Editor's Note: The embodied intelligent robot industry stands at a critical intersection of technological iteration and scaled deployment, accelerating its integration into various industries. The entire sector is swiftly transitioning from the "technology validation phase" to the "commercial popularization phase," with the outline of a new round of growth becoming increasingly clear. Meanwhile, more practical propositions are placed before every industry practitioner: where should technological breakthroughs be directed, where should the market be deeply cultivated, and how should the ecosystem be co-built... In light of this, on the occasion of the 2026 World Robot Conference, China Electronics News invited representatives from multiple robot companies to focus on core issues such as technological evolution, market deployment, ecological synergy, and future trends. By collecting voices from the front lines of the industry, we aim to preserve a vivid footnote for the next chapter of the robot industry.
Based on your company's position in the industrial chain and business layout, what do you consider to be the most prominent challenge facing the embodied intelligent robot industry in the transition from technology validation to scaled commercialization? Among technological maturity, cost control, scenario adaptation, and industrial ecosystem construction, which is the biggest challenge at present?
The four challenges are closely intertwined, and the most prominent systemic bottleneck is the "three mountains": first, the data closed loop. What the industry lacks is not data, but reusable physical interaction data. Manufacturers do not share data with each other, resulting in severe "data silos"; second, the world model. Most robots are still stuck at "recognition + preset actions," lacking physical reasoning capabilities; third, cost efficiency. Ultimately, it depends on whether the robot can help customers make money, reduce costs, and improve efficiency. However, these three mountains are layered: the commercial flywheel for closed, standardized scenarios has already been running smoothly. We should not deny the entire industry based on the ultimate challenge of general-purpose humanoid robots.
Chaoneng's solution is to offset the technological cycle through business model innovation: positioning itself as a one-stop complete machine customization and application scenario service provider, organizing the long-tail demands of various industries through 9S stores + online platforms. Since the opening of the world's first 9S store, it has received nearly 500 batches of customers and over 40,000 visitors, accumulated more than 140 business opportunities, and achieved a single-store sales volume of over 40 million RMB, verifying the value of this path.
We believe the most prominent challenge is the disconnect between technological maturity and scenario adaptation, with the core lying in bridging the closed loop of "hardware-perception-data-model-scenario."
The current industry suffers from structural shortcomings of "valuing demonstrations over deployment, and hardware over full-stack capabilities." Many enterprises only focus on a single dexterous hand or a single algorithm module. However, a robot being able to "see and walk" is merely the foundation; the ability to stably, precisely, and continuously manipulate objects is the key to embodied intelligence entering the industry, which requires full-stack capabilities of "hardware-perception-data-model-scenario." For example, in industrial precision assembly and high-risk power maintenance scenarios, robots require sub-millimeter operation accuracy, multi-modal tactile feedback, and reliability for long-term continuous operation. Behind this is the deep integration of hardware, data, models, and scenario engineering. Single-point breakthroughs are hard to solve systemic problems. We chose the full-stack in-house R&D route precisely to bridge the complete link from underlying hardware to scenario delivery, allowing technology to truly adapt to the stringent standards of the industry.
Zhang Dapeng, Assistant Vice President of Leju Robotics
Frankly speaking, these four challenges are not an "or" relationship, but an "and" relationship—they are intertwined and together constitute the systemic threshold for industrialization and deployment.
From Leju's practice, scenario adaptation is the most direct bottleneck at present. Industrial scenarios have stringent requirements for reliability, stability, and takt time efficiency. The engineering gap from "being able to run and jump" to "being able to work and work well" is much larger than the outside world imagines. Every specific scenario requires a massive amount of real-robot data training and repeated tuning.
Closely following is cost control. For humanoid robots to become "productivity tools," the cost of the whole machine needs to continuously drop. Through the country's first 10,000-unit-level humanoid robot automated production line, Leju is promoting mass production and cost reduction; meanwhile, the localization rate of Leju's Kuafu humanoid robot exceeds 95%. By joining hands with Huixi Intelligent and Weifan Intelligence to build a domestic edge computing power base, we are completing the "last piece of the puzzle" for the full-stack localization of embodied intelligence, achieving full-link autonomous control, and fundamentally building a cost advantage.
Industrial ecosystem construction is the accelerator. A single enterprise cannot cover all scenario demands. Relying on the OpenLET national-level embodied intelligence open-source dataset community, Leju has gathered more than 70 universities and industrial chain enterprises, and over 10,000 developers, building a complete open-source ecosystem of "data training-scenario pilot-achievement incubation," enabling an exponential increase in scenario deployment efficiency.
Moving from technological deployment to commercialization, the difficulties are actually multi-layered. The most intuitive short-term pain point is scenario adaptation. The effect in the lab is very good, but placing it on different factory production lines requires massive customized modifications, making the customer's payback period too long. Hardware cost and technological maturity are long-term hard constraints, but the most fundamental issue is the imperfect industrial ecosystem. Currently, hardware and software from different companies are incompatible, lacking unified standards, data cannot be interconnected, and upstream and downstream enterprises work in isolation. If the ecosystem problem is not solved, cost reduction and scaled deployment will continue to be hindered.
The most prominent challenge remains the scarcity of data and the capability of models. The primary factor currently restricting the commercial deployment of embodied intelligent robots lies in the intelligence side (i.e., models), which struggles to meet the requirements of high precision, high stability, and high robustness in unstructured scenarios. This is because the available data in the field of embodied intelligence is far less than the massive data accumulated on the internet over decades for large language models, making it impossible to train models that meet commercial deployment requirements. Therefore, the industry urgently needs a way to generate more data efficiently and at low cost as "fuel" for embodied model training, which is exactly the successful practice of Kuawei Intelligence over the years.
Veichi Electric has been deeply engaged in the field of electrical drive and industrial control for over 20 years. Relying on long-term accumulated underlying motion control technologies, it extends downward to layout the core robot components track, meeting the diverse needs of various robots in force control, perception, dynamic response, and long-term reliable operation.
In terms of scenario adaptation, we focus on three major directions: embodied intelligence, human-machine collaboration, and efficient mobility, to create scenario-based solutions. Different forms of robots have different demands for core components: humanoid robots have extremely high requirements for high precision and lightweight; collaborative robots emphasize torque precision, control capabilities, and safe interaction experience in complex environments; mobile robots focus more on low power consumption and highly integrated design. To this end, Veichi Electric conducts targeted optimizations for different application directions to ensure that every product can deliver optimal performance in real working conditions.
At present, I believe the more prominent challenge in the industry is industrial ecosystem construction, which requires the concerted efforts of the whole industry. By linking upstream and downstream partners, we can bridge the whole-chain coordination system of "raw materials—core components—complete machines—terminal scenarios," helping the entire industry achieve scaled, sustainable, and high-quality development.
Relying on the "Bionic Tensegrity Robot Theory and Technology," Moonbot is positioned around bionic humanoid robots, bionic dexterous hands, tactile perception, and related core components, standing at the intersection of key components and complete machine system synergy. From this industrial chain position, if I must choose a primary challenge among technological maturity, cost control, scenario adaptation, and industrial ecosystem, I think it is technological maturity; but what is meant here is not single-item performance in the lab, but the maturity of systems engineering under scenario constraints. A prototype being able to complete one action does not mean the product can run long-term. True scaled commercialization requires robots to simultaneously meet requirements such as task success rate, continuous operation capability, anomaly recovery, batch consistency, maintenance convenience, safety, and unit task cost. Cost control is the result of engineering maturity, scenario adaptation is the way to test technological maturity, and the industrial ecosystem determines whether mature solutions can be quickly replicated. Therefore, the core gap the whole industry needs to cross is to transform "demonstrable technology" into "products that can be stably delivered, continuously maintained, and create revenue for customers."
We believe these challenges are intertwined and coexist. Specifically, at the technological level, there is still room for optimization in the edge-side adaptation of general embodied large models and the integration of motion control and perception decision-making, and the supply of mature standardized computing power and algorithm solutions is insufficient; at the cost level, the costs of core computing power, sensors, actuators, and other components remain high, restricting large-scale popularization; at the scenario level, application demands vary greatly across different industries, with obvious fragmentation characteristics, high customized development costs, and weak replicability, making it difficult to quickly form economies of scale; at the ecosystem level, standards across the industrial chain are not unified, coordination efficiency among chips, algorithms, complete machines, and scenario providers is low, and an efficient positive cycle has not yet been formed.
Among key elements such as data, computing power, algorithms, and hardware, which one do you think is most likely to become the "choke point" for the long-term development of the industry? How should we layout in advance to cope with it?
In the long run, data is the deepest "moat" in the embodied intelligence industry and also the most likely "choke point." Model architectures can be replicated, and parameter scales can be caught up, but high-quality interaction data in the real world must rely on long-term, scaled, and systematic collection and precipitation, which cannot be copied or purchased in the short term. Whoever masters the scale and quality of data will take the initiative in defining the next generation of embodied intelligence.
Regarding how to layout in advance, our core idea is: abandon the "data silo" mindset, use the "total cost of intelligence" as the benchmark, and systematically build data barriers. The "total cost of intelligence" mentioned here consists of three parts: data cost, computing power cost, and R&D team cost—we should not just look at the level of single-item investment, but evaluate the long-term intelligence efficiency brought by the synergy of the three. Based on this judgment, Xinghaitu takes action from three levels:
First, firmly bet on real data to establish advantages from the source. All pre-training at Xinghaitu is based on real data. Before accumulating 1 million hours of real-world data, frankly, we do not see the need to adopt simulated data—because the improvement of model generalization ability by real data cannot be replaced by simulation at present. We do not believe in shortcuts, we only believe in "from the real world, to the real world."
Second, jointly build scaled data infrastructure. In June 2026, Xinghaitu, together with Yizhuang Robot and Yizhuang Guotou, jointly established the data company "Yishu Intelligence," and used it as a carrier to launch the "1 Million Hours of Ultra-High-Quality Real Data Plan," planning to complete one million hours this year and move towards tens of millions of hours in the next three years. To intuitively understand this magnitude, we can make an analogy: the total time a person interacts with the physical world from 0 to 18 years old in an awake state is about 120,000 hours—one million hours is equivalent to the complete learning volume of 8 people, and tens of millions of hours is equivalent to over 80 people. When its "brain" possesses the embodied interaction experience equivalent to 80 human lifespans, its understanding of the physical world will be on an entirely different level from today’s systems.
Third, build an open data ecosystem alliance. We have jointly initiated the Embodied Intelligence Data Ecosystem Alliance with 15 enterprises including Ant Digital Technologies, Baidu AI Cloud, and Yuanli Lingji. We believe that the long-term barrier of data does not rely on monopoly, but on the ecosystem—ultimately, data is not a cost item, but a strategic asset. Whoever accumulates it with sufficient determination and in a sufficiently systematic way today will define the ceiling of the industry tomorrow.
High-quality physical world interaction data is the key bottleneck for the long-term development of embodied intelligence.
Embodied intelligence does not lack beautiful demos, but lacks reusable, evaluable, and sustainably generated physical interaction data.
Computing power and algorithms can be iterated quickly through general technology platforms, and high-end hardware can also be tackled through precision manufacturing. However, there is no shortcut for real-robot operation data that truly reflects the complex interaction logic of the physical world; it must be precipitated relying on real scenarios.
Currently, the industry generally faces problems such as a large "Sim-to-Real" gap between simulated and real data, high costs of real-robot data collection, and data silos in the industry.
To this end, Lingqiao Intelligence has laid out two things in advance:
First, independently developed the DexStream full-link data collection pipeline, integrating multiple channels such as real-person teaching, real-robot operation, and non-wearable optical capture, and open-sourced the DexCanvas large-scale human hand operation dataset to provide data fuel for the industry;
Second, deeply participated in the construction of the national AI application pilot base, relying on the verification mode of real scenarios, real tasks, real people, and real robots, accumulating high-value data in real working conditions, and feeding the data back to model iteration.
We believe that data is the most easily overlooked but most likely to become a long-term bottleneck.
The essential difference between embodied intelligence and pure language large models lies in that it requires "embodied data"—that is, high-quality operation data generated when robots perform tasks in real physical environments. The collection of such data is costly, difficult to standardize, and requires broad scenario coverage, which is much more difficult to obtain than text data.
Leju's coping strategy is to "walk on two legs": on the one hand, build a national-level embodied intelligence dataset through the OpenLET open-source community, gathering power from industry, academia, and research in an open-source way to lower the threshold for data collection; on the other hand, relying on its scaled deployment in scenarios such as industrial manufacturing and scientific research education, accumulate high-quality real-robot operation data, forming a data flywheel of "real scenario driving, open-source ecosystem accelerating."
In terms of chips and core components, Leju adheres to the full-stack in-house R&D route, with the localization rate of core components exceeding 95%. Moreover, it has joined hands with Huixi Intelligent and Weifan Intelligence to build a domestic edge computing power base, completing the "last piece of the puzzle" for the full-stack localization of embodied intelligence to achieve full-link autonomous control, avoiding supply chain risks from the source. We believe that the real "choke point" is often not a problem of a single link, but the lack of overall coordination capacity of the industrial chain, which requires the joint efforts of the whole industry.
Among these elements of data, algorithms, computing power, and hardware, the one most likely to choke industry development in the long run is high-quality, compliant scenario datasets. Catching up in computing power and general algorithms is relatively faster; although domestic substitution of components is difficult, there is a clear tackling path. However, data ownership in industrial scenarios is complex, with many security restrictions, severe data silos among enterprises, and a gap between pure simulated data and the real environment. It is recommended to rely on national-level platforms to build hierarchically desensitized shared datasets, supported by public simulation platforms; at the same time, use actual scenarios to drive the continuous iteration of high-end hardware such as joints and sensors.
Computing power and algorithms can be caught up through investment and talent, but the 3D physical data required for embodied intelligence cannot be obtained from the internet, and it is difficult to solve it on a large scale by relying on real-robot collection. Kuawei Intelligence chooses a data engineering route centered on generative simulation and multi-source data collaboration, efficiently generating simulated data that conforms to physical laws in the virtual world through the domestically developed DexVerse™ simulation engine. We released the Dexterity-BEV unified spatial representation framework in June this year, and open-sourced the world's first spatio-temporal action alignment dataset Aligned DexWorld bridging "human-simulation-real robot" in July, precisely to solve the "choke point" problem of data from the source.
At present, the embodied intelligence chip industry as a whole is in the initial stage of incubation and accelerating evolution towards product deployment. The solutions in the market are still mainly general-purpose computing chips reused from the fields of intelligent driving and consumer electronics. Dedicated chips deeply customized for embodied scenarios are still in their infancy, and various manufacturers are exploring the optimal technical paths and product definitions.
Core opportunities are reflected in 2 points: First, downstream demand is rapidly expanding. As the commercialization process of embodied intelligence accelerates, the industry's demand for low-power, highly adaptable dedicated computing solutions is becoming increasingly urgent, and the market space is broad; Second, the foundation for technology reuse is mature. Domestic manufacturers have accumulated deep technical capabilities in the fields of intelligent driving and consumer electronics AI chips. The underlying architecture and toolchains can be efficiently reused in embodied intelligence scenarios, possessing the ability to iterate quickly;
Realistic pain points are equally obvious: First, scenario fragmentation makes chip definition difficult. The computing power, power consumption, and interface requirements of different robot scenarios vary greatly. It is difficult to cover all scenarios with one chip, and the fault tolerance space for product definition is small; Second, the lack of industry standards. Unified specifications for computing power, interfaces, and testing have not yet been formed, resulting in high adaptation costs between upstream and downstream; Third, the pace of commercial deployment remains uncertain. The R&D investment for dedicated chips is large and the cycle is long. It needs to match a clear commercial deployment pace to form a positive cycle of R&D and deployment.
From the practice of Black Sesame Technologies, we chose to enter through the path of homogeneous reuse of automotive-grade technology, launching a serialized computing module to cover the needs of all categories of robots, and quickly landing multiple benchmark cooperative projects. This is precisely a targeted response to industry pain points and a pioneering practice of domestic manufacturers exploring the deployment path of embodied intelligence chips.
Author | Editorial Department of China Electronics News Editor | Wu Lilin Art Editor | Malia Supervisor | Lian Xiaodong