CAICT and Tsinghua University Jointly Release the "Embodied Intelligence Development Report (2026)": The Industry Is at a Transition Point Toward Pilot Applications

During the 2026 Artificial Intelligence Industry Conference held on October 14–16, 2026, the Institute of Artificial Intelligence of the China Academy of Information and Communications Technology (CAICT), together with the Department of Electronic Engineering of Tsinghua University and the Embodied Intelligence Testing Laboratory, jointly released the "Embodied Intelligence Development Report (2026)" (Chinese version).

It is reported that this is the third time CAICT has released an embodied intelligence research report. The report is grounded in the critical window during which embodied intelligence is leaping from laboratory validation to large-scale application, focuses on the new changes and new challenges in industrial development over the past year, and offers an outlook on the future.

Over the past two years, embodied intelligence has been written into the Government Work Report for consecutive years, becoming one of the future industries cultivated as a national priority. Over the past year, technological and industrial iteration and achievements in embodied intelligence have advanced rapidly, yet a polarization between hot and cold has also emerged: on the one hand, capital market and policy attention continues to heat up, with new achievements in models, bodies, data and other links released intensively; on the other hand, large-scale deployment and a closed commercial loop still face multiple constraints, and a clear gap remains between technology demonstrations and real delivery. From the "laboratory" to the "training ground" and then to the "real-world venue," embodied intelligence in 2026 stands at the transitional node from a future industry to pilot applications.

Industry assessment: beneath the boom, hot and cold intertwine, and the contradiction between front-loaded expectations and lagging capabilities becomes apparent

In 2026, China's embodied intelligence industry reveals a chill amid continued warming. On the hot side, policy support, capital investment and product supply are all heating up across the board. The 15th Five-Year Plan sets the tone, and the real-world training special action proposes to push robots from "demonstration performances" to a normalized "work mode" by the end of the year, with more than 10,000 devices deployed; according to IT Juzi statistics, domestic financing in the first half of 2026 was about 93.5 billion yuan, roughly five times the same period last year, and the number of companies valued at over 10 billion yuan increased to 20; the 2026 World Robot Conference released 311 new products, more than double year-on-year, and global humanoid robot shipments in the first half of the year exceeded 22,000 units. On the cold side, technological cold, application cold and enterprise cold coexist: closed-source frontier models leave the capabilities available to the industry about two generations behind the frontier, and data utility limits the ceiling of model capabilities; shipment does not equal deployment, with a considerable proportion of devices powered on fewer than 10 days per month on average, low-value orders, and the price of the complete machine accounting for only 40%–60% of full life-cycle cost; daily rental prices for performance leasing have fallen by 80%–90%, the residual value of second-hand engineering prototypes has shrunk by nearly 90%, and most enterprises see revenue growth without profit growth. The intertwining of hot and cold is a concentrated manifestation of the contradiction between front-loaded market expectations and lagging industrial capabilities. The industry is in the expectation-inflation phase, and over the next two to three years the first round of industry consolidation will be completed through market competition.

Technology system: the four elements of models, data, bodies and computing power evolve in coordination, building a deployable and iterable system loop

Current technology routes have not yet taken shape, and the industry is seeking capability breakthroughs along the coordinated development of four key elements: models, data, bodies and computing power.

In terms of models, technology routes are far from converging, and the industry is exploring three directions around the goal of breaking through models' shortcomings in physical cognition and action. First, continuing to explore the model Scaling Law under data-driven approaches; second, advancing breakthroughs in cognition and action capabilities along multiple parallel routes, including continuing the VLA route to drive iteration of general behavior policies and specialized deployment iteration, integrating world models to strengthen physical cognition capabilities, and extending end-to-end policies to the underlying locomotion of whole-body mobile manipulation; third, exploring the agent paradigm, coordinating system access, task orchestration, data collection, model training and evaluation, as well as operational diagnosis and skill accumulation.

In terms of data, simulation synthesis and real data proceed in parallel on two lines to fill the gap. Embodied intelligence data is moving from scattered collection toward systematic supply; more than 70 training grounds have been built and put into use nationwide, data production is accelerating the construction of engineering platforms, a quality evaluation system has taken initial shape, and exploration of circulation, trading and compliance has begun.

In terms of bodies, the performance of core components such as joint drives, dexterous hands, tactile sensing, edge-side computing and batteries is accelerating in iteration; complete-machine configurations such as bipedal humanoids, wheeled arms, quadruped and wheel-legged robots are continuously being refined; and complete-machine manufacturing is shifting from component assembly to product definition.

In terms of computing power, three routes are evolving in parallel: unified scheduling of heterogeneous body resources, device-cloud split inference and communication-computing co-design, and edge-side energy-efficient inference with integrated computing and control. The system support capability of cloud-edge-device collaboration is continuously being improved, jointly driving embodied intelligence from single-point capability breakthroughs toward system deployment and continuous evolution.

Applications and business: shifting from demonstration performances to work modes, with industrial chain division of labor refining but a sustainable commercial loop still to be established

Embodied intelligence applications are deepening from display experiences and single-point validation toward actual operations and large-scale applications. Entertainment performances and competitive events continue to expand public attention, toys and companion products are the first to enter the personal consumer market, and pilot programs in manufacturing, logistics, energy, healthcare, commercial services and other fields are gradually expanding.

The division of labor along the industrial chain has become further defined, forming a development landscape with upstream data, computing power, basic software and hardware, and core components; midstream embodied intelligence products and technical services; and downstream industry applications. According to statistics from the Embodied Intelligence Industry Map of the Artificial Intelligence Industry Alliance (AIIA), as of October 8, 2026, there were 405 relevant domestic enterprises, including 180 upstream, 257 midstream, and 212 downstream, with 47.2% of enterprises covering two or more industrial chain segments simultaneously. The engineering chain for embodied intelligence from complete machine shipment to actual deployment is relatively long, giving rise to a group of secondary development service providers that rely on existing bodies and models to complete application delivery, becoming a prominent feature of the increasingly refined division of labor along the industrial chain. Business models are accelerating their differentiation along each segment: the data segment is the first to close the loop; models monetize indirectly through complete machine premiums, subscription services, and ecosystem lock-in; body sales and leasing proceed in parallel; scenario solutions are in the early stage of deployment; and the aftermarket, including robot 4S/6S/7S/9S stores and talent training, is beginning to take shape. Overall, most segments remain in an exploratory stage where "growth has been established, but profitability has not," and a sustainable commercial loop has yet to form.

Evaluation systems: Four types of benchmarks calibrate the technical level and application boundaries, while evaluation still faces three challenges

Based on statistics from approximately 800 publicly published embodied intelligence model papers in 2026 and the public benchmarks they used, existing benchmarks can be divided into four categories: embodied cognition, which mainly uses static question-and-answer evaluation to assess cognitive reasoning abilities such as spatial perception, physical common sense, memory, and planning; embodied navigation, which mainly uses dynamic simulation evaluation to assess exploration and mobility capabilities; embodied manipulation, which uses simulation or real-machine evaluation to assess manipulation and collaborative operation capabilities, with real-machine evaluation having developed distinctive modes such as unified exams, diagnostics, open interviews, and vocational skills tests; and embodied safety, which mainly uses simulation evaluation and is expanding from semantic safety evaluation to closed-loop execution and physical consequence assessment. Overall, controlled desktop manipulation benchmarks are becoming saturated, but model performance under environmental perturbations generally falls below 30%, and fine manipulation, long-horizon multi-step tasks, and open semantic interaction remain prominent shortcomings, with a clear gap between simulation and real-machine evaluation results. Embodied intelligence evaluation remains an open measurement science problem and faces three challenges. Evaluation realism, reproducibility, and low-cost input are difficult to reconcile; test variables are numerous and cannot be exhaustively enumerated, making it difficult for limited samples to form stable conclusions about models with similar performance; and existing benchmarks are misaligned with productivity indicators of concern to industry, such as continuous operation capability, task takt time, and operating costs.

CAICT Institute of Artificial Intelligence jointly releases the Embodied Intelligence Development Report (2026)

Over the past two years, embodied intelligence has been continuously written into government work reports, becoming one of the future industries prioritized for cultivation at the national level. Over the past year, embodied intelligence technology and industry have iterated and produced achievements at a rapid pace, but have also shown a polarized pattern of hot and cold: on the one hand, capital market and policy attention continues to heat up, with new achievements in models, bodies, data, and other segments released intensively; on the other hand, large-scale deployment and commercial loops still face multiple constraints, and a clear gap remains between technical demonstrations and real delivery. From the "laboratory" to the "training ground" and then to the "real-world venue," embodied intelligence in 2026 stands at the transition point from a future industry to pilot applications. During the 2026 Artificial Intelligence Industry Conference of the Artificial Intelligence Industry Alliance, held from October 14 to 16, 2026, the Institute of Artificial Intelligence of the China Academy of Information and Communications Technology (CAICT), together with the Department of Electronic Engineering of Tsinghua University and the Embodied Intelligence Testing Laboratory, jointly released the Embodied Intelligence Development Report (2026) (Chinese version). The English version of the Embodied Intelligence Development Report (2026) was also released at the seminar held alongside the third meeting of the International Telecommunication Union (ITU-T) Focus Group on Embodied AI Standards (FG-EAI) on October 16, 2026.

The report is positioned at the critical window in which embodied intelligence is leaping from laboratory validation to large-scale application. It focuses on the new changes and new challenges in industrial development over the past year and conducts systematic analysis across dimensions including clarifying the concepts of physical AI and embodied intelligence, cool-headed thinking amid the industrial boom, defining the four-element framework of embodied intelligence, analyzing the path evolution and business models from "performance" to "application," and progress in evaluation systems and international standards. Finally, it looks ahead to future development in four areas: reconstruction of the technical architecture, deepening of application scenarios, improvement of the evaluation system, and construction of safety and ethics.

Core viewpoints of the report

1. Conceptual connotation: Embodied intelligence and physical AI are different expressions of the same technological paradigm

Embodied intelligence (EAI) and physical artificial intelligence (PAI) are rapidly converging in core connotation, technological foundation, and object of action; in essence, they are different expressions of the same technological paradigm. First, both emphasize the core of "intelligence": both are intelligent technologies oriented toward the real physical world, with EAI emphasizing more "how intelligence is realized through a body" and PAI emphasizing more "how artificial intelligence enters and acts on the physical world." Second, they share a technological foundation, including multimodal foundation models, vision-language-action models (VLA), world models, robot learning, physical simulation, digital twins, synthetic data, and on-device intelligent computing. Third, robots are the ideal "test field," serving as the main carrier through which systems act in the physical world and bear the consequences in real time, and becoming a common validation ground for testing EAI and PAI. The development of embodied intelligence and physical AI is changing people's understanding of artificial general intelligence, and the physical world is becoming an important testing ground for examining whether artificial intelligence is truly moving toward general intelligence.

Source: CAICT

2. Industry assessment: Hot and cold intertwine beneath the boom, and the contradiction between front-loaded expectations and lagging capabilities emerges

In 2026, China's embodied intelligence industry is revealing a chill beneath its continued warming. On the hot side, policy support, capital investment, and product supply are all heating up across the board. The 15th Five-Year Plan has set the tone, and the real-scenario practical training special initiative proposes pushing robots from "demonstration performances" to a normalized "work mode" by the end of the year, with more than 10,000 devices deployed; according to IT Juzi statistics, domestic financing in the first half of 2026 was about 93.5 billion yuan, roughly five times the level of the same period last year, and the number of companies valued at over 10 billion yuan rose to 20; the 2026 World Robot Conference released 311 new products, more than double the year-on-year figure, and global humanoid robot shipments in the first half of the year exceeded 22,000 units. On the cold side, technological cold, application cold, and enterprise cold coexist: closed-source frontier models leave the capabilities available to the industry about two generations behind the frontier, and data utility limits the ceiling of model capabilities; shipment does not equal deployment, with a considerable proportion of devices powered on fewer than 10 days per month on average, low-value orders, and hardware prices accounting for only 40%–60% of full-lifecycle costs; daily rental prices for performance leasing have fallen by 80%–90%, the residual value of second-hand engineering prototypes has shrunk by nearly 90%, and most companies see revenue growth without profit growth. The interweaving of hot and cold is a concentrated manifestation of the contradiction between front-loaded market expectations and lagging industry capabilities; the industry is in the peak of inflated expectations, and the first round of industry consolidation will be completed through market competition over the next two to three years.

3. Technology system: The four elements of models, data, bodies, and computing power evolve collaboratively, building a deployable and iterable closed-loop system

Current technology routes have not yet been finalized, and the industry is seeking capability breakthroughs along the collaborative development of four key elements: models, data, bodies, and computing power. In terms of models, technology routes are far from converging, and the industry is exploring three directions around the goal of breaking through models' shortcomings in physical cognition and action. First, continuously exploring the model Scaling Law under data-driven conditions; second, advancing breakthroughs in cognition and action capabilities through multiple parallel routes, including continuing the VLA route to drive iteration of general behavior policies and specialized deployment iteration, integrating world models to strengthen physical cognition capabilities, and extending end-to-end policies to the underlying locomotion of whole-body mobile manipulation; third, exploring the agent paradigm, coordinating system access, task orchestration, data collection, model training and evaluation, as well as operational diagnostics and skill accumulation. In terms of data, simulation synthesis and real data proceed in parallel on two lines to fill the gap, and embodied intelligence data is moving from scattered collection toward systematic supply; more than 70 training grounds have been built and put into use nationwide, data production is accelerating the construction of engineering platforms, a quality evaluation system has taken initial shape, and exploration of circulation, trading, and compliance has begun. In terms of bodies, the performance of core components such as joint drives, dexterous hands, tactile sensing, edge computing, and batteries is accelerating in iteration; whole-machine configurations such as bipedal humanoids, wheeled arms, quadrupedal and wheel-legged robots are continuously being refined; and whole-machine manufacturing is shifting from component assembly to product definition. In terms of computing power, three routes are evolving in parallel: unified scheduling of heterogeneous body resources, device-cloud split inference and communication-computing co-design, and edge-side energy-efficient inference with integrated computation and control; the system support capability of cloud-edge-device collaboration is continuously being improved, jointly driving embodied intelligence from single-point capability breakthroughs toward system deployment and continuous evolution.

Source: CAICT

4. Applications and business: Shifting from demonstration performances to work mode, with industrial chain division of labor refined but a sustainable commercial closed loop still to be established

Source: CAICT

5. Evaluation system: Four types of benchmarks calibrate the technical level and application boundaries, and evaluation still faces three challenges

Based on statistics from about 800 publicly published embodied intelligence model papers in 2026 and the public benchmarks they use, existing benchmarks can be divided into four categories: embodied cognition, mainly static question-and-answer evaluation, examining cognitive reasoning abilities such as spatial perception, physical common sense, memory, and planning; embodied navigation, mainly dynamic simulation evaluation, examining exploration and mobility capabilities; embodied manipulation, using simulation or real-robot evaluation, examining manipulation and collaborative work capabilities, with real-robot evaluation having formed distinctive modes such as unified exams, diagnostics, open interviews, and vocational skills tests; and embodied safety, mainly simulation evaluation, expanding from semantic safety evaluation to closed-loop execution and physical consequence assessment. Overall, controlled tabletop manipulation benchmarks are becoming saturated, but model performance generally falls below 30% under environmental perturbations, and fine manipulation, long-horizon multi-step tasks, and open semantic interaction remain prominent shortcomings, with an obvious gap between simulation and real-robot evaluation results. Embodied intelligence evaluation remains an open measurement science problem and faces three challenges. Evaluation authenticity, reproducibility, and low-cost input are difficult to balance simultaneously; test variables are numerous and cannot be exhaustively enumerated, and limited samples make it difficult to form stable conclusions about models with similar performance; and existing benchmarks are misaligned with productive indicators of concern to industry, such as continuous operation capability, task tempo, and operating costs.

(1) The division of labor in the industrial chain is further clarified, and the role of system integration and services is enhanced

(2) Diverse entities are accelerating substantive deployment, and industrial collaboration is further deepened

(3) Multiple links in the industrial chain generate revenue, but a sustainable business model still needs to be established

(4) The construction of the industry's open-source ecosystem is accelerating, and an open collaborative innovation mechanism has initially taken shape

(5) Evaluation standard practices are continuously deepening, and systematic construction and application linkage are further deepened

V. Future outlook

(1) Technology trends: Model architectures continue to iterate, and integrated software and hardware jointly define intelligence

(2) Industry trends: Key scenarios lead the way, and infrastructure, industrial ecosystems, and employment systems are restructured in tandem

(3) Governance trends: The standards system moves from "building frameworks" to "implementation"

Head of the Domestic Counterpart Group for ITU-T Study Group 21, Secretary-General of the Artificial Intelligence Standardization Technical Committee of the Ministry of Industry and Information Technology, and Head of the Overall Group of the Artificial Intelligence Industry Alliance (AIIA). His research focuses on the field of artificial intelligence technology; he has published more than 20 papers, led the completion of more than 40 domestic and international standards, and participated in the drafting of a number of major national policy documents related to artificial intelligence.

Shi Lin, Senior Engineer and Director of the Embodied Intelligence and Robotics Department at the CAICT Institute of Artificial Intelligence

Mainly engaged in government support, technical research, standards development, testing and evaluation, and industrial ecosystem development in the field of artificial intelligence. Participated in the drafting of a number of international and industry standards in the field of artificial intelligence.

Zhang Weimin, Senior Engineer and Deputy Director of the Embodied Intelligence and Robotics Department at the CAICT Institute of Artificial Intelligence

Mainly engaged in artificial intelligence standardization and industry research, involving embodied intelligence, AIGC, benchmarking, and other areas. Led the construction of the EAI-bench embodied intelligence benchmarking system, published 10 papers and granted patents, and led the drafting of more than 10 international and domestic industry standards.

Bai Ruwen, Ph.D., Engineer in the Embodied Intelligence and Robotics Department at the CAICT Institute of Artificial Intelligence

Mainly engaged in research in the field of embodied intelligence; led the completion of more than 10 international/industry/association standards and technical specifications in the field of embodied intelligence, published more than 10 high-level papers in the AI field, and is an expert in the writing group of the monograph Future Industries: The Next Decade's Outlets and Breakthrough Paths.

Copyright statement: The copyright of this report belongs to the China Academy of Information and Communications Technology and the Department of Electronic Engineering of Tsinghua University, and is protected by law. When reprinting, excerpting, or using the text or views of this report in other ways, the source should be indicated as "Source: China Academy of Information and Communications Technology and the Department of Electronic Engineering of Tsinghua University, Embodied Intelligence Testing Laboratory." Violators of the above statement will be pursued by the editors for relevant legal liability.

Contact information for the writing team: Zhang Weimin, Institute of Artificial Intelligence, CAICT, zhangweimin@caict.ac.cn Report download

On this site's official account "Smart City Industry Analysis," enter the keyword "Embodied Intelligence Report" in the dialog box to obtain the full text of the report.

www.smartcity.team

Statement: Please indicate when reprinting that the source is "Smart City Industry Analysis (www.smartcity.team)" and the relevant author. Except for original content, some information on this site comes from the Internet. If the relevant content involves infringement of rights, please contact the contact information at the bottom for deletion.

[Digital Technology Special Research] Community

In-depth peer exchange, focused on smart city and digital technology industry research and material accumulation.

Continuously updating the latest industry white papers, industry research, policy developments, and implementation solutions to help industry practitioners efficiently accumulate knowledge and share experience.

Massive premium content continuously iterated, providing cutting-edge perspectives, connecting industry talent, carrying out case co-creation and resource exchange, and jointly building a high-quality industry exchange atmosphere.

#post-##

Recommended for you

National Public Data Operation Development Report (2025-2026): There are 76 public data resource authorized operation platforms or zones that have been publicly operated or are under construction nationwide, with 202 implementation institutions and 224 operating institutions identified

Embodied Data Development Report 2026: Detailed explanation of the development status, meaning and classification, industry chain and business models, distinctive cases, and development trends of the embodied data industry

Device-side Intelligent Industry Development Report (2026): Systematically reviews the overview, technology system, product forms, and application scenarios of the device-side intelligent industry

Green Token Factory Technical White Paper: Systematically elaborates the five-layer technical system of the green token factory: "energy infrastructure - computing infrastructure - token production - token operation - token circulation"

Guidelines for Building High-Quality Datasets for Agent Capability Construction (2026): Jointly released by CAICT, Lenovo, and the AIIA Data Committee, proposing the overall framework, key methods, and implementation paths for building industry high-quality datasets for agent capability construction

CAICT Planning Institute, Research Report on the Development of Intelligent Computing Infrastructure for Token Services (2026): Systematically reviews the paradigm reconstruction process of intelligent computing infrastructure from "computing factories" to "token factories"