The robot walks off the stage, and JD.com takes the stage.

Infrastructure creates possibilities. 

This was a key judgment Jeff Bezos made in the early days of Amazon. The emergence of AWS later essentially came from his discovery that many companies do not lack ideas, but the ability to build underlying infrastructure on their own.

This problem also appears in the embodied intelligence industry: technical capabilities are advancing rapidly, but for a robot to truly integrate into the physical world, far more than the ontology and model are required.

At the recent World Robot Conference, running, dancing and backflips are no longer new, and actions for life scenarios such as delivering water and grasping are becoming more and more skilled. In the past few years, the progress of humanoid robots in motion control, interaction and task execution has been visible to the naked eye. In the era when large models are trying to understand everything, robots seem to begin to understand the human world.

However, beyond the demonstration, robots still have to face a world far more complex than the stage: mixed noise and crowds in shopping malls, strict operation rhythms in warehouses, endless items in families, and a long cycle of a set of products from delivery, debugging to maintenance and recycling.

The stage demonstrates the boundary of capabilities, while the industry tests long-term real operation. There are many gaps between them — supply chains, data, delivery, operation and maintenance, and service systems.

This is an entry point to understand JD's current layout in the robotics sector.

Different from the technical demonstration of a single brand, JD presents a complete picture of the future city in the exhibition area: there is a full-range training system for robots, from data to models, from component procurement to maintenance services; there are real industrial park scenarios, such as industry, logistics, retail, housekeeping, etc.; there is also a future life where cyber sense and worldly atmosphere coexist, such as exoskeletons that allow the elderly to enjoy the fun of mountain climbing, small humanoid robots that can accompany children full-time and raise children with technology, etc.

No one knows when this picture will come true, in 10 years, 30 years or 50 years. Technological maturity is only one aspect, which depends more on when the industrial capabilities supporting large-scale entry of robots into the real world will mature. JD's goal is exactly to try to shorten this distance.

When robots leave the laboratory, growing from one unit to a thousand or ten thousand units, the problems will transform accordingly. If capabilities scattered in the industrial chain, including data, model training, component procurement, sales, performance and after-sales service, are still repeatedly built by each company, it will be difficult to form economies of scale. How to organize them into a reusable industrial infrastructure is the key for robots to move from "being able to be manufactured" to "being able to operate on a large scale".

Robots step off the stage, and JD officially steps onto the field.

The starting point of large-scale application, the challenge of industrial infrastructure

"Two years ago, people were asking whether this track was viable, and now they are more concerned about how to make this track better," the head of JD's intelligent robotics business summarized the change of industry sentiment to 36Kr.

The market has given some positive signals. On JD's platform alone, the sales of humanoid robots increased by more than 10 times year-on-year during this year's 618 shopping festival. This figure shows that consumers and enterprise customers are rapidly getting access to robots, but high growth does not directly equal industrial maturity. For a new category that is still in the low penetration stage, the rapid rise in sales means more that market attention has been formed, and whether product value, user experience and service system can be established in the long run still needs market verification.

Interestingly, the head of JD's intelligent robotics business is unwilling to call the current stage a commercial inflection point. He thinks a more accurate statement is that "large-scale application has just reached the starting point".

But the "starting point" does not mean that all robots stand on the same starting line. The technical maturity, application scenarios and commercialization progress of different categories vary greatly. Combined with platform data, he believes that the commercialization of robots can be divided into several levels.

Consumer-grade robots in fields such as education and family companionship have entered families and begun to undergo tests of price, interactive experience and after-sales service; delivery, cleaning and some industrial robots can already be used in vertical tasks, but the costs of customization, deployment and operation and maintenance still restrict the return on investment; general-purpose robots that can handle open home environments and long-sequence tasks are still constrained by model capabilities, data quality, and the coordination between the brain, cerebellum and ontology.

However, regardless of the difference in technical maturity, when they enter the market, they all face a common problem: there is still a lack of a sufficiently thick industrial undertaking layer between the prototype, the commodity, and stable service.

This gap was not prominent in the past. In the technical verification stage, the scale and volume of robot projects are relatively limited, and many problems can be solved by the founding team and engineers on site. With the increase of orders, the originally hidden costs begin to appear intensively.

In other words, once the industry enters large-scale application, it must rely on the system to solve problems.

Compared with mature consumer electronics, the complexity faced by robots runs through both production and usage. Upstream, the specifications of components have not been fully unified, the procurement scale is limited, and the mass production process still faces problems of cost and supply chain efficiency; downstream, after a robot enters the customer site, it often needs to go through survey, adaptation and continuous operation and maintenance, and the completion of transaction is only the starting point of the service chain.

The problem is that robots are still in the low sales stage, but they already need an asset-heavy and service-heavy industrial network. These capabilities have similar cost structures: it is necessary to invest in a network first, and then spread the cost by relying on sufficiently high usage density.

Even if the maintenance center only serves a small number of devices, it needs to be equipped with venues, engineers, testing equipment and spare parts; even if the passenger flow of offline stores is limited, it has to bear the costs of rent, prototypes and professional explanation. The annual delivery volume of leading enterprises is still counted in thousands or tens of thousands of units, and the equipment is scattered all over the country and even overseas, so it is difficult for a single outlet to form sufficient service density.

If every brand repeatedly builds maintenance outlets, spare parts warehouses, channels and service teams, it is difficult for limited sales to spread high fixed costs, which will eventually lead to low network utilization and high service cost per unit. If it is handed over to scattered third-party service providers, it will face the problems of inconsistent technical capabilities, accessory efficiency and service standards.

Repeated construction not only brings cost problems, but also further squeezes the R&D resources of robotics companies. The core competitiveness of robotics enterprises still comes from the ontology. Commercialization should have provided cash flow for technology R&D, but the closer to the market, the more resources sales, channels, delivery and services occupy, which may in turn consume the organizational capabilities of technology companies in the early stage.

When more and more companies are stuck in the critical leap from technology to the market at the same time, what is exposed is not only the short board of individual operation, but also the immaturity of industrial infrastructure.

The problem has thus changed from "how can robotics companies sell their products" to an industrial division of labor issue: who should build the underlying infrastructure that all enterprises need but can hardly be independently built by a single brand?

Platforms and enterprises, who will build the infrastructure

The value of platform-based enterprises is not to do all the work for robotics companies, but to connect the common links that were originally scattered inside each brand.

The first step is to organize the unformed market demand.

At the end of last year, Unitree Robotics opened the world's first offline experience store in Shuangjing store of JD MALL in Beijing. The two sides adopted the linkage mode of online JD self-operation and offline physical experience: Unitree provides products, prototypes and professional technical support, while JD provides venues, passenger flow, transactions, warehousing and distribution and after-sales services.

On the surface, this is a channel cooperation. But for a category with rapidly changing definitions, the value of channels is not limited to the sales entrance. Questions including what consumers are willing to pay for, which groups different robots are more suitable for, and what scenarios actually exist, all need to be verified.

Therefore, the platform connects brands and products on one end, and real users on the other, undertaking the task of making scattered demands identifiable.

According to Unitree's prospectus, JD is one of its largest customers during the reporting period, and the cooperation between the two sides has extended from consumer-grade robot sales to a wider range of commercial scenarios. This directly shows that large platforms are not only sales channels, but also begin to become important demand undertakers.

This is exactly why JD can become a sample worthy of observation.

On the one hand, the advantages of JD Retail's mixed mode of self-operation + third parties have made warehousing, logistics, supply chain and service networks a cross-brand shared system. Its entry into the robotics industry is equivalent to extending this organizational capability to a new category that has not yet completed large-scale application.

At present, JD has cooperated with more than 200 robot brands through the self-operation mode, with the platform undertaking commodity operation, customer service, warehousing and distribution, performance and part of after-sales services. The search, consultation, transaction and evaluation generated by about 700 million consumers and 8 million enterprise customers can be restored into product demand.

The head of JD's intelligent robotics business summarized this path as "going back to users", which actually answers a more prepositive question: what kind of robots do users really need?

A very interesting detail is that JD has divided robots with different uses on the platform into 16 detailed categories according to user demands. Sometimes when the brand itself has not figured out which customer group to target, JD first helps clarify its positioning. JD is becoming a pathfinder in this emerging consumer market.

After the demands are organized, the next step is to scale up the supply. JD therefore further extends its layout to industrial links such as supply chain, service and data.

Looking upstream, JD is trying to aggregate scattered common demands in the industry. Battery is a typical link among them. Different manufacturers have their own definitions in safety, capacity, size and communication protocols. The procurement volume of a single brand is limited, which not only leads to weak bargaining power, but also makes it difficult to establish stable inventory. As one of the officially announced strategic moves, JD is cooperating with more than 20 industrial partners to promote the standardization of robot batteries, and aggregate demands through centralized procurement, warehouse stocking and batch pickup.

For an industry that has not yet formed a million-level shipment volume, standardization is not only the result after maturity, but also the prerequisite for large-scale application to occur.

After the supply is scaled up, robots have to face another key problem: how to operate stably in the real world for a long time?

JD has currently built 8 robot maintenance centers in China, and the maintenance services have been expanded to Europe, the Middle East and North America. It also plans to build 80 RoboBase robot bases in the next five years, incorporating component production, complete machine pilot test, secondary development, maintenance and recycling into the full life cycle service system, and the relevant capabilities will cover more than 100 countries and regions around the world.

Its logic is similar to JD's construction of logistics infrastructure: first form a network with heavy investment, then allow multiple brands and categories to share, and establish a service scale that a single brand cannot support at the present stage in advance.

Whether this set of service capabilities can really run smoothly still needs to be tested in the on-site environment with high intensity and frequent sudden failures.

At this year's robot marathon, the "robot ambulance" running behind the 120 robots once became a hit on online media. At the upcoming Humanoid Robot Games, JD's "Robot Home" and robot maintenance personnel will also step onto the field to test their capabilities in actual combat.

Finally, to usher in its own "ChatGPT moment", robots also need to cross the data barrier.

Embodied intelligence requires a large amount of high-quality data from the physical world, but the acquisition of real data is far more difficult than that of Internet data. Internet models can learn from existing texts and images, but robot data must be generated through the operation of equipment in the real environment. JD happens to have a large number of real scenarios, such as actual business systems including JD MALL, 7FRESH, logistics supply chain and unmanned pharmacies.

In addition to robot operation data, high-quality human operation data is also equally important. For example, the operation process formed by JD housekeeping personnel in actual operations such as cleaning and sorting, after being collected and labeled, can become samples for robots to learn real tasks.

Obviously, JD has also realized this point. In its strategic layout, it also plans to collect more than 10 million hours of real scene data cumulatively within two years, and support the multi-modal interaction and scene adaptation of robots through capabilities such as JoyInside.

Putting these actions together, JD's robotics layout is not a simple superposition of several capabilities. On the one hand, it is reusing the existing retail, supply chain and service networks to connect market demands; on the other hand, it is rebuilding new infrastructure such as standards, RoboBase and real scene data for the robotics industry.

The latter in particular almost covers the full landing process of a robot from production to application: scenarios and data solve the problem of "how to train and iterate", standards and supply chains solve the problem of "how to manufacture on a large scale", and RoboBase solves the problem of "how to deploy and operate for a long time", thus truly building the industrial infrastructure of the industry.

With the large-scale application of robots, the industry's demand for infrastructure construction has begun to emerge. It not only needs to reduce the cost of commercialization, but also shorten the distance between technology iteration and market feedback.

Robots enter the real world

With the landing of robots, industry competition begins to shift from "single-point capability" to "continuously completing real tasks".

Embodied intelligence ultimately creates value through actions. Hardware is only the carrier of capabilities. Whether tasks such as handling, cleaning, tour guiding and production can be continuously completed determines whether the product is truly usable. Only when robots enter the real environment can they constantly expose problems, accumulate data and improve experience, thus forming a cycle between commercialization and technology iteration.

Judging from the layout disclosed by JD this time, its goal is not limited to becoming the largest robot sales platform, but to undertake the supply chain, scenarios, data and services required after products enter the real world.

This also gives JD's proposed "physical world operator" a more specific meaning. Robotics enterprises continue to determine the capability boundary of the ontology and model, while JD tries to make these capabilities find demands faster, enter scenarios, and complete iteration in