan) (Source: 36Kr Research Institute) 3. Multi-party Participation in the Practice of China's Embodied Intelligence Industry Many entities in China have participated in embodied intelligence industry, mainly divided into three categories: robot companies focusing on robot R&D, Internet technology companies that take a part through self-research and cooperation, and automobile companies dedicated to autonomous driving. Robot companies promote the integration of AI technology and robotics technology to make robots more intelligent. For example, Shenzhen UBTECH Corp., Ltd., a leader in the humanoid robot industry, has launched Alpha and Walker series robots. With its core customers and business scenarios customized for education, logistics, consumption and other industries, it is the humanoid robot company that has developed cooperation with the largest number of car manufacturers in the world. Internet technology companiesrely on their advantages such as capital and AI technology to participate in the embodied intelligence race directly or through capital contributions. For example, Beijing Xiaomi Technology Co., Ltd. has developed humanoid robots such as CyberOne and is committed to promoting the development of the consumer robot market. Tencent has set up the Robotics X robot laboratory to study the integration of multimodal large models and robots. Automobile companies promote readiness of autonomous driving technology to achieve commercialization. 15The White Paper on Exploration and Development Prospects of Embodied Intelligence in China Autonomous driving companies, such as Apollo Go, have begun commercial operations in many places. On the other hand, they provide capital support for robot companies. For example, SAIC Motor Corporation Limited participated in the A-round strategic financing of LimX Dynamics through its subsidiary Shang Qi Capital, and BAIC Group participated in the angel round financing of GALBOT through BAIC Capital. 16The White Paper on Exploration and Development Prospects of Embodied Intelligence in China II. System Architecture of Embodied Intelligence Industry The embodied intelligence industry chain is centered on the integration of physical entities and AI, presenting a three-level linkage structure of "upstream-midstream-downstream". The upstream focuses on core hardware and basic materials to form the physical carrier of embodied intelligence. The midstream is a hub for technology integration, which transforms hardware and algorithms into the "brain" and "cerebellum" of agents. The downstream promotes the scenario penetration rate with diversified products, revolutionizing productivity and lifestyle. The embodied intelligence industry chain is evolving in a coordinated manner of "hardware-software-scenario" to drive the global transformation from industrial automation to intelligence-enabled life. In the future, it will rely on large models, multimodal integration and policy support to accelerate large-scale implementation. Figure 6 China's Embodied Intelligent Industry Chain Map (Source: China Unicom Research Institute) 17The White Paper on Exploration and Development Prospects of Embodied Intelligence in China (I) Upstream Infrastructure and Raw Materials: Physical Reconstruction of Intelligent Base The upstream technology base of the embodied intelligence industry reconstructs the intelligent system architecture through physical entities. Core components such as chips, sensors, controllers, motors, communication modules, and energy management constitute the physical interaction system of agents. The computing power chip accelerates the neural network with a heterogeneous architecture; multimodal sensors integrate multiple information to build an environmental digital twin; servo motors achieve micron-level precision control with closed-loop feedback; industrial-grade communication modules ensure millisecond synchronization of equipment. The material, process and reliability verification of upstream hardware directly determine the upper limit of midstream system functions and the implementation efficiency of downstream solutions. 1. AI Chip: the Heartof Embodied Intelligence As the heart of embodied intelligence, chips mainly provide powerful computing power to support complex algorithm execution. At present, AI chips have made a series of technological progress, which is mainly reflected in the following aspects. First, architectural innovation includes storage and computing integration, breaking the bottleneck of "memory wall" and improving energy efficiency. The neuromorphic chip simulates the pulse neural network of the human brain and processes dynamic data with low power consumption. Heterogeneous computing, a CPU+GPU+ASIC combination, such as and Tesla Dojo D1 chips. Second, as for process and packaging, Taiwan Semiconductor Manufacturing Company Limited has announced mass production using its 3 nm process technology, improving chip density and energy efficiency; costs of Chiplet technology have been reduced. Third, with the rise of edge AI chips, lightweight models drive demand for low-power edge chips. Fourth, with algorithm-hardware collaborative optimization, the Transformer architecture has given birth to dedicated accelerators, such as Groq's LPU. Omdia expects the market value of robotic AI chipsets to reach 866 million U.S. 18The White Paper on Exploration and Development Prospects of Embodied Intelligence in China dollars in 2028. In the AI chip market, competition is extremely fierce, and global manufacturers are accelerating innovation to consolidate their position or seize market share.Tesla, with Dojo supercomputer + self-developed FSD chip, vertically integrates the full stack of automotive AI. Qualcomm/Intel Corporation enters the automotive and robotics market through acquisitions. The dual driver provided by China's policy suppor t and marke t demand is promp ting local AI chip manufacturers to accelerate their layout and strive to overtake on the curve. For example, Horizon Robotics' Journey series occupies 65% of the autonomous driving chip market share, competing with Tesla's Dojo supercomputer, and its partners include BYD and Li Auto. Huawei Ascend 910 chip is based on 7nm-enhanced technology, with a maximum power consumption of 350W and a computing power up to 256TFOPS. It has full-stack AI capabilities and focuses on smart cities and industry. The Cambricon's Siyuan series has deployed cloud training, but the progress on the edge is slow. China and other BRICS countries have the absolute advantage of mineral resource reserves and efficient production capacity that are most urgently needed in the global semiconductor industry. The current third-generation semiconductor materials are mainly represented by silicon carbide (SiC) and gallium nitride (GaN). China has a complete monopoly on germanium, antimony, gallium and graphite. In 2023, China produced 60% of the world's germanium, 48% of antimony, 98% of gallium and 77% of natural graphite. AI chips is still facing increasing challenges, which are mainly reflected in energy efficiency and power consumption. Embodied intelligence needs to run 24/7, but the power consumption of existing chips can still hardly meet the needs of passive devices. As for real-time performance and low latency, autonomous driving requires an end-to-end response of <100ms, and multi-sensor fusion places extreme demands on the parallel computing capabilities of chips. Dynamic environmental adaptability requires a hardware architecture that supports online learning, and traditional ASIC fixed algorithms cannot adapt to changing scenarios. For safety and reliability, automotive-grade/industrial-grade chips need to pass ISO26262 certification, and their anti-electromagnetic interference and physical attack capabilities are insufficient. As 19The White Paper on Exploration and Development Prospects of Embodied Intelligence in China the risk of AI data leakage is on the rise, differential privacy algorithms and hardware-level encryption technology should be combined, but the deployment cost is high. The software and hardware systems are fragmented ecologically. The algorithm framework PyTorch/TensorFlow has poor compatibility with the chip instruction set, and the developed tool chain is inconsistent. The essence o f the compe ti tion in embodied in telligence chips is the "scenario-defined hardware" capability. In the short term, autonomous driving and robotics are the main battlefields; in the long term, breakthroughs should be made in architectural innovation and ecosystem integration. Enterprises need to build differentiated advantages in the fields of automotive-grade chips, RISC-V ecosystem and Chiplet standards, while paying attention to the monopoly risks formed by large manufacturers through software and hardware coupling. 2. Sensors: the Basis of Perception Interaction Embodied intelligence sensors are key devices for realizing environmental perception and interaction of embodied intelligence systems. By integrating multimodal sensor data, such as vision, touch, force, and hearing, robots are endowed with human-like perception and decision-making capabilities. Vision sensors are mainly 2D/3D cameras, LiDAR, and mmWave radar. 3D vision sensors are a key focus of development given the need for high-precision spatial perception. As for tactile and force sensors, flexible tactile sensors have made breakthroughs in grasping control. For example, Tesla's Optimus Gen 2 can hold eggs without breaking them through a flexible tactile system. As for multimodal fusion, visual, tactile, and force data fusion technologies have become the mainstream. For example, the RoboFusion system combines reinforcement learning with multimodal data to improve industrial assembly accuracy. The global market size of vision sensors of embodied intelligence robots was approximately 198 million USD in 2024 and is expected to reach 1.1 billion U.S. dollars in 2031, with a compound annual growth rate of 28.7% from 2025 to 2031. In the global sensor market, some international giants such as Honeywell, ADI, TDK, 20The White Paper on Exploration and Development Prospects of Embodied Intelligence in China Bosch, Keyence and OMRON Corporation, have long dominated the high-end market of high-performance sensors, such as six-dimensional torque sensors and inertial sensors, and industrial automation sensors with their deep technical accumulation and extensive market influence. A number o f Chinese companies wi th core competitiveness in technology strengths and scenario-based applications have gradually emerged and occupied a certain share in the market. For example, Xindong Lianke, W-Ibeda, and MEMSensing Microelectronics have achieved good achievements in high-performance inertial sensors, and AKUSENSE has performed outstandingly in optoelectronic sensors. Silicon (Si) serves as a sensor substrate, and China is the world's second-largest supplier of silicon wafers. High borosilicate glass and alumina ceramics are used for temperature and pressure sensor substrates, and China leads the world in their production. The first challenge facing the sensor industry is the bottleneck of materials and processes. Flexible sensors rely on nanomaterials and microfabrication technology, which are costly and low yield rates. For example, the price of carbon fiber is 10-15 times that of glass fiber. The manufacturing of six-dimensional force sensors requires precision strain gauge mounting technology. At present, the equipment accuracy of Chinese companies is insufficient, resulting in poor product consistency. Second, standardization lags behind. Industry standards have not yet been unified, such as the lack of international specifications for tactile sensor sensitivity indicators, which restricts large-scale applications. The essence of the competition in embodied intelligence sensors is the integration of "multimodal perception + real-time decision-making" capabilities. In the short term, humanoid robots and medical scenarios are the main battlefields; in the l o n g t e rm, b r e a kt h r o u g h s s h o u l d b e m a d e i n m at e r i a l i n n o v at i o n a n d algorithm-hardware collaboration. Chinese companies need to create advantages in sub-sectors such as flexible touch and six-dimensional force control. In the future, sensors will not only be data collection tools, but also the core carriers of "human-like" agents. 21The White Paper on Exploration and Development Prospects of Embodied Intelligence in China 3. Motor: Core of Execution Control For embodied intelligence, the main role of the motor is to act as an actuator, converting the instructions of the intelligent system into specific physical actions and realizing the interaction between the machine and the environment. At present, the following progress has been made. First, high-precision and intelligent control. By optimizing the performance of drivers and encoders, servo motors have gradually achieved nanometer-level positioning accuracy and microsecond response speed, supporting embodied intelligence devices to complete complex actions. Second, breakthroughs in frameless motor and direct drive technology. The use of frameless motor design and direct drive technology reduces the transmission structure and improves energy conversion efficiency, with some products having an efficiency of more than 90%, while reducing mechanical losses. Third, lightweight and high power density. Carbon fiber materials, topology optimization design and other applications have reduced the motor size by 20%-30%, and increased the power density to more than 1.5kW/kg, meeting the needs of humanoid robots for compact power units. From the perspective of market structure, international giants dominate the high-end market. Japanese companies such as Yaskawa and Fanuc account for more than 60% of the global servo motor share. Chinese manufacturers such as Inovance Technology and Estun have a market share of more than 30% in the mid-to-low-end market and have made breakthroughs in sub-sectors such as frameless motors and integrated joints. As cross-border competition intensifies, technology companies such as Tesla and Xiaomi have impacted the traditional supply chain pattern by integrating their self-developed motors with AI algorithms. For example, the cost of Optimus robot joints has dropped by 40%. China controls the key raw materials for key servo motors. The amount of neodymium iron boron (NdFeB)-the core material of servo motors-required per humanoid robot is about 3.5 kilograms, which is 1.75 times the amount used in electric vehicles. China contributes 94% of the world's magnet materials production. The first challenge faced by the motor industry is that it lacks adaptability to 22The White Paper on Exploration and Development Prospects of Embodied Intelligence in China complex scenarios. Dynamic load changes pose challenges to the instantaneous torque response of the motor, which is difficult for the current technologies to fully meet. Second, there is a contradiction between cost and performance. High-end servo systems rely on imported components, resulting in costs accounting for 30%-40% of the entire machine, which restricts large-scale commercial applications. The essence o f the compe ti tion in embodied in telligence mo to rs is a three-dimensional game of "power density + control accuracy + cost control". In the short term, industrial and medical scenarios are the main battlefields; in the long term, breakthroughs should be made in material innovation and algorithm-hardware collaboration. With policy support and supply chain integration capabilities, Chinese companies are expected to establish differentiated advantages in sub-sectors, such as hollow cup motors and wheel hub drives. In the future, motors will not only be power units, but also core carriers for the execution of "human-like" actions by agents. (II) Midstream Software Development and System Integration: the Melting Pot of Emerging Intelligent Technologies As the technical core o f the indus trial chain, the mids tream link is the "technology melting pot" that drives the emergence of intelligent technology. This link integrates hardware components and intelligent functions to build a complete technical ecosystem consisting of four major technical centers: AI algorithm (intelligent decision-making center), operating system (resource scheduling center), cloud computing (computing power supply center) and middleware (system integration center). 1. Large Models: the "Brain" of Embodied Intelligence The embodied intelligence industry is undergoing a technological revolution driven by large AI models. Its core is to promote the evolution of robots from "tools" to "agents" through breakthroughs in multimodal fusion, autonomous decision-making and human-machine interaction capabilities, and to achieve 23The White Paper on Exploration and Development Prospects of Embodied Intelligence in China full-chain breakthroughs in technology from perception to action. The large model of the robot base will actively compete for the dominance of the robot "brain", which may accelerate the emergence of the next "Chatgpt moment". The algorithm scheme of embodied intelligence can be divided into two routes: hierarchical decision model and end-to-end model. At present, the mainstream is to use hierarchical large models for implementing scenario applications. The hierarchical model is based on the three-level architecture of "brain-cerebellum-limb". With its technological readiness and hardware adaptability, it occupies a dominant position in industrial manufacturing and service robots. The future development moves toward end-to-end large models, which can directly realize the transition from human instructions to robotic arm execution, that is, input image and text instructions and then output gripper end movements. Relying on breakthroughs in large models, a new paradigm is being built in consumption and services. The development pattern of embodied intelligence's large models presents US-led innovation in the general algorithm and end-to-end architecture, such as Google's RT series and Figure AI's Helix with a dual system model. China focuses on scenario adaptation and lightweight localization solutions, such as AgiBot's GO-1 open source platform. The two sides differ in technology paths, that is "the United States focuses on cloud-based large models, while China focuses on edge low costs". However, the underlying architecture still relies on international basic models, and China is accelerating the implementation of vertical fields with its supply chain advantages. The development of embodied intelligence-specific large models faces multiple challenges. At the technical level, there are still bottlenecks in the deep collaboration between multimodal perception and motion control. For example, real-time decision-making efficiency is insufficient in dynamic environments, and the migration gap between simulation environments and the real physical world leads to a policy failure rate of up to 30%; at the algorithmic level, there is a contradiction between the generality of large models and the specificity of embodied tasks, easiness for policy generation to deviate from human preferences, weak cross-domain generalization 24The White Paper on Exploration and Development Prospects of Embodied Intelligence in China ability, and difficulties in adapting to sudden changes in environmental parameters; Whether the robot architecture chooses a hierarchical model or an end-to-end model is essentially a trade-off between "explainability-generalization" and "real-time-reliability". The robot architecture will not move towards an either-or substitution, but will undergo a spiral evolution of "layered foundation building → end-to-end boundary pushing → dimensionality upgrading from multiple aspects". In the short term (2025-2027), hierarchical architecture will still dominate industrial scenarios, accounting for more than 70%, but the penetration rate of end-to-end models in high-frequency scenarios, such as home services and logistics sorting, will increase to 25%. In the medium term (2028-2030), the end-to-end model will break through the hardware bottleneck through neuromorphic chips, quantum sensing and other technologies, and the proportion of its industrial applications will exceed 40%. In the long-term run, the world model will be combined with reinforcement learning, promoting end-to-end architecture to replace hierarchical models in complex tasks. Chinese companies can build differentiated advantages with policy support and vertical scenario innovation (such as special robots and the open source ecosystem). In the future, large AI models will not only be codes and parameters, but also serve as the core driving force for "human-like" cognition and actions of agents. 2. Data Collection and Training: Embodied Intelligence's Accelerator It has become an industry consensus to accelerate the construction of large-scale, high-quality, multi-scenario data sets by creating a training ground for embodied intelligence data. At present, the United States and China are accelerating the release of various open-source data sets on humanoid robots to promote the training and optimization of large models of embodied intelligence "brain" and "cerebellum" of humanoid robots and advance the development of the humanoid robot industry. Regarding the global competition of the humanoid robot training ground, "China attaches great importance to the implementation of large-scale scenarios, while the United States excels in algorithms and cross-platform versatility". China has accelerated the integration of industry, academia and research through 25The White Paper on Exploration and Development Prospects of Embodied Intelligence in China policy support. The Shanghai Zhangjiang High-tech Park Heterogeneous Humanoid Robot Training Facility is China's first large-scale training base for heterogeneous humanoid robots. More than 100 heterogeneous machines are deployed, and each collects more than 500 trajectory data per day. It plans to build the embodied intelligence corpus with 50 million word entries by 2025. The National and Local Co-built Embodied Artificial Intelligence Robotics Innovation Center established in Beijing has built a closed loop of "data collection-model training-real machine verification" to promote application piloting in industrial scenarios. And foreign giants continue to take a lead by relying on their first-mover advantage in technology. Tesla's Optimus training ground reuses the scenario of the electric vehicle production line to collect industrial data such as battery packaging and handling. Google RT-X training ground integrates 22 types of robot data to build the Open X-Embodiment data set, supports zero-shot transfer to new tasks, standardizes data set production across platforms, and promotes general robot basic models. Data acquisition and simulation technology face serious bottlenecks. First, the data cost is high. The cost of real environment data collection is 100 times that of simulation, and heterogeneous robots have great difficulties in data standardization due to cross-brand adaptation and motion accuracy. Second, there is a gap between simulation and reality. The friction, material deformation and other physical properties in the virtual world can hardly fully match those in the reality, resulting in the failure rate of strategic transfer of over 30%. At present, the embodied intelligence training ground has shifted from initial single data collection to the 2.0 stage of heterogeneous collaboration, virtual-real integration, and ecosystem co-building. In the future, it is necessary to overcome the bottleneck of d at a a n d s im u l at i o n t e c h n o l o g y, st r e n gt h e n t h e d e e p c o o r d i n at i o n of hardware-algorithm-scenario, and accelerate the commercialization process through policy guidance. 3. Operating System: System Integration Hub As the core hub connecting physical entities and digital intelligence, the 26The White Paper on Exploration and Development Prospects of Embodied Intelligence in China embodied intelligence operating system is undergoing a profound transformation from instrumental to intelligent. The Robot Operating System (ROS) is currently the mainstream solution. The ROS is a quasi-operating system that mainly provides a series of tools and components for robot development. It is usually adapted to the underlying real-time operating system or Linux system. Since the ROS's source code is completely open, anyone can freely use, modify and distribute it. Therefore, it has quickly gathered global robot developers and technology enthusiasts and become the mainstream solution at this stage. China has launched an out-of-the-box embodied intelligence operating system for intelligent robots, the Embodied Operating System (EOS), which has recently been open sourced. EOS is different from the well-known foreign robot framework ROS. Rather than limited to the functional implementation at the middleware level, it is designed and implemented with a full consideration from the operating system kernel to the middleware, as well as the embodied intelligence and spatial intelligence algorithm library, supporting real-time kernels. The operating system is the strategic commanding height of robot development, and is expected to usher in a "counterattack" in the era of humanoid robots. Overseas companies continue to increase their investment in robot operating systems. Japan has proposed the national strategy to develop a robot operating system, and under the guidance of the advanced technology department, it will form an OpenRobot platform. The United States gave birth to Microsoft's development platform ROBOTICS, PlayerStage and the most famous ROS. The open source ecosystem ROS developed by Stanford University provides a series of tool libraries running on Linux systems and has quickly brought together major developers around the world. The essence of the competition in embodied intelligence operating systems is a three-dimensional game of "real-time-ecology-cost". In the short term, industrial and medical scenarios are the key areas for systems to make breakthroughs; in the long term, it is necessary to break through the bottleneck of hardware-algorithm collaboration and build an open source ecosystem. Chinese companies are forming a differentiated path with policy support and vertical scenario innovation, but we need to 27The White Paper on Exploration and Development Prospects of Embodied Intelligence in China pay attention to the ecosystem monopoly that international giants may form through soft and hard coupling solutions such as CUDA+Jetson. In the future, the operating system will not only be a control platform, but also the core infrastructure for "human-like" cognition and action of agents. 4. Cloud Service: Computing Resource Platform The cloud service industry in China is playing a key role in supporting the technology iteration and scenario implementation in embodied intelligence. The elastic supply of computing resources has been realized. Embodied intelligence cloud services, relying on the GPU clusters offered by leading cloud vendors, provide distributed training and inference acceleration capabilities. Huawei Cloud ModelArts supports Ascend chips to accelerate reasoning and meet the needs of training the large model as humanoid robot's "brain" on hundreds of billions of parameters. Simulation and data form a closed loop. The cloud-based high-fidelity simulation platform can generate more than one million trajectory data, shortening one algorithm iteration to minutes. Tencent Cloud improves the collection of heterogeneous robot data, achieving 500+ real trajectory data per day generated by a single machine. Regarding algorithm-hardware collaborative optimization, cloud service providers develop in-depth cooperation with hardware manufacturers through opening up their interfaces. For example, Tencent Cloud cooperates with AgiBot to optimize large model serving on the end side and reduce hardware energy consumption by 30%; Alibaba Cloud joins hands with UBTech to develop edge lightweight models, extending the robot's battery life to more than 4 hours. Leading cloud vendors accelerate industry integration through technology alliances. Huawei Cloud releases the Atlas robot development kit to strengthen end-cloud collaboration capabilities. Tencent Cloud, together with DOBOT and AgiBot, has built a "brain-cerebellum-limb" hierarchical architecture covering industrial control, home services and other scenarios; led by Alibaba Cloud, a "Community with Collaborative Development Plan for Embodied Intelligence" has been established, fueling the collaborative rise of intelligent manufacturing industry chain and carrier 28The White Paper on Exploration and Development Prospects of Embodied Intelligence in China cloud in the Beijing-Tianjin-Hebei region. Operator cloud is on the rise. eSurfing Cloud relies on 5G private networks to build a "robot as a service" (RaaS) platform. In the first half of 2024, the market share of government cloud infrastructure topped the ranking, providing a low-latency network guarantee for embodied intelligence. The development of embodied intelligence cloud services features "technologyecosystem-scenario". With the arrival of 2025 when humanoid robots are starting to enter mass production, cloud services will become the core hub connecting AI large models with the physical world, accelerating the deep penetration of embodied intelligence in intelligent manufacturing, medical care and other fields. (III) Main Downstream Products and Application Scenarios: Scenario Revolution of AI Adoption The downstream industry of embodied intelligence, centered on humanoid robots and service robots, has gained high-speed development. Humanoid robots are the major trend of future development and have broad prospects for applications in industrial manufacturing, service industry and housekeeping. Non-humanoid embodied intelligence products has been commercialized first. Delivery robots have been mainly applied in three scenarios: people's livelihood services, industrial logistics and special scenarios. 1. Humanoid Robots: the Best Form of Embodied Intelligence Humanoid robots show promising potential for applications in many fields, and particularly, the applications in industrial manufacturing, service industry and domestic services have been piloted for realizing commercialization. In industrial manufacturing, humanoid robots strong generalization ability. For example, AgiBot can participate in automobile chassis assembly and appearance inspection, even replace human labor in high-risk environments. In the service industry, humanoid robots are competent for tasks such as customer service and cooking. For example, Astribot S1 can cook meals and play musical instruments. In terms of domestic services, facing 29The White Paper on Exploration and Development Prospects of Embodied Intelligence in China the soaring demand for elderly care in the future, it can do housework. For example, MobileALOHA, apart from cooking, cleaning, and organizing, can also provide emotional support and companionship for the elderly. There is huge potential for development. Although many humanoid robots, such as Tesla Optimus and Boston Dynamics Atlas, have made technological breakthroughs, the industry has not yet achieved applications at a large scale due to constraints posed by difficulties in technology, cost and commercialization. Technically, the development of humanoid robots faces multiple challenges. It needs to replicate the mechanical structure of the human body. It should maintain strength while ensuring small and lightweight skeleton joints. Before achieving flexible grasping movements, we need to precisely design more than 20 joints of the hand. At the same time, developing a drive device that combines power and precision to ensure endurance and integrating sensors for precise perception in complex environments are both urgent technical difficulties to be addressed. In terms of cost, high-precision components drive a price up. Joints, as core components, account for about 60%-70% of the total cost. More than 20 joints in a single hand correspond to the same number of motors, making a single robot quite expensive. In the process of commercialization, high prices are beyond the reach of ordinary consumers and hinder market expansion. In addition, safety and ethical issues also make it less acceptable to the public. For example, the risk of robots getting out of control or privacy leaks has become a key obstacle preventing the industry from scale-up. �. Non-humanoid Embodied Intelligent Products: Emerge Like Blossom of a Hundred Flowers Delivery robots are experiencing constantly widened application scenarios and fierce market competition, mainly from three types of players. The delivery robot is an intelligent terminal device based on multimodal perception, autonomous path planning and the high-precision execution system. Its application scenarios have broken through the boundaries of traditional logistics and extended mainly toward 30The White Paper on Exploration and Development Prospects of Embodied Intelligence in China three directions. First, in people's livelihood services, it extends to catering scenarios, hotel delivery, and community group buying; second, in industrial logistics scenarios, it includes warehousing sorting and production line transfer; third, breakthroughs have been made in special scenarios, such as medical supplies transportation, and post-disaster rescue. The industry presents a trend of co-operation among three core players. One is the ecosystem-based tech giant. Alibaba's "Xiaomanlv (Little Workaholic Donkey)" has built a closed loop of "warehouse-station-vehicle", with the daily delivery peak exceeding one million orders. X Business Group of JD Logistics has developed a six-axis robotic arm for sorting, increasing sorting efficiency by 400%. The second is technology companies in vertical areas. Geek+ has deployed more than 500 smart warehouses around the world, and the market share of its storage robots ranks first. Quicktron has launched the "QuickBin" system, with bin-to-person picking efficiency reaching 1,200 pieces/hour. The tour guide robot has multiple functions integrated, a wide range of applications, and has diverse market players. As a carrier representing the combination of AI and robotics technology, tour guide robots realize autonomous positioning, dynamic obstacle avoidance and personalized services by integrating multimodal perception, natural language interaction and intelligent path planning. Currently, they have been applied in many fields, such as tourism, education, medical care, and commercial exhibitions. In terms of market competition, there are mainly three types of players. Technology giants, such as Alibaba and Tencent, rely on their technological advantages and user resources to actively deploy innovative products; professional robot manufacturers, such as UBTech and iFLYTEK Corporation, rely on technology accumulation to create high-performance and high-reliability products; customized service providers focusing on specific scenarios such as museums and scenic spots occupy the market segments with targeted solutions. Forces from multiple sides are working together to accelerate the development of the guide robot market. As an innovative travel mode driven by autonomous driving technology, Robotaxi 31The White Paper on Exploration and Development Prospects of Embodied Intelligence in China has performed a pilot run of commercial operation with full autonomy, with its application scenarios continuing to expand. More than 10 Chinese cities, including Beijing, Shanghai and Wuhan, have opened pilot zones in commercialization. Among them, Wuhan takes a lead with 3,000-square-kilometer pilot areas, while the areas in Beijing and Shanghai also cover 600 and 912 square kilometers respectively. Currently, about 67% of users use this autonomous driving as a primary or auxiliary mode of commuting. Robo taxi marke t par ticipan ts include au tomakers, autonomous driving technology companies and mobile transportation platforms. Automakers, responsible for vehicle manufacture and capitalizing on their design and production capabilities, cooperate with autonomous driving companies to create compatible models, such as GAC Toyota and XPeng Motors. The latter uses NGP technology to conduct business and cooperate with mobile transportation platforms. Autonomous driving technology companies, such as Apollo Go and PonyAI, continue to output intelligent driving solutions; mobile transportation platforms, such as Didichuxing and Ruqi Mobility, rely on their user and operational advantages to get services delivered. 32The White Paper on Exploration and Development Prospects of Embodied Intelligence in China III. Typical Practice and Exploration of Embodied Intelligence in China In recent years, embodied intelligence has been infiltrating the core scenarios of human production and life at a speed far exceeding expectations. This process is accelerated by two key driving forces: one is the leapfrog breakthrough in underlying technologies, such as AI, IoT and robotics; the other is the feasibility of large-scale applications brought by decreased hardware costs and improved algorithm efficiency. From a technical perspective, the evolution of deep learning frameworks has significantly enhanced the real-time processing capabilities of embodied intelligence for multimodal data. For example, models based on the Transformer architecture enable robots to synchronously parse visual, voice and environmental feedback information, thereby achieving efficient operation of the "perception-decision-action" closed loop in dynamic scenarios. At the same time, the maturity of edge computing and 5G technology has further reduced response delays, providing a technical foundation for autonomous navigation and group collaboration of mobile intelligent agents such as unmanned vehicles and drones. From a cost perspective, the miniaturization and cost reduction of high-precision sensors and bionic joint actuators have turned embodied intelligence "luxury" to "affordable luxury", significantly shortening the product transformation cycle and providing an economic foundation for large-scale scenario applications. (I) Embodied Intelligence Applications Will Be First Realized in Industrial Scenarios 1. Industrial Scenario: Reconstructing the Paradigm of Human-machine Collaboration and Promoting Dynamic Production Optimization In the industrial field, embodied intelligence is gradually becoming a core force in promoting the upgrading of intelligent manufacturing by deeply integrating perception, 33The White Paper on Exploration and Development Prospects of Embodied Intelligence in China decision-making and execution capabilities. With the advancement of industrial automation and intelligence, enterprises have raised their expectations for robots' adaptability and flexibility in complex environments. Embodied intelligence utilizing real-time task decomposition and multimodal perception fusion technology enables robots to autonomously construct environmental topology maps, dynamically decompose process constraints, and generate optimal action sequences based on reinforcement learning (RL). For example, in the electronic component mounting scenario, a collaborative arm equipped with six-dimensional force feedback can identify material deformation online and automatically compensate for assembly errors through an online trajectory optimization algorithm. At present, the application of embodied intelligence in the industrial field has formed a progressive model of "single point breakthrough-production line collaboration-whole plant intelligence". Its development will go through three stages. In the early stage, it coexists with humans in the same p roduction envi ronment and focuses on solving the p roblem of human-machine partnership. The medium stage will witness more efficient and intelligent human-machine collaboration, but it still relies on human operation. In the long term, embodied intelligence will be able to complete tasks independently, with human participation in front-line production gradually decreasing. As this participation decreases to zero, "unmanned factories" will be formed. (1) Representative Scenarios and Outcomes In the intelligent inspection scenario, through multi-robot collaboration and distributed intelligent systems, embodied intelligence robots can optimize task allocation and path plannin. For instance, they can combine infrared night vision functions to monitor temperature and humidity in equipment rooms, identify personnel intrusion, and warn of gas or liquid leakage. At the same time, it supports multi-terminal remote management, and multiple robots work together to cover areas such as factory workshops. In flexible manufacturing scenarios, industrial robots equipped with embodied intelligence can continuously observe their surroundings and automatically update decisions and optimize actions during task execution. They can adapt to 34The White Paper on Exploration and Development Prospects of Embodied Intelligence in China different production line layouts and workstation requirements, enabling rapid adjustment and reorganization of production lines, making them particularly suitable for multi-variety, small-batch, fast-paced production models. This production model is not only highly adaptable, but also has higher production efficiency and manufacturing precision. In the precision assembly scenario, automobiles and electronic products with complex internal structures contain a large number of parts, while the embodied intelligence robot equipped with high-precision mechanical arms and dexterous hands can perform precise grasping, placement, assembly and other operations and also complete precision assembly tasks, such as circuit board welding, screen installation, and battery fixation. (2) Business Model For industries having a high requirement for system integration and stability, such as automobile manufacturing, 3C electronics, and other high-precision assembly industries, a hardware + software bundled sales model is adopted, where a one-stop solution is provided with intelligent robots, sensors, and other hardware as carriers and through integrating software systems, such as autonomous decision-making algorithms and digital twin platforms. For small and medium-sized enterprises or industries with changing demand, an on-demand subscription model is adopted. Customers pay for their usage time, task volume, or effect without bearing the cost of hardware purchase and maintenance, while manufacturers continue to optimize algorithms through the cloud and charge subscription fees. In response to the complex needs in multiple scenarios across many industries, we create an ecosystem platform model, where we build an open technology platform to connect hardware manufacturers, algorithm developers and end customers, and generate profits through standardized interface commissions or value-added services. (3) Future Development Trends In terms of technology, the evolution of embodied intelligence in the industrial field does not necessarily move toward pursuing large and comprehensive general models, 35The White Paper on Exploration and Development Prospects of Embodied Intelligence in China but specialized small models focusing on market segments are equally important.. The trend shows the collaborative evolution of both large and small models, which meet the actual needs of industrial scenarios and demonstrate unique advantages in cost, efficiency and professionalism. In terms of applications, even though embodied intelligence has some relatively mature applications in industrial manufacturing, it still requires time to adapt to large-scale industrial application scenarios. The reliability requirements of industrial manufacturing scenarios and the cost control of enterprises have become the primary issues hindering the widespread adoption of embodied intelligence. Embodied intelligence is first implemented in scenarios that do not require high accuracy but have primary requirements for flexibility/generalization, such as box handling, packaging, and quality inspection. In the future, embodied intelligence will gradually realize tasks that require accuracy and flexibility to a certain degree in industrial scenarios, such as equipment assembly testing and debugging. In terms of the ultimate form, humanoid robots are not the only ultimate physical form for embodied intelligence. In the future, embodied intelligence can empower diversified equipment (such as industrial machine tools) through the "one-brain multi-machine" model. For example, a single AI core can simultaneously dispatch drone swarms and ground robots to achieve efficient task collaboration. 2. Applications Cases of Implemented Industrial Practice Applications: (1) UBTech's Industrial Humanoid Robot Walker S1 UBTech has entered the Stage 2.0 training phase. In March 2025, based on BrainNet, a software architecture designed for humanoid robot groups, and the Internet of Humanoids (IoH), dozens of Walker S1 groups applied swarm intelligence technology to conduct collaborative training in Zeekr's 5G smart factory. In the factory, Walker S1 is deployed in multiple complex scenarios, including the general assembly workshop, SPS instrument area, quality inspection area and door assembly area, successfully 36The White Paper on Exploration and Development Prospects of Embodied Intelligence in China enabling multi-task collaborative operations, such as collaborative sorting, collaborative handling and precision assembly. In the collaborative sorting link, UBTech's industrial humanoid robot Walker S1 applies cross-field pure visual perception technology and intelligent hybrid decision-making technology. Through cross-field pure visual perception technology, dynamic targets can be continuously perceived and tracked across fields. Robot groups collaborate to build global maps and realize "group building and group sharing". At the same time, intelligent hybrid decision-making technology based on large multimodal reasoning models combines semantic VSLAM navigation with dexterous operation requirements to dynamically allocate tasks between the cloud brain and local brain, thereby realising end-to-cloud and cloud-combined group intelligent decision-making. In the collaborative handling scenario, humanoid robots face numerous difficulties in cooperatively handling large loads and large-sized workpieces, including uneven load distribution, complex motion trajectories and adapting to dynamic environment adaptation environments. UBTech has developed a joint planning control system utilising multi-machine collaborative control technology to enable multi-machine collaboration in trajectory planning, load identification and compliance control, ensuring that the robot can dynamically adjust its posture and strength during handling, significantly improving the stability of handling large-sized and heavy-load workpieces. In precision assembly scenarios, Walker S1 demonstrates its superior dexterity in manipulating soft objects. Regarding small-sized soft film objects prone to deformation, Walker S1 dynamically adjusts the gripping force and posture through high-precision perception and adaptive control technology, ensuring that the film object remains undamaged or free from deviations during assembly. This technology showcases the flexibility and reliability of UBTech's humanoid robots, featuring a tactile five-fingered dexterous hand in complex industrial scenarios. Figures 7 & 8 Industrial Humanoid Robot Walker S1 Working Scenario in Zeekr's 5G Smart Factory (Source: UBTech) 37The White Paper on Exploration and Development Prospects of Embodied Intelligence in China (2) China Unicom Humanoid Inspection Robot China Unicom has developed a core capability system, including ubiquitous connection and multi-network standard intelligent adaptation, robot multi-protocol adaptation, multimodal perception data fusion analysis and robot equipment intelligent upgrade with end-cloud training and promotion collaboration, as well as robot agent building and the open sharing of generality. It closely combines 5G edge private networks, large industry models, small general models and embodied intelligence to launch humanoid inspection robots to promote the transformation of enterprises from labour-intensive to intelligence-driven. Cu r rently, bipedal humanoid robots a re p rima rily responsible fo r handling abnormalities in IPQC tasks. The system classifies alarm levels for abnormal events that may occur during the inspection and automatically summons the bipedal humanoid robot to perform tasks related to correcting abnormal events, such as monitoring 5S standard-related abnormalities, conducting operation safety-related inspections, and intervening in abnormal event error correction actions. In the future, training of humanoid robot operation models and bionic arms, as well as dexterous hand design will be utilised to enhance bipedal mobility, support multi-station rapid response in narrow spaces, and facilitate precise material handling, flexible loading and unloading, and precision assembly and debugging scenarios. Figure 9 Working Scenario of Humanoid Patrol Robot (Source: China Unicom) (3) AlphaBot, Universal Intelligent Robot Launched by AI² Robotics In the fi r st half of 2025, Zhejiang Geene r Mic roelect ronic s Co., Ltd. and Shenzhen-based AI² Robotics concluded a strategic cooperation agreement. The two 38The White Paper on Exploration and Development Prospects of Embodied Intelligence in China sides will jointly develop solutions for universal, embodied intelligence robots in p recision manufactu ring, based on the p rocess requi rements of high -end semiconductor manufacturing scenarios. The universal embodied intelligence robot AlphaBot, a new generation of intelligent terminals defined by the large-scale embodied model Alpha Brain, independently developed by AI2 Robotics, possesses excellent spatial intelligence and powerful learning ability. AlphaBot can complete multiple tasks without training, quickly master new tasks and stably adapt to various changes. It is now applied in semiconductor factories. By continuously learning from multidimensional data in semiconductor production scenarios, AlphaBot will gradually automate high-precision process links, such as wafer loading, intelligent consumable replacement, and precision part sorting. As more experience is gained, robotic systems can continue to improve like humans. The universal embodied intelligence robot solution provided by AI2 Robotics solves the automation challenges of the semiconductor industry with an innovative model. Its universal and flexible characteristics provide a new idea for the large-scale implementation of "unmanned factories". Figure 10 Application of Universal Embodied Intelligence Robot AlphaBot in Loading and Unloading Work Scenarios (Source: AI2 Robotics) Exploration of Other Products and Their Applications in Industrial Scenarios: (4) AgiBot's Flexible Manufacturing Robot Yuanzheng A2-W The AgiBot Yuanzheng A2-W flexible intelligent manufacturing robot is designed for flexible intelligent manufacturing scenarios. It has multiple capabilities, such as grabbing, placing, handling, and plugging. It has self-developed embodied intelligence 39The White Paper on Exploration and Development Prospects of Embodied Intelligence in China algorithms, open interfaces, and provides tools to facilitate secondary development. Yuanzheng A2-W, featuring a bionic seven-degree-of-freedom dual-arm, supports parallel and asynchronous operations, can relay and change hands in the air, and can easily cope with complex workpieces and specific posture operations. Equipped with a four-wheel drive system, it supports zero turning radius and crab walking. The waist can elevate and tilt, offering 22 degrees of freedom to cover the human working space fully. It features an integrated design of the chassis, arm, and perception system, supporting installation, debugging, and deployment within a single day. Based on UniGrasp, Uni6DPose, UniPlug, and other embodied atomic capabilities, as well as 275T computing power, it can achieve millisecond-level real-time object recognition, pose estimation, and operation decision-making, dynamically adapting to flexible working environments. What's more, through 3D model synthesis training and reinforcement learning, the replacement of operational objects can be shortened to hours at the fastest. The robot features a multi-mode perception function that integrates multiple sensors, including a 360 ° Lidar, four AI vision sensors, and two six-dimensional force sensors, enabling millisecond-level perception and intelligent obstacle avoidance. The arm collision detection system, along with 360 ° real-time environmental sensing and redundant perception and control design, effectively ensures the safety of both personnel and the environment. Expedition A2-W can be applied to loading and unloading, terminal plug-in, logistics transfer and other scenarios, enhancing the automation level of production lines and facilitating flexible production. Figure 11 Product Parameters of AgiBot's Flexible Manufacturing Robot Yuanzheng A2-W (Source: AgiBot) 40The White Paper on Exploration and Development Prospects of Embodied Intelligence in China (II) Embodied Intelligence Explores Potential Scenarios and Applications in the Service Industry 1. Healthcare: Connecting Health Data Throughout the Entire Cycle to Realize an Active Care Service Ecosystem Embodied intelligence is reshaping the capability boundaries of medical AI by building a "perception-cognition-action" closed-loop system. It provides AI with human-like perception and execution capabilities through the deep integration of multimodal perception, action control, decision planning and memory abilities, enabling it to adapt to complex medical environments. The application of embodied intelligence has made inevitable progress in market segments, such as clinical intervention, nursing companionship, and R&D. For example, in the field of R&D, AI algorithms are used to accelerate data analysis and experimental automation, providing power for medical breakthroughs. Even though embodied intelligence has achieved certain milestones in the medical field, its overall development is still in its infancy and is facing multiple challenges. For example, embodied intelligence development is usually carried out on a simulation platform, which often cannot exactly replicate the real-world environment. In addition, embodied intelligence systems rely on large datasets, but privacy regulations and complex clinical workflows make it difficult to access ethical real-world datasets. In the future, embodied intelligence will need to promote smart medical care to new heights through continuous technological innovation and interdisciplinary collaboration. (1) Representative Scenarios and Outcomes In clinical intervention scenarios, embodied intelligence has been widely utilised throughout the entire cycle of clinical intervention, including pre-intervention, intervention and post-intervention stages. Before intervention, the embodied intelligence system supports the medical team in developing personalised treatment plans through accurate diagnosis and evaluation. For example, intelligent image analysis helps doctors detect potential diseases (such as tumours and stones) early 41The White Paper on Exploration and Development Prospects of Embodied Intelligence in China through the automated interpretation of medical images, thereby improving the accuracy and efficiency of diagnosis. During the intervention, embodied intelligence can improve the accuracy and safety of surgery. For example, in minimally invasive surgery, embodied intelligence can accurately locate the operating field and perform the surgical procedure, effectively reducing the error rate. After the intervention, embodied intelligence systems, such as intelligent rehabilitation robots, can help patients recover their motor functions through personalized rehabilitation training, particularly for those with stroke and spinal cord injuries. The robot can dynamically adjust the training content and intensity according to the patient's recovery situation to ensure the effectiveness and safety of rehabilitation training. In the nursing and companionship scenario, embodied intelligence provides patients with comprehensive life support and emotional companionship. For example, social assistance robots improve the social skills of autistic children through interaction; feeding robots assist patients in completing tasks such as eating and dressing through perception and adaptive technology, and provide patients with real-time feedback on and encouragement for rehabilitation exercises; exoskeleton devices provide gait training for spinal cord injury patients; AI wheelchairs support users with limited mobility to move independently through intelligent navigation. (2) Business Model Embodied intelligence in the medical field mainly includes two business models: technology empowerment and data value-added. Unde r the technology empowerment model, in addition to directly purchasing embodied intelligence hardware, a subscription system and effect-based sharing model can be adopted, charging according to the amount of work completed by embodied intelligence or the specific medical effects achieved. In the data value-added model, establishing a data trading platform enables the direct sale of selling medical professional data or the provision of corresponding data products, such as insulin response curves, thereby generating profits. 42The White Paper on Exploration and Development Prospects of Embodied Intelligence in China (3) Future Development Trends In the future, embodied intelligence will realize cell-level diagnosis and treatment in the medical field. Through the R&D of computer vision, new magnetically controlled materials, and image tracking equipment for micro-magnetic objects, as well as the development of visual micro-nano manipulation technology and targeted drug delivery technology in vivo of magnetically controlled micro-nano swimming robots, nanorobots can complete targeted treatment work within cells, but also achieve operations, such as th rombus removal, d rug delive ry, and cance r t reatment. Additionally, brain-computer interface technology will be utilised in medicine and rehabilitation, particularly for restoring patients' mobility. Brain-computer interfaces can help patients with strokes or spinal cord injuries reconnect with the outside world, improve their quality of life and independence, and also show innovative potential in precision treatment and monitoring, early disease diagnosis, and promoting the recovery process. 2. Logistics and Transportation: Reconstructing the Autonomous Navigation System to Achieve Efficient Cargo Transportation In logistics and transportation, embodied intelligence is expected to reduce transportation costs and become a key factor in forming an efficient, fast and intelligent logistics system. Embodied intelligence enables logistics robots to possess stronger autonomous decision-making and learning capabilities, allowing them to adapt to more complex and diverse tasks. The current application of embodied intelligence in the logistics field primarily includes picking robots, forklift robots, handling robots, and material handling robots. The empowerment of embodied intelligence technology enhances work efficiency and management levels in warehousing, loading and unloading, handling, sorting, packaging, distribution, and other related processes. For example, Amazon has recently been testing the humanoid bipedal robot Digit, developed by Agility Robotics and invested in by Amazon, in its warehouse operations. Digit comprehensively completes tasks, mainly including unloading trucks, moving boxes, and managing shelves, which greatly improves the efficiency of warehouse operations. 43The White Paper on Exploration and Development Prospects of Embodied Intelligence in China (1) Representative Scenarios and Outcomes In the flexible sorting scenario, traditional automation technology or industrial robot technology face certain challenges in case of unstructured scenarios. However, embodied intelligence builds dynamic maps using LiDAR, camera, and SLAM technologies, which will significantly enhance its environmental perception, route planning, and motion navigation capabilities. In this way, it can better adapt to variable environments, identify multiple targets, adjust paths autonomously, and avoid obstacles in time, thereby empowering sorting, handling, and other tasks. In the unmanned transportation scenario, through unmanned trucks, drones, unmanned delivery vehicles, and other carriers, the entire supply chain from trunk line transportation to terminal distribution is unmanned. Through 24-hour uninterrupted operation, it improves logistics distribution efficiency and reduces corporate operating costs. In special logistics scenarios, embodied intelligence enables the material transportation in extremely cold and high-temperature environments by utilising cold-resistant and high-temperature-resistant materials and multimodal positioning. Through explosion-proof design and radiation shielding technology, its applications are implemented in hazardous scenarios, such as the autonomous handling of radioactive materials in nuclear power plants. (2) Business Model Embodied intelligence is reshaping the core competitiveness of the logistics industry by combining physical carriers with intelligent algorithms. From warehousing to distribution, its application scenarios are constantly expanding, while the business model has also shifted from single equipment sales to data-driven integrated services. The business model of embodied intelligence in the field of logistics and transportation can be summarized into the following four categories: One is the combination of hardware sales and subscription services. The company earns income by selling smart devices, such as unmanned forklifts and AGV/AMR, while providing subscription services, including software scheduling systems and maintenance services. In this way, a sustainable profit model has formed. 44The White Paper on Exploration and Development Prospects of Embodied Intelligence in China For example, Mooe Robot and Hangcha Intelligence have achieved a compound annual revenue growth rate of more than 30% through hardware sales and software service subscriptions. The second is all-in-one solution integration. Logistics system integrators (such as Shenzhen New Trend International Logis-Tech Co., Ltd (NTI) and OMH SCIENCE Group Co., Ltd) provide a full-process solution of "equipment + software + operation" and charge fees by service items. A case in point is the AGV system deployed at the BMW Leipzig plant in Germany, which reduces downtime waiting time by 70% by seamlessly connecting production and logistics links. The third is platform resource sharing. Based on the IoT technology, social transportation capacity and warehousing resources are integrated to build a sharing platform that pays on demand (such as Huolala and Yunmanman), reducing the fixed investment of small and medium-sized enterprises. JD Logistics utilises the "Atomic Skill Library" framework to break down tasks into reusable modules, thereby improving resource utilization efficiency and reducing data requirements. The fourth is data-driven value-added services. Through the collection and analysis of logistics full-link data, high-value-added services such as supply chain optimization and demand forecasting are provided. For example, Amazon Prime Air’ s drone delivery system optimizes route planning based on real-time data analysis, reducing the cost of single-piece deliveries by 40%. In the future, the business model will further develop towards ecosystem collaboration. Through cross-regional joint efforts (for example, Beijing technology research and d e v e l o pm e nt c o l l a b o r at i n g w it h G u a n g d o n g ’ s i n d u st r i a l c h a i n ) a n d "manufacturing-logistics-retail" data interoperability, an end-to-end intelligent supply chain system will be built to promote industry cost reduction, efficiency improvement, and green transformation. (3) Future Development Trends In the future, the core development trend of embodied intelligence in the field of logistics and transportation is as follows: 45The White Paper on Exploration and Development Prospects of Embodied Intelligence in China The first is the deep integration of technology and the construction of a three-dimensional network. AI, 5G, IoT, and digital twin technologies will be further integrated to enhance the real-time perception and autonomous decision-making capabilities of logistics systems. For example, based on the reasoning capabilities of large models such as DeepSeek-R1, logistics companies can optimize path planning and warehouse management to achieve dynamic adjustments in their supply chains. At the same time, drones, eVTOL (electric vertical take-off and landing aircraft) and autonomous vehicles will form a "space-air-ground" three-dimensional logistics network, breaking through terrain restrictions and shortening delivery time. For example, JD Logistics has piloted drone delivery for the last mile in both urban and remote areas, improving time efficiency by more than 30%. The second is greening and sustainable development. New energy equipment (such as hydrogen forklifts and electric unmanned vehicles) and intelligent scheduling algorithms will be popularized at an accelerated pace to reduce carbon emissions. Government policies support intelligent connected new energy vehicles and low-altitude economy, contributing to achieving the goal of carbon neutrality in the logistics industry. The third is globalization and ecosystem synergy. Cross-border logistics optimizes timeliness through multimodal transport and the establishment of overseas warehouses. For example, Cainiao collaborated with Emirates Airlines to deliver services within 2 hours in the Dubai Free Trade Zone. Leading enterprises will rely on technology output to expand their global market share, while promoting data interoperability across the "manufacturing-logistics-retail" industrial chain to form an end-to-end intelligent supply chain ecosystem. The fourth is flexible and professional services. Driven by consumer demand, customized logistics services (such as cold chain temperature control and high-safety delivery) will become the focus of competition. The combination of large models and agent technology will reshape software forms to support personalized solutions. Additionally, blockchain technology enhances the traceability of goods, improving the transparency and security of the supply chain. 46The White Paper on Exploration and Development Prospects of Embodied Intelligence in China The fifth is policy and infrastructure support. The government promotes the development of a unified and open transportation market, encouraging the use of rail and water transportation, as well as smart logistics technology to reduce transportation costs. The aim is to further reduce the proportion of logistics costs in GDP from 14.06% in 2024. In summary, embodied intelligence will promote the logistics industry's acceleration towards efficiency, greenness and globalization, and technological iteration and ecosystem synergy will become key driving forces. 3. Family Services: Reshaping the Environmental Perception Paradigm and Realizing Full-scenario Smart Housekeeping In the field of family services, embodied intelligence enables the delivery of real, customized services through advanced cognitive and action capabilities. The development of home service robots has evolved from basic robot vacuum cleaners to multifunctional robots that can now perform floor cleaning, item handling, and basic household tasks. For example, 1X and OpenAI have been collaborating to develop an embodied intelligence humanoid robot EVE that can achieve cognitive understanding of the human daily work environment, learn, correct, and collect data during interactions with the environment, and complete autonomous home and office helper tasks. Embodied intelligence makes home service robots real human-friendly intelligent assistants, providing intelligent and personalized full-scenario family services. (1) Representative Scenarios and Outcomes In the housework execution scenario, embodied intelligence enables long-distance operation in a large space through a mobile base and simulates human hands to complete tasks such as preparing dishes, stir-frying, serving from a pot, washing clothes, playing with cats, and watering flowers. For example, the AgiBot Juechen C5, launched by AgiBot, integrates multiple cleaning capabilities, including sweeping, washing, and dust removal, and can autonomously complete tasks such as charging, 47The White Paper on Exploration and Development Prospects of Embodied Intelligence in China adding drainage, and cleaning sewage tanks in complex environments with minimal human intervention. In the child companionship scenario, embodied intelligence corrects children's writing posture through gesture recognition to enhance their learning efficiency. By adopting UWB+Bluetooth AoA positioning technology (accuracy ±5cm), it enables electronic fence cross-border alarm functionality. In home security scenarios, embodied intelligence can scan doors and windows and identify abnormal opening and closing through computer vision. It can monitor smoke and water immersion, and upon detection, quickly cut off gas and water sources to reduce the probability of leakage accidents. Robot dogs patrolling at night can identify intruders through thermal imaging and activate sound and light deterrence to prevent illegal intrusion. (2) Business Model The business models of embodied intelligence in home services field primarily include the following. The first is the combination of hardware sales and subscription services, which aims to achieve continuous profitability by selling intelligent robot hardware (such as home service robots and elderly care companion robots), along with data subscription or function upgrade services (such as health monitoring and emotional interaction modules). The second is ecosystem cooperation between B2B and B2C, such as cooperating with home appliance manufacturers and smart home platforms to incorporate embodied intelligence robots into the home IoT ecosystem and provide scenario-based solutions. The third is the integration of software and hardware in vertical fields, developing customized robots for specific family needs (such as elderly care and childcare), and enhancing user stickiness through the deep integration of hardware and data services. Fourth, it involves technology authorization and API services, providing other companies with core algorithms or modules of embodied intelligence (such as perception and navigation technologies), thereby lowering the industry entry threshold and expanding the scope of technology application. In addition, some companies obtain policy support and market resources through projects in cooperation with the government (such as smart 48The White Paper on Exploration and Development Prospects of Embodied Intelligence in China elderly care pilots). (3) Future Development Trends In the future, universal embodied intelligent robots will be able to perceive anthropomorphically, utilise traditional tools, and perform tasks autonomously in uncertain environments, thereby realizing full-scenario home assistants. They will enter thousands of households like cars and become indispensable life partners and helpers for every family, assisting with tasks such as rehabilitation and household chores. First, technological breakthroughs drive the expansion of scenarios. The integration of multimodal large models and continuous learning technology will enhance the adaptability of robots in complex family environments, promoting an upgrade from single tasks (such as cleaning and handling) to comprehensive services (including health management, education, and companionship). Second, it will experience cost reduction and large-scale applications. With the maturation of the supply chain and breakthroughs in mass production technology, the price of home robots is expected to drop significantly, accelerating their popularization among consumers. Third, policies and ethical norms will be improved. The government will strengthen the formulation of security standards (such as data privacy and human-computer interaction ethics), while promoting the intelligent transformation of family scenarios through industrial policies (such as "AI +" actions). Fourth, it moves toward ecosystem synergy and cross-border integration. Home service robots will be further integrated into smart home, medical health and other ecosystems to form a full-chain value network of "hardware + service + data". Ultimately, embodied intelligence may become the core of smart terminals in family life, redefining the future form of human-machine collaboration. 4. Others: From Empowerment to Change, Build a Cognitive Evolutionary System of Human-machine Collaboration In addition to the above areas, embodied intelligence applications in scientific 49The White Paper on Exploration and Development Prospects of Embodied Intelligence in China research, emergency response, and other fields will also bring profound changes. In the field of scientific research, embodied intelligence can autonomously conduct scientific experiments and operate continuously for extended periods, thereby accelerating the scientific research process. In extreme environments, such as deep sea and outer space, embodied intelligence robots can replace humans to explore and discover unknown scientific mysteries. At the same time, they can also take on high-risk or tedious tasks to ensure personnel safety and improve work efficiency. Additionally, embodied intelligence robots can efficiently collect and analyze data to provide valuable information for scientific researchers. Regarding emergency response, embodied intelligence has become a key technology for ensuring personnel safety and optimizing work processes by performing high-risk tasks that humans cannot undertake. In search and rescue scenarios, embodied intelligence robots can be used to search for missing persons, deliver medical supplies and perform rescue missions to reduce casualties. In explosive ordnance disposal and mine clearance scenarios, embodied intelligence robots can carry explosive detectors to detect and handle explosives. Apart from that, they can also perform mine clearance operations in minefields to avoid casualties. 50The White Paper on Exploration and Development Prospects of Embodied Intelligence in China IV. Exploration of New Paths for the Future Development of Embodied Intelligence (I) Build a Solid Hardware Foundation and Promote the Interconnection of Embodied Intelligence Standards Among BRICS Countries At the hardware level, durability, battery energy efficiency and the need for deep integration with software are the main challenges to the development of embodied intelligence at this stage. In the future, regarding the development of embodied intelligence hardware, BRICS countries should consider not only breakthroughs in key technologies and cost-effectiveness, but also such factors as hardware versatility and standard consistency of embodied intelligence. During the development process, the specific operational levels mainly include the following: First, the BRICS countries should encourage building a unified standard for the use of embodied intelligence hardware, share data sets universally, enhance the versatility and reliability of hardware operation and maintenance, and build a collaboration network of embodied intelligence for BRICS countries. In the process of developing embodied intelligence, there are problems with technical differences and standard fragmentation. In terms of hardware configuration, there are various types of robotic arm designs; in terms of joint modules, the performance is uneven due to inconsistent friction damping and back-driving force; in the field of sensors, there is a lack of unified specifications for performance parameter setting and data interface design, and the selection of multimodal sensors is fragmented; in terms of end effectors, there are differences such as two-finger hand clamping and three-finger gripper. To eliminate these differences, it is recommended that the BRICS countries establish a unified standard system. Specifically, by formulating standardized interface protocol specifications and modular architectures, the difficulty of hardware maintenance is reduced and the service life of equipment is extended. By 51The White Paper on Exploration and Development Prospects of Embodied Intelligence in China building modular hardware parameter benchmarks, it improves interoperability between hardware. It should unify the communication protocol and human-computer interaction specifications of teleoperation systems and use standardized frameworks to reduce cross-platform collaboration costs, promote the universal sharing of data sets from industry manufacturers. It also promotes the formation of an international collaborative network covering the entire process of technology R&D, production and manufacturing, and operation and maintenance services among BRICS countries. Second, the BRICS countries should take technology autonomy and industrial collabo ra tion as thei r co re s t ra tegies, suppo r t ups t ream componen t manufacturers, focus on breakthroughs in underlying hardware technology and generalized large-scale applications. Through the BRICS Institute of Future Network and other platforms, a cross-border cooperation mechanism will be established to support and guide generalization of upstream component manufacturers of embodied intelligence. They should "survive" first, ensure stable and reliable component delivery capabilities and effectively collaborate with embodied intelligence to complete machine factories, while achieving key technology breakthroughs in multiple fields. In the field of chips, scenarios define hardware capabilities, with the focus on key scenarios for autonomous driving and humanoid robots. The existing computing power foundation of all countries should be integrated. With the combination of RISC-V open source architecture and Chiplet technology, the key task is to break through the energy efficiency bottleneck of heterogeneous computing architecture, and reduce the end-to-end delay of scenario applications through collaborative optimization of algorithms and hardware. In the field of sensors, with a focus on segmented applications such as flexible touch and six-dimensional force control, it should support the batch preparation process R&D of flexible tactile sensor nanomaterials, and establish open protocol standards for multimodal perception data fusion. In the field of motors, we will give full play to the advantages of supply chain integration, promote the effective three-dimensional integration of "power density + control accuracy + cost control", and establish leading advantages in segmented fields such as hollow cup motors and wheel hub drives. 52The White Paper on Exploration and Development Prospects of Embodied Intelligence in China (II) Strengthen Perception Ability and Promote BRICS Embodied Intelligence Fusion Innovation At present, the embodied intelligence development of BRICS countries faces certain challenges in achieving breakthroughs in core technology, data resource integration and implementation of large-scale applications. On the one hand, breakthroughs are still yet achieved in the full-modal perception capability of embodied intelligence systems; on the other hand, the high cost of high-quality data collection in dynamic interactive scenarios and the significant difference between simulation and real data restrict the improvement of model generalization capabilities. In addition, the fragmentation of robot operating systems further hinders the formation of scaling effect. To this end, in the future, a full-stack solution covering technology R&D, data supply and industrial implementation will be formed from the three dimensions of technical collaboration, data closed loop and platform ecosystem through deep linkage between industry, academia and research, building of virtual-real integration training grounds and building of general development networking platforms. The specific path includes: The first is to build a deep collaborative ecosystem of industry, academia and research, breaking through the bottlenecks of embodied intelligence full-modal perception and end-cloud collaboration technology. First, we should improve the full-modal perception capability and achieve breakthroughs in end-to-end large models. At present, the embodied intelligence large model still has insufficient technology reserves in all modalities of vision, touch, hearing and smell, especially in force feedback and tactile feedback. The current modeling ability for complex perception is still weak. In the future, it is necessary to focus on achieving b reakth roughs in end -to -end la rge model capabilities and build a unified perception-decision-control framework through multimodal data fusion technology. Neural symbolic systems are introduced to enhance the semantic alignment of multimodal information and self-supervised learning is used to reduce dependence on high-quality labeled data. Secondly, in response to end-side computing power 53The White Paper on Exploration and Development Prospects of Embodied Intelligence in China limitations, we will promote the building of end-side intelligent computing platforms, conduct in-depth technical research, and solve the communication delay problem in end-cloud collaboration. Limited by the bottleneck of end-side computing power, it is difficult to compress cabinet computing power to terminal equipment in the foreseeable future. Therefore, "end-cloud collaborative computing" has become an inevitable choice. The cloud is responsible for complex model training and global knowledge updates, while the end side focuses on perception and control tasks with high real-time requirements. In terms of decision-making and control technology, we will make up for the shortcomings in motion coordination control capabilities, strengthen simulation training based on physics engines and small sample reinforcement learning technologies, and improve the generalization of control modules. At the forward-looking scientific research level, BRICS countries can promote the industry and universities to jointly deepen basic research, develop a low-latency ROS control framework, and solve the communication delay problem in end-cloud collaboration; and in terms of project implementation, promote cooperation among manufacturers of end-side computing chips and communication chips to improve the end-side computing power of software and hardware collaboration. The second is to build a virtual-real fusion training ground system for BRICS countries and promote the deep integration of embodied intelligence simulation and real data. Data is a key barrier to breakthroughs in embodied intelligence capabilities in the future. The data of embodied intelligence applications involves complex interactions in dynamic environments, and regarding data, we face such challenges as high costs, insufficient data quality and diversity, large differences between simulated data and real data, and data training biases in data collection and training. In response to these bottlenecks, there are three major ways to break through in the future. First, the BRICS countries can build an international joint training ground system that integrates virtuality and reality. Through the physical training grounds that are internationally linked in many BRICS countries and regions to cover the main scenarios, combined with digital twin technology to build a virtual training ground for software and hardware collaboration, all parties in the industry work 54The White Paper on Exploration and Development Prospects of Embodied Intelligence in China together to build a high-quality multimodal corpus, achieving a dual improvement in data collection efficiency and scenario coverage. Secondly, through industry alliances and cross-border cooperation, we will jointly build high-quality, large-scale embodied intelligence data sets to promote the improvement of multimodal data closed loops. Open source data sets reduce R&D costs, accelerate technology iterations, and attract more developers to participate. Key breakthroughs include the collection of end-to-end data in key scenarios, such as two-arm collaboration, tactile feedback, and force control, combined with self-supervision technologies such as comparative learning (SimCLR) and mask autoencoders (MAE) to reduce the current dependence on high-quality labeled data. Finally, simulation data is integrated with real data to improve the quality and proportion of simulation data in applications. To improve the authenticity and diversity of the simulation environment, high-density and controllable training data is generated through a large-scale simulation platform, with transfer learning used to bridge the "reality gap". Synthetic data generation technology based on physics engines is developed, and the complementary iteration of simulation data and real data gradually realized. We should give full play to the advantages of simulation data, integrate simulation data into training, let the end-to-end model learn the optimal time and path scheme, and avoid the error of non-optimal data acquisition path of real machine. The third is to build a platform ecosystem of universal development and network connection for BRICS embodied intelligence, breaking through the bottleneck of operating system fragmentation and multi-machine collaborative scheduling. The future development of the embodied intelligence operating system platform will revolve around generalization, collaboration and scaling. At present, the fragmentation problem of robot operating systems is prominent. The mainstream solutions in the market, such as ROS or Linux-based customized systems, rely on a large number of open source components, resulting in compatibility conflicts and poor scale-up effects. Therefore, the development of general embodied development platforms and robot networking platforms will be the key. First, the BRICS countries can encourage leading enterprises and scientific research institutions to establish a universal open platform for 55The White Paper on Exploration and Development Prospects of Embodied Intelligence in China embodied intelligence and build a technology open source community. Through a unified central control system, multi-robot collaborative scheduling is realized, and the dual-core mechanism of large model-driven task planning and data-driven end-to-end skill execution is integrated to provide standardized interactive interfaces and development frameworks for robots of different brands and forms. We should solve the compatibility problem of heterogeneous systems and support cross-scenario task integration through modular function libraries. For example, the shared development template for industrial assembly and home service scenarios provides a technical foundation for industry reuse. Secondly, it should promote the building of a networking platform for embodied intelligence devices by telecommunications and cloud vendors in BRICS countries to release the potential for large-scale applications of embodied intelligence. We should promote the connection, management and operation and maintenance of hundreds of embodied intelligent devices, support human-machine collaboration in various scenarios and business systems, create full life cycle management for embodied intelligence, support remote operation and maintenance, health monitoring and OTA upgrades, and lower the threshold for equipment use. (III) Focus on Key Scenarios to Drive the Sound Development of BRICS Embodied Intelligence As a key technology for the deep integration of AI and the physical world, embodied intelligence can not only improve manufacturing efficiency, but also reshape the indust rial fo rm. The BRICS count ries can use thei r unique advantages in manufacturing, scientific research and other aspects to systematically promote the large-scale application of embodied intelligence technology in key industries, such as a ut om ob i l e m a n uf a ct u r i ng, t h r o ug h d iff e r e nt i at ed d e v e l opm e nt r o ut e s, multi-dimensional technology integration, collaborative innovation mechanisms and industrial fund support, build a complete industrial ecosystem from technology R&D to large-scale applications, form an internationally competitive embodied intelligence 56The White Paper on Exploration and Development Prospects of Embodied Intelligence in China development pattern, and create new momentum leading the global intelligent manufacturing transformation. First, the BRICS countries can take a differentiated development route in the application of embodied intelligence industry, and implement it in stages in multiple forms and scenarios. In terms of form, at the early stage, the focus is on quadrupedal and wheeled robots, and in the later stage, the focus is on full-size humanoid robots; in terms of application scenarios, classified policies are implemented to guide the scenario, and following "small applications, large scale-up", promoting applications gets started from highly structured scenarios. In the selection of key industries and business scenarios, it follows the principle of "gradual progress, from easy to difficult". First, in terms of key industries, it will be introduced from the automotive manufacturing industry and gradually penetrate into the semiconductor manufacturing and pharmaceutical industries. In the long term, it is expected to be applied to commercial and home service environments, focusing on breakthroughs in eight scenarios such as assembly, inspection, cutting, sorting, handling, unpacking, cropping, and testing. In the manufacturing industry, the task execution and process of embodied intelligence devices are highly regula r and can be t rained to efficientl y complete repetiti ve ta s k s. The process-oriented characteristics of the manufacturing industry give it a first-mover advantage in the application of embodied intelligence technology, such as urgent demand for flexible production and high degree of structured working environment. Industrial embodied intelligence robots can effectively improve the flexibility and adaptability of factory production tasks, realize autonomous learning during operation, and continuously enhance their ability to perform complex tasks and operational accuracy. Secondly, in te rms o f key scena rios, the ea rly la rge -scale implemen ta tion focuses on the final assembly link in the au tomo tive manufacturing industry. According to front-line research, large-scale applications are carried out in loading and unloading, anodizing, and battery port plugging and unplugging. Combined with existing mature technologies, force/torque control is used to repair inadequate situations. In the future, with the maturity of technology, we will 57The White Paper on Exploration and Development Prospects of Embodied Intelligence in China implement the mature applications of dual-arm coordination in manufacturing. For different process links, the key focus is different, and the rhythm and manufacturing speed are improved in a targeted manner. The second is the application of embodied intelligence in manufacturing, the core of which lies in multi-dimensional technology integration and full-link closed-loop connection. First, relying on the collaborative innovation of the industrial chain, it opens up the entire link of "technology-data-industry" and forms a virtuous development cycle of "flywheel effect". The BRICS countries have promoted the transformation of core technological achievements such as AI chips, multimodal perception, and dexterous operation into large-scale manufacturing through policy support. Secondly, at the technical level, it is necessary to focus on integrating capabilities such as autonomous navigation, foot control, upper body movement planning, force control and tactile perception to form a technology stack for closed-loop control and environmental interaction. Autonomous navigation needs to be combined with dynamic environmental perception and path planning algorithms, enabling robots to achieve flexible obstacle avoidance and accurate positioning in complex industrial scenarios. It relies on multimodal perception technology. In the future, it requires the coordination of vision, LiDAR and tactile feedback and breakthroughs in high-order intelligent driving algorithms. Force control and tactile perception are the key links to achieve refined operation. The flexible pressure sensor array can obtain the three-dimensional distribution data of contact force in real time. Combined with the decision-making ability of multimodal large models, it enables robots to accurately adjust the operating strength when assembling precision parts or performing collaborative tasks. The third is to integrate the scientific research resources of BRICS countries and establish an industry fund for embodied intelligence. First, we should build a cooperation mechanism for collaborative innovation in embodied intelligence among BRICS countries. Relying on the cooperation framework of the BRICS Business Council, we should promote the establishment of a cooperation platform for embodied intelligence, integrate scientific research institutions, enterprises and 58The White Paper on Exploration and Development Prospects of Embodied Intelligence in China government resources, and carry out cross-border joint research. Through regular technical seminars and achievement report meetings, we will promote knowledge sharing and standard consistency in core technology fields, such as embodied perception, human-computer interaction, and autonomous decision-making, and focus on breakthroughs in common technical bottlenecks such as multimodal data fusion and environmental adaptation. At the same time, we will deepen BRICS digital education cooperation, implement the "Joint Training Program for Embodied Intelligence Talents", set up cross-border laboratory rotations, technology-themed competitions and industry-university-research joint mentor programs to reserve versatile talents for development of various fields. Secondly, a joint investment fund for BRICS embodied intelligence industry is established. Drawing on the experience of building an ecosystem for industrial cooperation among BRICS countries, an investment fund for embodied intelligence will be established with joint funding from member countries' government guidance funds, social capital and international financial institutions. The fund focuses on supporting robot autonomous learning, multimodal large models, multimodal data collection and training, embodied general development platforms, robot networking platforms and other fields, promoting the fo rmation of a full -c ycle suppo rt chain of "technolog y re sea rch and development-scenario verification-commercial promotion", accelerating the application of embodied intelligence in the field of intelligent manufacturing, and enhancing the competitiveness of BRICS countries in the field of embodied intelligence. 59The White Paper on Exploration and Development Prospects of Embodied Intelligence in China There is an old Chinese saying, "When we work together, hundreds of rivers will return to the sea; when we pool our wisdom, ten thousand trees will become a forest." This is exactly what we believe when we face challenges. China is willing to work hand in hand with BRICS countries to achieve breakthroughs in both the depth and breadth of Embodied intelligence applications in consumer and industrial sectors. We will continue to promote the construction of an open world economy and build a community with a shared future for mankind, making the world more prosperous and people's lives happier. 60 |
|
|
|
