Hongqi Chain Terminates Robot Unmanned Retail R&D; Robot Retail Remains at Proof-of-Concept Stage

Veteran convenience store chain's robot sales venture fails: honed tens of thousands of times in the lab, yet still drops bottles in front of the shelf

English text is being prepared.

Punching and dancing are old news for robots; folding clothes and making ice cream are the new highlights.

Nowadays, they sell drinks in retail warehouses and ice cream cones in ice cream shops, doing their jobs quite convincingly.

But one veteran retailer has taken a hard fall amid this wave of enthusiasm.

In March 2025, Hongqi Chain, the leading community convenience store chain in Sichuan, announced that it would jointly launch a project with a robotics company to develop a "robot unmanned vending system."

As soon as the news broke, the capital market was abuzz; the company's stock price hit the daily limit the next day, and as of the close on April 23, the share price had risen about 35.0% from before the announcement.

The market's imagination of "traditional convenience store + robot clerk" was thoroughly ignited.

However, after 16 months of R&D, the project was declared terminated.

Because they found that even after being honed tens of thousands of times in the lab, when it came to the real shelf, the robot still behaved clumsily and helplessly.

But in the embodied intelligence industry, robot retail has become a favored deployment scenario for many embodied AI companies, with "working" robots appearing in popular scenic spots and bustling commercial districts.

Have the robots successfully taken up their posts? What is the real situation?

Hongqi wanted robots to sell goods, but "this matter was still thought of too simply"

Faced with the enthusiasm of the capital market, Hongqi had to pour a basin of cold water itself.

In early July 2026, Hongqi announced that it was terminating the R&D of the robot unmanned vending system.

Hongqi Chain is not just a small supermarket; it is the first convenience store chain listed on China's A-share market and is regarded by the industry as the leader among convenience store chains in Southwest China.

So why did this listed company's robot project fail?

The project was launched in 2025.

In March, Hongqi announced that the company was conducting R&D and innovation on a robot unmanned vending system.

Soon after, Hongqi issued a risk announcement, mentioning that multiple robots were still unable to "smoothly take up their posts" and reminding investors to be rational.

Hongqi mentioned three major shortcomings in the robots' capabilities.

When facing dense shelves and irregularly shaped goods, the robotic arm's "grasping error rate is relatively high, and stability urgently needs improvement."

This exposes that the robot's ability to autonomously grasp goods is still far from sufficient.

For example, when facing products of different materials such as glass bottles, tissues, and metal cans, the robotic arm's grasping force, angle, and method are all completely different.

This places stringent demands on the robotic arm's motion precision, reaction speed, and the algorithm's real-time computing capability.

Under different lighting conditions, the robot's visual recognition system is prone to deviations, misjudging products or positions.

In addition, the system must ensure that the robot can flexibly avoid obstacles while maintaining stable data transmission to ensure smooth transactions, which poses a great challenge to hardware performance, software algorithms, and network communication.

At the time, Hongqi said it had completed the software and hardware testing of the robot in a laboratory environment, and the prototype was expected to officially roll off the production line in May.

But "after all, there is a big difference between the laboratory environment and actual application scenarios," so "there is still some distance before the product is officially put on the market and sold to consumers."

Sixteen months later, the prototype never arrived; instead, what came was Hongqi's announcement terminating the system's R&D.

**△**Hongqi Chain announcement

The announcement stated that the system has not yet reached the standard for large-scale commercial deployment. Based on considerations such as the input-output ratio, the company plans to terminate the R&D of the robot unmanned vending system and shift to the scaled operation of 24-hour cloud monitoring services.

English text is being prepared.

Moreover, when it comes to convenience stores, every real scenario the robot encounters is more complex than the experimental scenario. For example, the aisles between shelves are relatively narrow, and the robot also has to work with customers or store clerks around.

After sufficient experimentation, Hongqi found that the robot could not achieve the expected results, so it reluctantly halted the project.

This industry insider believes that robots do well in performances, but there is still a long way to go before they can do work in complex scenarios.

Hongqi did not disclose exactly how much it invested in this project, but its 2025 annual report shows that Hongqi's R&D expenses were about 2.8668 million yuan.

English text is being prepared.

From this, it can be inferred that the R&D expenditure for this project is most likely around 2.8 million yuan.

English text is being prepared.

Most robots selling goods only serve the function of "picking up goods"

Although Hongqi's attempt failed this time, robot retail is being deployed as a hot track in the embodied intelligence industry.

The most common form at present is the standalone robot retail warehouse. Such stores are suitable for industrial parks, commercial districts, and cultural tourism sites, and at the same time serve both to sell goods and to attract passersby to stop.

At present, there are more than 100 robot retail warehouses or retail setups across the country, distributed in more than 30 cities, and even deployed in Hong Kong.

The other category is robots entering semi-open scenarios to work, such as coffee shops and ice cream shops. There are also robots that "quietly contribute" behind the scenes, such as picking up goods in pharmacies.

In open-plan supermarkets, both the tasks and the environment are more complex. Most robots only handle greeting, explanation, and promotion functions, such as handing out samples, guiding customer flow, and providing product information.

According to QbitAI ROBO statistics, at least a dozen embodied intelligence companies are currently running retail projects, attempting vending, freshly made drinks, pharmacy picking, and in-store shopping guidance.

Judging from past media reports, robots from 17 companies have already entered real operating stores or standalone consumer locations, with projects spanning regular operations, trial operations, and in-store demonstrations.

In retail warehouses and convenience store scenarios, robots appear to already have "jobs." How are they actually performing at work?

QbitAI ROBO visited several robot retail warehouses.

In a plaza in Zhongguancun, Haidian District, Beijing, a silver robot retail warehouse of about 10 square meters is in operation.

When you pass by the retail warehouse, it speaks in the voice of a young man: "Friends on the right, come take a look. Spring Festival Gala stars are here working too. The retail shop has lots of good things—come play with me."

Approaching the sales window, you see a silver-white robot waving its arm at the window to attract customers. When pedestrians pass by, it can actively recognize them and greet them.

The robot's hands are in the shape of two-finger grippers. Its two arms swing flexibly up and down, and its lower body is wheeled.

On the two shelves behind it are neatly stacked bottled water and beverages of similar specifications, while another shelf holds children's smartwatches in different colors and tube-packed potato chips. There is also an automatic coffee machine at the rear right.

Compared with the 2,000 to 3,000 SKUs (stock keeping units) of a traditional convenience store, the retail warehouse's products appear very limited.

With its arms clamped to both sides of its body, its wheeled lower body first moves it to the shelf where the product is located, then it extends its right gripper to grasp a beverage bottle and slowly places it on the table at the sales window.

The whole process is relatively slow, taking about 20-plus seconds.

Getting coffee works on the same principle. After receiving an order command, the robot moves to the automatic coffee machine, uses its gripper to pick up the finished coffee, and brings it to the sales window.

Throughout the process, the robot only plays the role of moving and retrieving goods.

After the pickup process ends, it pauses for a second or two, then returns to a new round of interaction, still using the same greeting script mentioned at the beginning, as if it has forgotten everything it has already interacted about—like a goldfish with no memory.

But the robot here is not able to sell goods independently inside the warehouse as one might imagine.

A staff member on duty inside the warehouse told QbitAI ROBO that the retail warehouse operates from 10 a.m. to 10 p.m. daily and requires someone to be on duty at all times.

The on-duty staff member needs to restock at any time, charge the robot, and report when the robot malfunctions.

After a malfunction is reported, operations personnel repair and test it remotely online.

According to the staff member, the products on the shelves need to be arranged neatly by people in advance. If they are placed crookedly or out of stock, the robot is unable to handle it and can only get stuck in place or wait for the clerk to restock.

At the checkout counter of another convenience store chain in Zhongguancun, a robot is waiting for customers.

Although it is in a convenience store with a more complex environment than a retail warehouse, its usage logic is still the same as in the retail warehouse. It can only move within the narrow space at the front counter, and the only things it can pick up are several types of uniformly packaged beverages and chewing gum in the cabinet behind it. The newly added function is that it can now grill sausages.

Why does the robot only operate at the front counter? Zhou Yun, a tester on duty, told QbitAI ROBO that letting it move around the whole store would delay the clerks' work. "It moves relatively slowly, and I'm still debugging it little by little."

Zhou Yun said that when the robot was first deployed in the store, it could not go on duty directly. "Its gripping force or position needs to be debugged slowly to find the most suitable position."

At present, the robot's hands in the store are in the shape of grippers. "The softness and hardness of different product packaging vary. We have it grasp relatively hard products, such as chewing gum or bottled drinks. For soft-packaged products, we recommend using dexterous hands to grasp them, but the cost is relatively high, so dexterous hands are not used in this scenario."

Clerk Xiao Tao admitted that the robot is still "quite far from working like a person. It is not as flexible as a person, but the more it is used, the better it gets." At first, it took nearly a minute for the robot to pick up goods, and now it has improved to a dozen or so seconds.

Although efficiency is low, one has to admit that the robot's interactive ability is strong.

This may be one reason many retail stores are still willing to bring in robots.

When picking up goods, from the moment a customer appears until the transaction ends, it can keep "finding topics."

Its vision extracts customers' clothing features for interaction: "This friend in the red hat, please come and stroll around the space warehouse as you like."

When picking up goods, it reminds customers, "Is your phone (camera) ready?"

When the order ends, a string of auspicious phrases comes out: "Thank you for your support. May everything go as you wish and all your wishes come true. Thank you for coming to support us. I wish you happiness every day. Bye-bye."

It provides plenty of emotional value. If you go to other convenience stores, no clerk will interact this frequently.

With emotional value comes more foot traffic. Clerk Xiao Tao said the in-store robot's display and visit functions are greater: "Many people make appointments with the (robot) company to visit."

QbitAI ROBO saw two women pulling suitcases in the store filming a video and interacting with the robot to buy bottled water. One of them said they were in Beijing on a business trip and it was their first time seeing a robot, which they found especially novel, so they filmed a video to record it. "The experience was quite fun. I think (the robot) is very agile."

Is robot retail still in the PoC stage?

QbitAI ROBO's visits found that most of the retail robots currently on duty cannot independently handle stocking and selling. So, is it cost-effective to "hire" a robot to work now?

Most robot retail projects on the market are developed and deployed by embodied intelligence companies on their own initiative. There are standalone retail pods or robots that handle retail work. Embodied intelligence companies cooperate with supermarkets, cultural tourism sites, chain ice cream shops, etc., pushing robot retail pods or individual robots into various scenarios.

Embodied intelligence companies doing retail can be roughly divided into three categories by delivered product.

There is the robot product type, centered on self-developed models and complete robots, paired with sales pods and modular mini-stores; the core of delivery is the complete robot and its manipulation capabilities.

There is also the embodied brain and operation solution type, centered on the embodied brain, combining models with robots and store systems, delivering picking and fulfillment solutions, including hardware adaptation and on-site deployment.

There is also the retail operating system type, connecting the capabilities of different robots to product recognition, orders, transaction fulfillment and store operations, providing operators with deployable mini-store solutions.

All three types of companies may develop "brains," and retail has thus become a stage for these companies to showcase their "brain" capabilities.

A single order command can get a robot to string together understanding the request, recognizing goods, planning actions, grasping, verification and delivery; in the process it can also interact with people in real time. It does indeed look fairly intelligent.

But can robot retail projects actually make money?

First, let's look at the cost of a robot retail pod.

On e-commerce platforms, a retail robot model in use costs nearly 700,000 yuan. Including the retail pod as a whole, industry insiders estimate the per-unit cost at nearly one million yuan.

But a retail pod probably carries only a few dozen SKUs, and the types are very limited.

According to public media reports, one robot mini-store generated 21,000 yuan in weekly revenue per store during the 2026 "May Day" holiday; some locations of another robot company exceeded 10,000 yuan in single-day sales during the 2026 Spring Festival.

Another listed AI group company's 2025 performance announcement shows that new business revenue, including robot retail, accounted for 6.0% of the group's revenue.

And this new business segment also includes three other businesses, so the share of retail revenue is actually even lower.

Jia Chuan, a veteran industry practitioner and former industry sales manager at UBTECH, believes the current robot retail business model is not yet mature, and many sales scenarios remain at the PoC (proof of concept) stage.

Small-scale testing is used to prove whether a robot company's technology, products or business model actually work in reality.

Simply put, the robot's capabilities are still "not very valuable."

Since it doesn't make money, why are robot companies still keen to send robots into various retail scenarios?

The truth behind it may be that robots need deployment scenarios more than (retail) scenarios need robots.

Zhou Xiang, head of retail business at a company that delivers robot retail brains and solutions, told QbitAI ROBO, "Lots of robot hardware companies come to us, because right now they have no real scenarios to deploy in."

In Zhou Xiang's view, robots dancing is becoming more and more common; factories have extremely high precision requirements; home scenarios are still out of reach for now, as robots lack generalization ability and safety is not guaranteed; data collection factories are now evolving more toward collection without a physical body.

Robots are facing the reality of supply exceeding demand.

And retail scenarios, this intermediate form with "high visibility and high fault tolerance," are a very good choice in this awkward period.

Jia Chuan revealed that factories have very low fault tolerance for robots: "Not a single cent or inch can go wrong."

But in retail scenarios such as cultural tourism sites and commercial districts, consumers are very tolerant of robots: "They're afraid you won't mess up. If something goes wrong, I'll quickly record it and post it online so everyone can take a look."

In the view of multiple industry insiders, retail scenarios generate little profit and cannot truly support robot companies' revenue and profits; they are products with a "marketing attribute."

The distribution of robot retail in cities has a common feature: it specifically targets places with heavy foot traffic, all in the core commercial districts of various cities or must-visit photo spots for tourists, such as the Summer Palace in Beijing, Chunxi Road in Chengdu, West Lake in Hangzhou, and Gulangyu in Xiamen. These places have very large visitor flows.

For embodied companies, every pod and robot is an advertising slot.

These robot companies can reach many people without appearing on the Spring Festival Gala or running large-scale ads.

English text is being prepared.

Robot retail, at most a cultural tourism project?

So who is paying for the retail pods?

Jia Chuan mentioned that cultural tourism scenic spots and city landmarks are important deployment scenarios for robot retail pods, accounting for 60 to 70 percent.

QbitAI ROBO learned that some local cultural tourism entities have signed strategic cooperation agreements with embodied intelligence companies, engaged in asset-light cooperation, and even set up joint ventures with them to "jointly expand cultural tourism applications."

Equipment may be purchased by scenic spots, operated by robot companies themselves, or run by local operators.

Jia Chuan has communicated with cultural tourism authorities in many places multiple times. He said that cultural tourism has a real demand for robots because they can "attract eyeballs." "They can also show off that the local area is actively embracing new technology. Traditional cultural tourism projects, such as handsome men and beautiful women dancing and singing, are no longer enough to attract attention. Robots are a relatively novel thing; there are always people who have never seen them, so they can drive traffic and attract tourists to take photos."

But both Zhou Xiang and Jia Chuan believe that selling robot retail pods is not a sustainable business model. "It's just a cultural tourism project, not a retail commercial project."

First, the repurchase rate of retail pods is very low.

Lin Qi, a salesperson at an embodied intelligence company that makes pharmacy pickup robots, believes that retail robots in cultural tourism or commercial districts mainly serve to display rather than to work. Customers buy goods just for novelty and to take photos, and repeat purchases are relatively rare. "Most people go to experience it because a robot is making ice cream, not because they really want to eat ice cream. (An ice cream) takes 6 minutes to make."

Second, if only cultural tourism scenarios are deployed, the number of scenic spots nationwide is finite, so the number of retail pods needed for scenic spots also has an upper limit.

"A retail pod costs several hundred thousand or even nearly a million yuan, but apart from local cultural tourism entities, no commercial customer will pay for it," Zhou Xiang said.

But embodied intelligence companies are still working hard to cooperate with local cultural tourism entities. "This is the same logic as the original data collection factories. Local governments build data collection factories, purchase a batch of robots, and generate revenue for robot companies."

In Lin Qi's view, if it were not for telling stories and attracting capital, retail robots would not need to be humanoid, because as long as they can pick up goods, flexible dual arms are enough.

According to his observation, no embodied company's robot shipments have reached more than 100 units in any single retail scenario; "they are all exploring."

"Everything comes down to cost." In Jia Chuan's view, the value that robots can currently provide is not enough to cover their purchase cost and operating cost.

Only when the cost of a robot providing services is cheaper than the cost of hiring a real store clerk will merchants be willing to pay. "If there is no clear time point for this, I don't think it counts as entering commercialization."

But the other side of the coin is that problems can only be discovered through continuous operation in real production environments. "When 99% of the problems robots encounter in exploration scenarios can be solved, that will be the commercialization node."

During the sales process, Lin Qi often hears customers say their expectation for robots is that "the robot can at least replace one store clerk."

Replacing humans with robots will take another two to five years

So what problems still need to be solved before robots entering retail scenarios can truly replace human sales clerks?

First is the safety issue.

Lin Qi believes that the biggest problem in the embodied industry is safety.

Robots in open scenarios need to interact closely with people. If they move too fast or have too large a working radius, they will create safety hazards. For example, if a child walks back and forth in front of a robot, current technology cannot guarantee that it will not hurt anyone.

Customers walk around, and robots also need to avoid obstacles and replan routes while working, which greatly reduces efficiency.

Lin Qi said that robot commercialization deployments are not standardized projects but customized projects, requiring more or less modification of the site.

For example, the floors of pharmacies and convenience stores are uneven, and wheeled robots cannot pass through. In addition, there is no unified standard for shelf heights and distances between shelves in each store, so the robot's movement trajectory, navigation position, and grasping motions need to be fine-tuned.

Therefore, at present, robots taking up posts require "a packaged solution, not a standard robot."

How will robot retail commercialization develop in the future?

Lin Qi believes that the prices of robots and retail pods must be low enough. The cost of one robot needs to be lower than the cost of hiring one person, and a pod should ideally pay for itself within a year.

The variety of goods in the pod should be rich, making it more like a mature store. Only in this way can a profit model be formed, which is in line with the essence of retail. "The endgame should be similar to the newspaper kiosks of the past: more locations, then lower the cost of opening a single store and increase density."

Several industry insiders mentioned above all said they believe the robot industry has not yet reached the critical point of commercialization.

Robots' capabilities are still at a preliminary stage, the technology has not converged, and generalization ability is insufficient, so they do not meet the needs of real scenarios. "Robots cannot yet fully replace human labor or achieve the same level as humans. This takes time."

In the future, how can robots develop better in retail scenarios?

Zhou Xiang mentioned that his company's robots already have restocking capability, but efficiency is very low. "At present, restocking is only a means of data collection." In addition, battery standby time is short, so robots still cannot get rid of human supervision and operate retail pods independently.

The next goal is for retail pods to get rid of fixed manpower, allowing one operations and maintenance person to "cover" multiple stores within a radius of several kilometers, moving around to check for abnormalities, replace robot batteries, and restock goods, ensuring the normal operation of retail pods.

He believes that various devices (microwaves, juicers, coffee machines) can continuously be added to the pod, allowing the robot to learn to operate one new device every day. "First learn to operate various devices in closed scenarios, then enter open scenarios. This technology path is smoother, rather than suddenly jumping from point A to point B. Before robots have learned enough actions, simply doing open scenarios is meaningless."

In Zhou Xiang's view, retail robots will definitely move from unmanned pods to semi-open scenarios (staying at counters) in the future, and ultimately enter open convenience stores and supermarkets to truly replace human store clerks.

And in retail scenarios, what robots need to possess is definitely not just grasping capability.

It also needs to expand many capabilities, including making non-standard items such as grilled sausages, selling tea eggs, preparing instant noodles, frying chicken cutlets, and oden, and its skills will gradually become stronger.

So how long is it before robots truly replace human labor?

Multiple industry insiders believe that this time is about two to five years. "The embodied industry needs a GPT moment, and that time has not yet arrived."