The fate of embodied intelligence will not be decided by the size of anyone’s data pool, but by whether that data is physically reasonable and whether the data pipeline behind it can remain stable. That assessment, delivered by Dong Kai, director of the Science and Technology Division at the CCID Research Institute under the Ministry of Industry and Information Technology, set the tone for the embodied intelligence sub-forum at the 2026 JD Global Technology Explorers Conference, where industry leaders gathered to confront a sector that is expanding rapidly on paper while struggling painfully in practice.
The contrast on display at the forum was stark. Embodied intelligence is riding a wave of extraordinary enthusiasm, with revenue scaling at speed and a growing roster of specialized suppliers. Yet beneath those headline figures, the translation of embodied intelligence from laboratory demonstrations into frontline industrial work remains far less smooth than the surrounding hype suggests. Executives, researchers, and end users repeatedly described the same set of obstacles: hardware that does not match real operating environments, systems that cannot adapt to extreme conditions, and economics that fail to close. The result, as several participants put it, is a technology that can perform on stage but struggles to work on the floor and rarely earns its keep.
At the embodied intelligence sub-forum of the 2026 JD Global Technology Explorers Conference, the debate was marked by bottlenecks, unresolved disputes, and a shared recognition that embodied intelligence has entered a deeper and more demanding phase of industrial development, one with higher entry barriers and tougher competitive dynamics than the era of prototype showcases.
- Momentum at the Top, Friction at the Bottom
The growth story for embodied intelligence and the broader robotics industry is genuinely impressive. According to figures shared by Dong Kai at the forum, industry revenue grew at an average annual rate of nearly 25 percent during the 14th Five-Year Plan period, with total revenue breaking through RMB 300 billion. More than 700 specialized and sophisticated “little giant” enterprises have been cultivated, the installed base of industrial robots has surpassed 2 million units, and applications now cover 71 major industry categories.
Those numbers sketch an industry with considerable depth, a widening supplier base, and a footprint that reaches into nearly every corner of manufacturing and services. They also explain why capital, talent, and policy attention continue to flow toward embodied intelligence. The promise is that embodied intelligence will allow machines to perceive the physical world, reason about it, and act within it, turning general-purpose artificial intelligence from a software phenomenon into a source of physical labor.
Yet the industrialization reality is far harsher than the growth curve suggests. As large models and robotics technologies converge faster, embodied intelligence is accelerating from the laboratory toward the industrial frontline, but the conversion into deployed, revenue-generating systems is nowhere near as smooth as many expected. Frontline players at the forum acknowledged the same structural problems: mismatches between hardware and genuine application scenarios, difficulty adapting to extreme environments, and an imbalance in the economics of putting systems into production. Those factors leave most technologies trapped in a loop in which they can demonstrate well, cannot perform reliably, and do not make money.
That is the paradox of embodied intelligence in its current phase. The technology has never been more visible, and the gap between visibility and value has never been more obvious.
- Three Mountains Blocking the Road to Industrialization
In his keynote address, Dong Kai summarized the obstacles facing embodied intelligence as three mountains that the industry must cross before industrialization can begin in earnest: data, models, and applications. The framing was widely echoed throughout the forum, and it captures why embodied intelligence is proving so much harder to deploy than purely digital artificial intelligence.
| Mountains | Core Constraint | What the Industry Faces |
|---|---|---|
| Data | Scarcity of compliant real-world physical interaction data | Severe shortage of failure data such as falls, slips, object breakage, and human intervention; a proliferation of data collection methods with no consensus on which fits which model; data that does not transfer across different robot bodies; and missing unified annotation and alignment standards |
| Models | Model-driven capability built on both data and knowledge | Each model family has distinct strengths and weaknesses, and no final architecture has emerged; an embodied intelligence “ChatGPT moment” is at best three to five years away, and possibly a decade if defined as autonomous closed-loop completion of new tasks in unfamiliar environments |
| Applications | Integration, reliability, and return on investment | Robots cannot work independently and must integrate with existing deterministic industrial systems; cycle time and reliability gaps are enormous; return on investment remains hard to calculate; batch consistency is poor; and standards and ecosystem support are missing |
Dong Kai described the application mountain as the most seriously underestimated of the three. His reasoning was systematic. First, robots cannot work independently; they must be integrated into existing deterministic industrial systems. Embodied intelligence, however, produces probabilistic model outputs, and reconciling those outputs with fixed process parameters and system debugging requirements is extremely difficult.
Second, the gap in cycle time and reliability is vast. Traditional industrial robotic arms achieve a mean time between failures of 60,000 to 80,000 hours. Embodied robots today struggle to reach even 1,000 hours, and their adaptability to long-tail conditions is weak. Third, return on investment cannot be calculated simply by counting how many workers a machine replaces. A full system accounting, including operations, maintenance, and service costs, is required, and at present that accounting does not produce a clear answer.
Fourth, batch consistency is poor. Many embodied robots are assembled largely by hand, and when dozens or hundreds of units are produced in sequence, differences emerge in joint calibration and customized points. Fifth, standards and ecosystem support are missing. Every industry has its own requirements for explosion protection, fire safety, and other specialized conditions, yet a delivery and integration ecosystem has not formed, which drives the cost of secondary development extremely high.
- Seventy Percent of Products Need Deep Reconstruction
The mismatch between how robot bodies are designed and what application scenarios actually require surfaced repeatedly during the forum’s roundtable discussion. Xing Boyang, chief technology officer of the National and Local Co-built Humanoid Robot Innovation Center, offered one of the most striking estimates of the day: roughly 70 percent of products require a round of deep reconstruction before they can genuinely satisfy the real demands of a scenario.
He traced the problem back to design logic. Early robot body development largely followed a research-oriented logic. Decisions about payload capacity, end effectors, and even whether a dexterous hand was necessary were not derived from the actual operational requirements of the user. Once a robot enters a factory, non-standard hardware amplifies the difficulty of data collection still further. Dexterous hands differ from manufacturer to manufacturer in degrees of freedom, configuration, and sensor modality, which means data collected on one platform cannot be shared with another.
This is a defining challenge for embodied intelligence. Data is the fuel of the learning process, yet the hardware that generates the data is fragmented. When every dexterous hand speaks a different language, the resulting datasets cannot be pooled, and the economies of scale that have driven progress in other areas of artificial intelligence never materialize.
- When the Factory Floor Refuses to Be Standardized
In construction and manufacturing settings, the discreteness of the environment and the frequency of anomalies make standardized models even harder to apply. Yang Li, deputy director of the Digitalization Division of the Science, Technology and Equipment Business Unit at China Communications Construction Company, told the roundtable that the engineering construction industry is highly discrete, with enormous variation in geographical conditions, processes, and work objects. Standardized scenarios that a laboratory can barely handle run into dust, weak ventilation, and rough road surfaces on site, making industrial-grade reliability difficult to achieve.
The implications for embodied intelligence are significant. A model trained in a controlled environment learns assumptions that the real world does not honor. Dust interferes with perception. Poor ventilation affects thermal management and sensor performance. Uneven ground disrupts locomotion and stability. Each of these factors is manageable in isolation, but together they compound into a reliability problem that no single algorithmic improvement can solve.
Xiao Yongqiang, chief technology officer of EFORT Intelligent Equipment, added the manufacturing perspective. Frequent production changeovers, inconsistent incoming materials, and the most difficult problem of all, anomaly handling, create conditions that embodied intelligence is not yet equipped to manage. A human worker handles an unexpected situation as a matter of course. For a machine, the same situation represents an enormous challenge, because it falls outside the distribution of everything the model was trained on.
- The Economics That Keep Embodied Intelligence Off the Floor
Beyond technical obstacles, the economics of deployment are proving just as decisive. Yang Li noted that it is not only the user who must run the numbers; robot developers face their own difficulty in closing the commercial loop. Research and development spending must be recovered, and without volume production, neither side’s profit requirements can be satisfied. This economic constraint makes widespread adoption extremely difficult to achieve.
The problem is circular, and that circularity is precisely what makes embodied intelligence so hard to commercialize. Users hesitate to commit to expensive systems whose reliability is unproven. Developers cannot reduce unit costs without volume. Volume requires user commitment. Breaking that loop demands either a dramatic improvement in reliability, a dramatic reduction in cost, or a scenario in which the value is so clear that users are willing to absorb the risk. For embodied intelligence, all three remain works in progress.
- The Data Mountain: Missing the Failures That Matter Most
Among the three mountains, data is the one that most directly determines whether embodied intelligence can learn to cope with reality. Dong Kai explained that compliant data drawn from real physical interaction scenarios is extremely scarce. The shortage is most acute for failure data, including falls, slips, object breakage, and human intervention. Without such data, a model cannot learn how to handle anomalies, which is precisely where human workers earn their value.
Compounding the shortage is fragmentation in how data is collected. Approaches vary widely, and there is still no conclusion about which data collection method best suits which model. Data does not transfer across different robot bodies, and unified standards for annotation and alignment are absent. For embodied intelligence, this fragmentation is more than an inconvenience. It means that effort spent collecting data on one platform may not benefit another, and the cumulative learning that has accelerated progress in other domains of artificial intelligence is difficult to reproduce.
The data problem also has a physical dimension that purely digital systems never encounter. Text and images can be scraped, licensed, or synthesized at scale. Physical interaction data must be generated by machines moving through the world, and the world pushes back. Every fall, every failed grasp, every collision is an event that must be safely produced and carefully recorded. That makes data collection for embodied intelligence slower, more expensive, and more dangerous than data collection for language models.
- The Model Mountain: Searching for the Embodied Intelligence Breakthrough
At the model layer, Dong Kai identified the fundamental gap between embodied intelligence and earlier generations of intelligent systems: the former must be driven by models built from both data and knowledge. Current model families each have their own strengths and shortcomings, and none has established itself as the final answer.
He offered a candid timeline for embodied intelligence. The most optimistic estimate for an embodied model’s “ChatGPT moment” is three to five years. If the definition is raised to autonomous, closed-loop completion of new tasks in unfamiliar environments, the horizon stretches to roughly a decade. In the meantime, the practical goal is to make models usable within clearly defined scenario boundaries. The central direction is to perceive the physical world more precisely, understand physical laws, and produce actions that are efficient and consistent with those laws.
This framing matters for how the industry should judge progress in embodied intelligence. A system that performs impressively in a narrow demonstration is not the same as a system that performs reliably across the range of variation found in a real workplace. The former can be achieved with enough data from a controlled setting. The latter requires models that generalize across conditions they have never seen, which is a far harder problem and one that no participant claimed to have solved.
The knowledge component is what makes embodied intelligence distinct. A robot must not only recognize that an object is heavy; it must understand what happens when a heavy object is lifted at a particular angle with a particular grip. That understanding draws on physics, materials, and task context, and it cannot be extracted from raw interaction data alone. Building models that combine data-driven learning with structured knowledge remains an open research problem, and the absence of a settled architecture explains why the industry has not converged.
- The Application Mountain: Integration, Reliability, and Standards
The application layer is where the ambitions of embodied intelligence meet the hard requirements of industrial operations, and it is here that Dong Kai believes the difficulty is most underestimated. His analysis identified five distinct obstacles, each of which is sufficient on its own to slow adoption.
| Application Obstacle | Description |
|---|---|
| System integration | Robots cannot work independently and must be integrated with existing deterministic industrial systems, while embodied intelligence produces probabilistic model outputs, making process fitting and system debugging extremely difficult |
| Cycle time and reliability | Traditional industrial robotic arms reach a mean time between failures of 60,000 to 80,000 hours, while embodied robots struggle to reach 1,000 hours, with weak long-tail adaptability |
| Return on investment | Counting only how many workers are replaced is insufficient; a full system accounting including operations, maintenance, and service does not yet produce a clear answer |
| Batch consistency | Units are largely hand-assembled, so differences in joint calibration and customized points emerge across dozens or hundreds of units produced in sequence |
| Standards and ecosystem | Industries maintain their own requirements for explosion protection and fire safety, yet a delivery and integration ecosystem is missing, driving secondary development costs extremely high |
The reliability comparison is especially revealing. Industrial automation earned its place on the factory floor by achieving uptime measured in tens of thousands of hours. Embodied intelligence, by contrast, is still working to reach a fraction of that. For a production manager deciding whether to place a robot on a critical line, the difference is not a technical detail but the entire business case. Downtime is expensive, and a system that requires frequent intervention carries hidden costs that rarely appear in a purchase proposal.
Integration adds another layer of complexity. Factories are built around deterministic systems with predictable timing and behavior. Embodied intelligence is probabilistic by nature. Bridging the two requires engineering work that sits outside the core competencies of most robot developers, and the missing integration ecosystem means each deployment becomes a bespoke project rather than a repeatable product. That is why Dong Kai placed the application mountain last but described it as the most underrated, since it is the layer where all the other progress must ultimately be converted into value.
- The Numbers That Define Commercial Viability
Against this backdrop of structural obstacles, companies at the forum did not wait passively. During the roundtable discussion, participants from robot developers, application-side organizations, and platform companies shared the lessons and solutions they had developed through real deployments.
Wang Kun, founder and chief executive officer of Daka Robotics, distilled the lessons of three entrepreneurial attempts into three numbers: 100, 99, and 15. Each captures a distinct requirement for embodied intelligence to succeed commercially.
| Number | Meaning | Practical Interpretation |
|---|---|---|
| 100 | 100 percent of the original job | A robot must fully take over all work of the position it replaces, not just a portion of it |
| 99 | 99 percent success and reliability | This is only the minimum passing grade; at large processing volumes, even a 99 percent success rate leaves an intervention burden that cannot be ignored |
| 15 | Payback within 12 to 15 months | The target is to allow users to recover their investment within roughly a year to fifteen months |
The first number reflects a hard lesson. Wang Kun described earlier efforts, including a security patrol robot that could inspect but could not respond, and an unmanned delivery vehicle that could reach a residential compound but could not deliver to the door. Satisfying only part of a requirement delivers far less user value than the technology suggests. For embodied intelligence, partial capability is often indistinguishable from no capability at all, because the human worker who remains necessary to handle the uncovered portion is still on the payroll.
The second number sets the commercial threshold at 99 percent success and reliability. Wang Kun emphasized that 99 percent is a very low passing grade, not an achievement. When a warehouse processes a large volume, even a 99 percent success rate produces a number of interventions that must be absorbed by human staff, and those interventions erode the value proposition.
The third number establishes the payback horizon at 12 to 15 months. All three numbers ultimately point to a single outcome: repeat purchase. Without repeat purchase, there is no commercialization. Even if 5,000 or 8,000 units are sold, the cost of generating further growth the following year becomes extremely high if customers do not come back. For embodied intelligence vendors, this means the first sale is not the finish line but the beginning of the only test that matters.
- The Application Side Takes a Step Forward
The application side is also moving toward the technology rather than waiting for it. Zhao Huawei, chief engineer of the Beijing Metro Innovation Research Institute, proposed that application-side organizations work with partners on a three-part package. His argument was that the burden of preparing an environment for embodied intelligence cannot rest entirely with the technology provider, because only the user truly understands the operational context.
| Component | Purpose | Example |
|---|---|---|
| Foundation layer | Build the datasets and knowledge structures that define correct and incorrect conditions before robots are deployed | For inspection scenarios, prepare datasets covering normal meters and abnormal environmental states, build a professional knowledge graph, define what is right and wrong, and migrate that structure to agents and back-end systems |
| Standards system | Establish openness and operating standards for procuring robot bodies, including safety grading standards | Including procurement openness and operating standards for robot bodies, and in particular grading standards for safety fallback in inspection robots |
| Openness at the body level | Allow partners to integrate knowledge into the system rather than building closed, single-vendor stacks | Creating an open body-side ecosystem so partners can plug knowledge into the system in a modular fashion |
Zhao Huawei argued that completing this three-part package is what makes it possible to close the loop from perception to decision, execution, and iteration, and to fully work through a scenario. Once that is done, application-side organizations need not worry as much about cost levels, and robot partners can deliver knowledge into the system in a plug-and-play manner. Only then, he said, is the model genuinely mastered rather than merely piloted.
The logic here is important for the industrialization of embodied intelligence. Much of the current difficulty stems from the fact that each deployment starts from zero. Data must be gathered, standards must be negotiated, and integrations must be built one at a time. If application-side organizations prepare their environments and define their requirements in advance, the marginal cost of each additional deployment falls, and embodied intelligence begins to behave like a product rather than a project.
- The Platform View: Real Scenarios as the Training Ground
The platform perspective, presented by the head of JD Group’s intelligent robot business, connected the technical debate to commercial reality. In the laboratory, the measure of success is whether a task can be completed. In a commercial environment, the measure is whether the system responds to genuine needs, runs stably over the long term, and is cost-effective. In the effort to bring embodied intelligence to the ground, real scenarios are the best training ground.
JD is turning its physical spaces into training environments for robots. The company has designated locations including JD MALL, Qixian, unmanned pharmacies, logistics warehouses, and automotive service stores as training grounds where technology is tested and iterated rapidly in real commercial conditions. At the same time, JD is building an embodied intelligence data collection center and has already open-sourced a human-perspective dataset. Drawing on its supply chain capabilities and actual business scenarios, the company intends to accumulate real data on a continuing basis. Over the next two years, JD plans to complete the collection of 10 million hours of high-quality operational data to provide a data foundation for training and continuous iteration of embodied intelligence models.
That commitment speaks to the scale of the data challenge facing embodied intelligence. Ten million hours is a figure that reflects both the ambition of platform companies and the sheer volume of interaction required before models can be expected to behave reliably in the physical world. It also illustrates why platforms may occupy a distinctive position in the embodied intelligence ecosystem: they control environments where real work happens, they generate data continuously, and they have a direct financial interest in the reliability of the systems they deploy.
The convergence of platform-scale data collection with robot development is one of the more promising developments described at the forum. If data is the first mountain, then the entities capable of producing high volumes of physically reasonable, well-annotated interaction data are positioned to shape the pace of progress for embodied intelligence as a whole.
- A Long Race That Requires Both Technology and Collaboration
The industrialization of embodied intelligence is destined to be a long-distance race rather than a sprint. Crossing the three mountains of data, models, and applications requires not only technological breakthroughs but also industrial collaboration and deep cultivation of application scenarios. Every speaker at the forum returned in some form to the same conclusion: the problems of embodied intelligence are systemic, and they cannot be solved by any single company working alone.
The data mountain cannot be climbed without standardized collection methods, shared annotation frameworks, and platforms willing to invest in the unglamorous work of recording failures as carefully as successes. The model mountain will not be crossed until research produces architectures that combine data-driven learning with structured knowledge of the physical world, and until the industry accepts a realistic timeline rather than expecting an immediate breakthrough. The application mountain demands integration ecosystems, reliability engineering, consistent manufacturing, and standards that allow deployments to be repeated rather than reinvented.
Each of these requirements implies coordination across the boundaries of individual firms. Robot developers must design bodies around real operational needs rather than research convenience, which means working backward from what users actually require. Application-side organizations must prepare their environments, define their standards, and invest in the datasets and knowledge structures that make automation possible. Platform companies must build the data infrastructure and training grounds that allow embodied intelligence to improve continuously. Component suppliers, integrators, and standards bodies must fill the ecosystem gap that currently makes every project bespoke.
The encouraging signal from the forum was that these conversations are no longer theoretical. Frontline players are describing specific numbers, specific reconstruction rates, specific reliability thresholds, and specific data volumes. Those details may be less inspiring than the broader narrative of embodied intelligence, but they are the substance from which industrialization is built. The era of demonstrating capability is giving way to the era of proving value, and the companies that survive the transition will be those willing to work through the unglamorous details of reliability, integration, and repeat purchase.
As more enterprises prove willing to chew on the hard problems in real scenarios rather than controlled demonstrations, embodied intelligence can move from the laboratory to genuine productivity. The path runs through the three mountains, and there are no shortcuts around them. But the participants at the 2026 JD Global Technology Explorers Conference made clear that the industry now understands the terrain, and that understanding is the first requirement for crossing it.