AI News

Embodied intelligence robots take up posts in scenic areas: who is making a lot of noise, and who is crunching the numbers

36Kr · AIUpdated Nov 16 18:48

AI summary

Nearly 30 embodied intelligence and robotics companies have entered cultural tourism, retail, and consumer service scenarios. AgiBot deployed more than 300 units at once at Chimelong in Hengqin, and Galbot has landed more than 200 space capsules in 50 cities nationwide. The three interviewed companies Zhidong Weilai, Zhi Pingfang, and SMYZE believe that whether robots can work stably for long periods, whether they can handle operational dirty work such as snapping on cup lids, and whether they can dilute costs through scale are the three hurdles for robots taking up posts. SMYZE says that when a single city reaches 100 stores, efficiency will comprehensively surpass staffed stores; it currently has 40+ locations deployed in China.

Body · AI translation

The long holiday has been over for a while now, but the holiday effect it set off still lingers. People who travel far and wide can probably hardly miss it: besides the crowds, scenic spots and shopping malls now have some rather different figures—those greeting guests at the entrance, expertly making coffee, handing over a drink, serving a table of fine dishes, playing guessing games with children... embodied intelligence. This group of "new new workers" has appeared in bulk in cultural tourism, retail, hotel and other commercial service scenarios, embedding itself into ordinary people's daily lives in a way that cannot be ignored.

According to incomplete statistics, nearly 30 embodied intelligence and robotics companies have now entered cultural tourism, retail and consumer service scenarios. AgiBot deployed more than 300 units at once in Hengqin Chimelong; Galbot has landed more than 200 space capsules in 50 cities nationwide; Zhiyuanfang's "Zhi Mofang" has already been deployed in cultural, commercial and tourism landmarks across more than ten provinces and cities in China, including Beijing's Chaoyang Park and Beijing Huaxi LIVE·Wukesong, Shanghai's Oriental Pearl Tower, Shenzhen's MixC INNO100, Shenzhen Talent Park, Haiancheng Park, Guiyang's Colorful Guizhou City, and Zhangzhou's Longjiang Suiyue Beer Plaza; Guangzhou Zhidong Future's Xiaozhi S3 has entered Xi'an's Grand Tang Mall and Hong Kong's Wong Tai Sin Temple, and deployed 16 robots at 8 main points across all venues of the 140th Canton Fair (in three phases), receiving a cumulative 189,000 business visitors from around the world (counting only the first phase, October 15–20); there is also the 7×24-hour SMYZE (Shangyin i-Shop) smart drink bar, which uses robotic arms to provide more than 100 categories of beverages, located in terminal buildings and subway stations in the core first-tier cities of Beijing, Shanghai, Guangzhou and Shenzhen, as well as cultural tourism scenic spots such as the Guangxi Museum and Wuxi Nianhua Bay... At the same time, the unified line from all relevant persons in charge is that more cultural, commercial and tourism landmarks are still being continuously expanded.

The bustle is real. But behind the bustle, a more real question is surfacing: are these silicon-based workers actually competent? And besides that, have they made money?

With this question in mind, we talked with three companies: Zhidong Future, Zhiyuanfang and SMYZE (Shangyin i-Shop). The three companies' products currently have different forms and focus on different scenarios, but their answers unexpectedly point to three common hurdles.

From "can demo" to "can deliver": is embodied intelligence really doing work?

The reason embodied intelligence is collectively rushing toward commercial service scenarios is not complicated.

Robots in factories do work that "cannot go wrong." If the cycle slows, capacity is dragged down; if yield drops, it means real money lost. Without several years of validation, it is impossible to fully let go. The home is another extreme: environments vary enormously—today a child throws toys on the floor, tomorrow an elderly person moves a chair—and the robot has to deal with a mess that would give most humans a headache just looking at it.

By comparison, scenic spots, hotels and shopping malls have instead become a relatively friendly "middle ground of the workplace": making a cup of coffee, handing over a bottle of water, giving directions, recommending a product... The task boundaries are clear, and a small mistake will not stop production or shut down a line. It can also accumulate data and polish products amid tourists' watching and tolerance, and even incidentally do brand exposure and make money at the same time.

But this is not just the industry's wishful thinking. One noteworthy signal is demand-side attitudes toward robots. Last year, when robots entered scenic spots, most tourists stopped to look at the novelty and left after taking photos; this year, when they enter scenic spots again, some people have already started asking, "How come you don't even have a reception robot here?" When "there's actually a robot" becomes "there should be a robot," and when onlookers taking photos becomes routine, this public acceptance and sense of expectation is simply a "qualitative leap" compared with last year.

On this, Gao Shan, co-founder and CMO of Zhidong Future, told Xiaguang AI Lab: "Cultural tourism and commercial business, in our own assessment, is a very good scenario in the current stage of humanoid robot development. Customers truly have demand and are willing to pay for it, and there is positive feedback between this and the market."

With real purchases, positive feedback, room for error and data accumulation—the first scenario for large-scale normalized deployment of embodied intelligence was thus chosen.

But "being chosen" is only a ticket to entry. The real first hurdle is: can robots work continuously, stably and autonomously in real commercial environments, rather than being stunning in demos and collapsing on the job?

Zhiyuanfang's judgment is clear: listing "products and services stable enough" as the first condition for robot retail to shift from novelty to efficiency-driven. For example, whether coffee or ice cream, what consumers ultimately buy is the product itself. The robot must be able to complete operations continuously and stably, ensuring output quality and service experience.

Take one detail as an example: in January 2026, Zhiyuanfang opened its first robot coffee shop (Zhi Mofang) in Shenzhen Talent Park. That robot was not only responsible for picking up finished products, nor was it an automatic vending machine with a robot shell added; it directly operated coffee equipment to complete the full production process. More than nine months later, "that robot is still that robot," and it has successively learned ice cream, matcha, milk tea, beer and cocktails, and can also sell lottery tickets, cultural creative products and blind boxes.

Photo shows Zhiyuanfang's general-purpose intelligent robot Aibao operating a coffee machine and ice cream machine. Image source: Zhiyuanfang official WeChat

Cheng Xiaoyu, senior vice president of SMYZE, analyzed that many robot projects are cool in demos and stunning when executing a single task, but when they actually go on duty working 24 hours, the manpower configured on the back end may exceed the scenario they were initially meant to replace. According to her, SMYZE's current comprehensive stability in beverage production is an error rate of 1‰, meaning only one cup out of every 1,000 has a problem.

She particularly emphasized that this stability is not about the robotic arm alone, but about the entire system—the coordination between the robotic arm and the various modules, and its integration with retail operation scenarios. If you only look at making the robotic arm move to a certain coordinate point to perform an action, problems are actually unlikely to arise. The real difficulty lies in whether the coordination between the robotic arm and other systems is reliable during uninterrupted operation.

Photo shows the SMYZE (Shangyin i-Shop) experience store providing consumers with a full range of beverage services

But stability is not just something that can be achieved by running in a lab. Gao Shan said that Xiaozhi from Zhidong Weilai faces disturbances in real scenarios that simulation environments can never replicate.

Stability is first reflected in whether it can work continuously for long periods. At this year's autumn Canton Fair, Zhidong Weilai's Xiaozhi S3 served continuously for more than 8 hours a day from opening to closing at 8 locations, during which it had to receive visitors, provide guidance, deliver coffee, and answer questions without breaking down midway. Gao Shan said this is also the practical reason why Zhidong Weilai currently chooses wheeled robots rather than bipedal ones—bipedal humanoid robots have a standby time of only about one hour longer, and simply cannot handle all-day tasks such as scenic area services, restaurant waiter duties, or the Canton Fair. Wheeled robots can currently achieve a standby time of more than 8 hours, and this is also the baseline requirement that Zhidong Weilai believes service robots should possess.

But battery life alone is not enough. In the robot restaurant at the world's first robot 6S store in Longgang, Shenzhen, Xiaozhi S3 faces another kind of test, such as a child suddenly running over while it is serving food. Such situations cannot occur in a lab, but they happen frequently in real restaurants. Currently, S3's approach is to make an emergency stop, wait for the child to pass, judge that the surroundings are safe, and then continue delivering food. "At the beginning, it also dropped plates," Gao Shan laughed. "But after 6 months of stable operation, the phenomenon of dropping plates is now close to zero."

Photo shows Zhidong Weilai's Xiaozhi S3 providing food serving service

Being able to work does not equal being able to deliver; only being able to deliver creates a real "rice bowl." The long-term stability, adaptability, and ability to handle non-standard situations of embodied intelligence in real scenarios are the first qualification certificate for robots to take up posts.

When robots work as waiters, who does the dirty and tiring work behind the scenes?

After passing the first hurdle, a more trivial question emerged: after embodied intelligence becomes a "waiter," who does the dirty and tiring work behind the scenes?

Cheng Xiaoyu shared a highly vivid detail: putting on lids.

Food safety requirements allow dine-in drinks to be served without lids, but takeout or quick-pickup modes must have lids. For putting on lids, all robotic arm companies' solutions use a joint turned upside down to "smash" the cup lid. But the lid is hard and the cup body is soft, so smashing makes it easy to tip the cup over, causing the drink to spill onto the table. Sugary liquids easily attract mosquitoes and insects, creating corresponding food safety risks and directly affecting user trust and purchase rates.

SMYZE did not ask robotic arm manufacturers to solve this problem, but instead solved it with another module in the entire beverage station system. Cheng Xiaoyu emphasized: "We are the only company in the industry that has solved the problem of achieving a 100% lid-pressing success rate."

Putting on lids sounds trivial, but it precisely illustrates the real difficulty of service robot operations: it is not about whether the robot can perform an action, but whether that action can be completed continuously, stably, safely, and at low cost in real operating scenarios.

Photo shows the SMYZE AI beverage station providing smart beverage selection services for large crowds at an exhibition

"Problems that need to be solved from the robot side may account for only 30%~40% of all operational problems," Cheng Xiaoyu added. Many problems are not entirely solved by robots, but if these problems are not solved, the system cannot be scaled. If it cannot be scaled, it will never become mainstream compared with stores operated by traditional manpower. "If we simply did not want to do this very down-to-earth dirty and tiring work, then robots might remain at the performance stage, or just be a passing trend."

Zhipingfang used the overseas launch of "Aibao" as an example to confirm that the real challenge of robots going overseas is not shipping equipment abroad, but enabling robots to adapt to local environments and provide services stably over the long term. Scenario adaptation, operational adaptation, localized services, and compliance and data requirements all need to be implemented one by one. For "Aibao" to take up a post in Lan Kwai Fong, Hong Kong, every step—from hardware export and customs clearance compliance to integrating into the operations of locally licensed commercial venues in Hong Kong—is a real threshold. Lan Kwai Fong itself is not exactly a "friendly" scenario either: high-density dynamic crowds, multiculturalism, noisy sound fields, complex lighting, and immediate social needs all pose comprehensive tests for the robot's perception, decision-making, and interaction. In bars, sudden lighting changes, people blocking the view, and swaying glasses—every variable can cause an action to fail. The NeuroVLA brain-inspired architecture carried by Aibao can reduce robot motion jitter by more than 75% and achieve a reflex response within 20 milliseconds after a collision, which is what enabled it to gain a foothold in "one of the hardest examination rooms."

Photo shows Zhipingfang's Aibao providing cocktail-making and interactive services for customers in Lan Kwai Fong, Hong Kong. Image source: Zhipingfang official WeChat

Gao Shan confirmed this from the delivery side. What Zhidong Weilai makes is work-performing robots, which face scenario delivery right from the start and do not have much room for trial and error. Gao Shan said the biggest feature of cultural and tourism scenarios is that they are "non-standard," and every scenic area requires secondary development, which also means delivery is extremely difficult.

In the Canton Fair case, Zhidong Weilai also did not roll out a large-scale deployment at the Canton Fair right away. Instead, from last year's spring session to the autumn session and then to this year's spring session, it provided lightweight services for three consecutive sessions, accumulating experience and debugging processes, before expanding at this year's autumn Canton Fair to 8 sites and 16 Xiao Zhi S3 robots, taking on tasks such as retail, guided tours, coffee delivery, meeting reception, and even assisting band performances. This year's challenge is greater than in previous years—overseas buyers have noticeably increased, which requires the Xiao Zhi S3's multilingual corpus to cover accents from dozens of countries, and voice interaction, navigation and guided tours, and reception and seating to all work in real crowds. Gao Shan said the Canton Fair is co-hosted by the Ministry of Commerce and the Guangdong Provincial People's Government, and the screening and requirements for entry are very high. Being able to provide continuous service from opening to closing in such a scenario is itself a comprehensive stress test.

The photo shows Zhidong Weilai's humanoid robot Xiao Zhi S3 providing guided tour services at this year's autumn Canton Fair.

Hotel scenarios are more complex. Gao Shan revealed that Zhidong Weilai began debugging and testing with hotels near the Canton Fair the year before last. Lobby services have already been connected, but room tidying can currently only handle the scenario of folding quilts, "and is not yet sufficient for all-round deployment." Gao Shan added, "If room tidying can be thoroughly connected, then we will basically be not far from entering households."

Precisely because delivery is so heavy, Zhidong Weilai can only do secondary development itself at this stage. They particularly want partners with this capability and are willing to open up resources, but at the current stage they can only seek stability, first debugging the scenarios themselves, and then look forward to subsequently replicating them to partners.

So service robots stand in the foreground, while what truly supports the show is often the invisible "dirty work and hard labor" behind the scenes: R&D systems, operations systems, supply chain management, scenario debugging, and data closed loops.

From "single-point showmanship" to "making the numbers work": how does this business actually pay back?

The cost of dirty work and hard labor is right there. Calculated at a single point, robots do not have an advantage. So how can this account be made to work? The answers given by the three companies point in the same direction: large-scale operations.

Cheng Xiaoyu told a common story in the industry. A friend brought in a robot. It performed well during the leader's first inspection, but later it stopped working well. However, the robot's position was conspicuous, and the leader often passed by, so it could not be shut down. They could only find engineers to work shifts and stand guard, repairing it immediately when problems arose.

"You saved the money of three store clerks, but added the money of three engineers. This account does not hold."

This sentence can almost be seen as the most plain account book for the commercialization of embodied intelligence as a whole. It is not that robots cannot work, but that the comprehensive cost of having them work is not yet low enough to replace human labor. Single-point accounting is not an area of advantage.

SMYZE's approach is to work backward from operations to products. They have studied humanoid robots. The advantage of a humanoid form lies in the sense of atmosphere, but the extra challenge brought by that sense of atmosphere is operational stability. Coordinated humanoid arms may collide, and operating speed must be slowed down. Yet users' core rigid demand is a tasty and affordable drink, and the added atmospheric value brought by a humanoid form is not enough to cover its increased costs and risks. So they chose robotic arms—precisely because they are more stable, more controllable, and lower in comprehensive cost, entirely out of commercial considerations.

More importantly, SMYZE does not calculate the account of a single device, but TCO (total cost of ownership). SMYZE is self-operated across the entire chain, from collaborative robot design and R&D to site selection and store opening, brand marketing, and beverage R&D, all completed in-house. All costs for equipment, personnel maintenance, restocking, marketing, and shrinkage are amortized internally.

In Cheng Xiaoyu's view, single-point operation is not cost-effective, but after grid-based operations, the logic changes. She introduced that currently one operations and maintenance person can manage sites in one city, and cleaning, capacity, operations and maintenance, warehousing, and delivery can all be shared across the store network. The marginal cost of adding a new site is very low.

100 is a number Cheng Xiaoyu repeatedly emphasized. "When we reach 100 stores in a city, our efficiency comprehensively surpasses that of staffed stores." Before reaching this inflection point, SMYZE will not expand stores particularly quickly. It is reported that SMYZE currently has laid out 40+ sites in China and is trying to be the first to prove out the business model of 100 stores in a single city in Beijing.

Gao Shan gave a more direct indicator from the customer side: customer recognition and repeat purchases. "Customers such as the world's first 6S robot restaurant and Hong Kong's Wong Tai Sin Temple all went through layer upon layer of comparison and screening, spent real money to buy our robots, and then debugged scenarios together with us. For us, this is a very precious real demand in B-end scenarios."

At present, the restaurant scenario of Zhidong Weilai's world's first 6S robot store in Longgang, Shenzhen, has been proven out at a single store. It is reported that a second high-end robot restaurant model is about to launch in Guangzhou. But Gao Shan also admitted that the closed loop of the business model still needs time, and it will probably be only by the end of the year or next year that the market share of robots in commercial scenarios can be reviewed, including judgments on ROI and store expansion rates.

Of course, after a single store is proven out, whether it can be replicated depends on the degree of standardization. Gao Shan clearly broke down the scenario lines for Zhidong Weilai's product applications: after several months of debugging, the restaurant line's ordering and serving functions have become basic functions in the product line; the guided tour scenario, after being tempered through four Canton Fairs, has also accumulated voice interaction, navigation, reception and seating, and actions such as bending over, nodding, and making finger hearts into standard capabilities. When facing different customers, these scenarios only require lightweight secondary development based on space and content, which will further accelerate delivery.

Zhijifang, for its part, said that ultimately what matters is real orders, repeat-purchase performance and total lifecycle cost. Getting a single site running is only the starting point; only when equipment reliability, restocking, operations and maintenance, quality management and back-end operations form a standardized system can it be replicated to more stores and markets.

Once the accounting reaches this point, it is no longer a question of how much one robot costs, but whether an operating system can spread costs thin and build up repeat purchases. Getting a single site running is just the appetizer; scaled operations are the precondition for the business to hold up.

Money is easy to make when things are buzzing; money is hard to make on ordinary days

During holidays, tourists are willing to gather around and take photos, and scenic areas are willing to spend marketing budgets in exchange for exposure. But as robots become more numerous, this value is being rapidly diluted. Rental prices already reflect this change. In the first half of 2025, the daily rental price of a popular humanoid robot was still generally 10,000 to 20,000 yuan; by 2026, basic models had dropped to 3,000 to 5,000 yuan, and on some platforms prices including engineer service had even fallen below 2,000 yuan. The emergence of low-priced packages is due on one hand to different versions of the equipment itself and to engineers spreading labor costs thin through remote support and one person covering multiple sites, and on the other hand to price competition after rental providers increased and supply expanded rapidly. Behind the ebbing of prices is the fact that the novelty premium is giving way to more pragmatic operating calculations.

So the industry's next step still has to return to that watershed: from "people are willing to pay for novelty" to "even when it is no longer novel, people are still willing to pay for efficiency."

Cheng Xiaoyu put it directly: no matter whether a robot is high-tech or low-tech, it must be able to provide services to people. Gao Shan always emphasizes that recognition from customers and real repeat purchases are the true sign of having cracked the model. Zhijifang said it hopes to drive robots to gradually move from eye-catching "novelties" to productivity tools that can operate stably, improve efficiency and create commercial value.

Put these three statements together, and they precisely form the three sides of embodied intelligence commercialization: technical capability determines whether a robot can take the job, operating capability determines whether a robot can continue to be needed, and the commercial closed loop determines whether this business can hold up.

The cultural tourism scenic-area robots during the long holiday have already proved the first point. But what truly determines whether this business can last is still the latter two. It is that capping action by SMZYE that must succeed 100 percent of the time, the judgment Xiaozhi S3 must make to avoid children in a restaurant, and Aibao's ability to keep expanding its business scope on the basis of long-term operations...

"What seems most ordinary is often the most extraordinary; what looks easy is in fact arduous."

Robots entering scenic areas and shopping malls, handing you coffee, selling ice cream, playing rock-paper-scissors with you—it looks like a lively and simple thing. But those who have actually done it know that from "can demonstrate" to "can make money," there are countless trivial and difficult operational details in between.

Money is easy to make when things are buzzing; money is hard to make on ordinary days. But only when stable cash flow can be generated even on ordinary days is it truly a business. Robots taking up posts is only the first step; the system-level operations that follow are what truly determine who can be "employed" over the long term.

This article comes from the WeChat public account "Xiaguang AI Lab," author: Vinco, published by 36Kr with authorization.

Source

36Kr · AI