By Tiya, Edited by Yang Xuran
Grand narratives are inevitably receding.
As a focal point of industry discussion, embodied AI has shifted from debating what the technology can do to exploring how it can truly operate, continuously iterate, and ultimately achieve profitability in more specific scenarios.
This narrative shift is a true reflection of reality and a response to capital demands. In the second half of 2026, multiple mainstream manufacturers made significant moves in the 'robot + retail' sector. From commercial districts in first-tier cities to community convenience stores, 'robot employees' are penetrating diverse business formats at a visible speed.
Beyond industrial production lines and exhibition stages, retail stores are becoming the most intensive direction for the deployment of embodied AI. A clear signal has emerged: retail is no longer just a 'testing ground' for embodied AI, but a direct scenario for verifying its commercial value.
This means that in the current embodied AI market, the definition of value has almost been reconstructed: 'technical feasibility' and 'commercial pragmatism' are both indispensable. All robots that have entered the retail scenario to 'work' are striving to balance these two points.
The quietly emerging robot boom in retail scenarios should not be simply viewed as a short-term trend. It acts as a mirror, more specifically reflecting that the entire industry is moving from mere 'capability showboating' to 'value realization.'
This article is an in-depth value piece from the content team of 'Juchao WAVE'. We welcome you to follow us across multiple platforms.
Exploration
From the perspective of scenario verification, the retail scenario follows the basic path of robots integrating into human work in the two major scenarios of stage performances and factory operations. However, it also breaks through and even surpasses the limitations of original scenario verification to some extent.
Whether in stage performances, industrial production, or retail services, the deployment of robots in diverse scenarios is essentially reshaping the production and living patterns of the world. The deep logic of this entire process is not simply to pursue 'unmanned operations,' but to use robots as a hub to build a new human-machine complementary collaboration model.
From the technical and product perspectives, the requirements for embodied AI in these three scenarios have diverged.
Most industrial production robots are confined to horizontal, vertical, or 'grid-like' enclosed spaces for operations. Even forklifts and mobile robots roaming in factories mostly operate autonomously throughout the entire process, with relatively little direct interaction with humans.
Robots in stage performances appear more agile, but they mostly represent simple functional demonstrations with insufficient commercial monetization capabilities.
In contrast, retail stores fall exactly between the two. The retail scenario has standard SKUs and standard SOPs, which can meet the underlying needs for robots' continuous learning and capability iteration. At the same time, it features real customer traffic, random orders, and on-site interference, serving as a 'touchstone' to test the generalization ability and adaptability of robots in fixed scenarios.
The 'unmanned' trend in the retail industry has a long history. From vending machines and unmanned shelves to AI visual settlement, the retail industry has been seeking solutions to reduce reliance on human labor, but previous explorations have focused more on 'equipment automation.'
A report by the China Commercial Industry Research Institute shows that in 2025, the market size of China's unmanned retail exceeded CNY 48 billion, and is expected to reach CNY 53.626 billion in 2026. Vending machines were the earliest terminals in the industry to achieve an unmanned closed loop. The current 'robot + retail' trend is more like a further endogenous in-depth exploration, rather than just a fad that started in recent years.
However, some current robot retail initiatives merely overlay a robot shell and simple Q&A onto the existing unmanned retail system. The overall state looks more like a technology demonstration or 'performance,' which cannot be called true robot retail services.
The real turning point lies in the store-based deployment of robot retail and its integration with the retail system: robots in stores truly participate in the entire business process within the authorized scope, and 'selling things' becomes an extension of corporate business behavior.
Policy support, technological optimization, and commercial maturity jointly drive this transformation towards a smarter retail industry.
On the policy front, support for smart retail is increasing, and policy dividends are promoting the development of AI (Artificial Intelligence)-empowered unmanned services. Technologically, world models and tactile perception have successively crossed the usability threshold, allowing robots to directly take up posts in existing stores. Commercially, just as people are gradually getting used to the 24/7 service of palm-held smart assistants, after a period of calculation, enterprises have a clearer understanding of the dividends of robot employees being 'on duty around the clock' compared to before.
On this basis, the sustainable development of robot retail is almost certain. But this certainty belongs to the track, not to individuals.
Currently, this field has not yet fallen into the quagmire of involution. Therefore, the breakthrough samples of those who take the lead in breaking through are more worthy of research and reference.
Business Models
Driven by capital acceleration, embodied AI manufacturers in the retail track are growing wildly with mixed quality, and the reshuffle period has started ahead of schedule.
The wild growth is mainly reflected in the debate over forms—humanoid, dual-arm collaboration, wheeled composite, desktop-level, and other forms coexist. They make coffee, mix drinks, and pick items from retail shelves. The choice of form depends on functional and scenario needs, and there is no universally recognized optimal solution yet.
The coexistence of mixed quality and the reshuffle period is because the speed of capital influx often far exceeds the cycle of technological iteration and commercial verification, causing a 'time compression' effect in the industry. Many enterprises are forced to start scaling narratives before completing the product closed loop, which also means that enterprises that cannot achieve commercial self-consistency can only be forced out.
Overall, what has truly gone viral is still the humanoid robot, because the visual presentation of robots is more likely to arouse people's curiosity and represents a larger space for imagination. Manufacturers have different practical directions, giving rise to differentiated deployment paths and business models.
Galbot did not do excessive technical demonstrations, but directly carried out technical verification and commercialization transformation in real offline stores. Its Galbot G1 humanoid robot is based on a wheeled humanoid structure, focusing on the optimization and realization of the three most basic operational capabilities: 'grasp, hold, and deliver.'
This robot can rely on its wheeled chassis to quickly respond to orders. Its soft gripper has more stable grasping performance. The closed-loop transmission of 'brain-cerebellum-tactile perception' is relatively smooth, enabling fully autonomous operation without the need for remote control.
It can be seen that such overall planning aims to provide replicable smart retail solutions for offline stores.
Part of the capital's bet on Galbot is the expectation that it can transform technology into 'standardized retail labor.' However, the degree of unstructured nature in current real retail scenarios, the actual friction costs of human-machine collaboration, and the most realistic willingness of terminal customers to pay for 'labor substitution' all require further verification.
Sharpa, on the other hand, has taken the technical route of systematic dexterous manipulation, enabling robots to learn how to make ice cream.
In August this year, Sharpa collaborated with Dairy Queen (hereinafter referred to as DQ), a globally renowned ice cream and fast-food chain brand, to create a 'zero-renovation, fully autonomous, year-round' robot Blizzard restaurant, which was established at the DQ Wujiang Road store in Shanghai.
Like Galbot, Sharpa does not excessively pursue making robots look like humans. Instead, it focuses more on shaping more human-like thinking and action modes for robots, enabling them to transition from 'delivery personnel' to 'product makers.'
The advanced form of Physical AI is to deeply understand and complete complex long-horizon tasks based on physical laws and social norms in open and dynamic environments. For robots, this process includes prediction and multi-task decomposition before action, dynamic adjustment during action, and memory 'archiving' after task completion, to continuously optimize the next action.
Breaking down the technology, the robot is equipped with the Sharpa Wave five-finger tactile dexterous hand, which has 22 active degrees of freedom and can flexibly switch among a complete set of actions including 'grasp, pinch, push, pull, scoop, and pour.'
Compared to just 'looking like a human hand,' the true significance of this dexterous hand lies in its ability to converge retail operations that originally required multiple sets of equipment, multiple adaptations, and multiple debugging sessions into a reusable, scalable, and seamlessly connected physical interface for external interaction.
Sharpa's self-developed hierarchical end-to-end model, CraftNet, introduces mechanisms such as tactile, force, and state feedback in the underlying control. The robot adjusts its output force based on changes in friction, sliding, vibration, and force. Tactile and force perception is then fed back to the omni-modal world model, which 'saves' historical task data while planning and simulating the preparation of the next cup.
Sharpa's ice cream-making process is very much like a 'dynamic experiment.' It needs to face multiple physical entities such as cups, spoons, cabinet doors, and ice cream machines, fully autonomously completing 55 consecutive tasks from ordering to delivery, and it is not a flash demonstration.
It can be seen that in retail scenarios, robots will continuously touch and cross the 'critical point' of their ability to ultimately achieve long-horizon tasks.
This full-stack technological retail approach may also be learned and promoted by Sharpa and its embodied AI peers to richer food preparation such as other ice creams and burgers. The 'ripple effect' is clearly evident.
However, different food and beverage categories have significant differences in process flows, hygiene standards, and equipment compatibility. These practical constraints may, in turn, drive the accelerated emergence of more adaptable next-generation 'Sharpa-like' store formats.
SenseTime Shanhui is also an important player in this track. The 'Shaomai Go' robot store is the core carrier of its vision for AI (Artificial Intelligence) physical services for consumers.
Through the SenseMart OS retail physical operating system, SenseTime Shanhui hopes to form an open, shared, and 'plug-and-play' platform foundation. Different stores can not only access different robots and devices but also connect to brand, operator, and venue portals, shortening the launch cycle of new products, new devices, and new business formats through standard interfaces.
The complex smart retail chain is broken down into four types of capabilities: perception, interaction, decision-making, and execution, all scheduled by a unified retail brain. Each action can be independently trained and verified. In-store devices and shelves are also lightly modified to facilitate robot access. Robot employees are assigned to different stores to perceive service touchpoints such as coffee machines, ice cream machines, and cabinet doors, continuously completing designated work tasks.
Yi Shuai, co-founder of SenseTime Shanhui, frankly stated that a 'robot store' is not entirely equivalent to an 'unmanned store.' The essence of a robot store still needs to return to 'operation,' which obviously requires much more comprehensive capabilities than just robots selling goods.
According to the design, the 'robot store manager' will transmit interaction data back to the decision-making hub, becoming a key foundation for retail store operations. However, in key decision-making links such as new product introduction, brand co-branding, and large-scale promotions, human judgment will still be retained. Only in this way can the technical tension and commercial potential of the entire service system be maximized.
Embodiment capabilities, operation precision, and business systems jointly constitute the necessary support for the deployment of robot retail. The synergy of the three integrates perception, decision-making, operation, and business into a complete system for sustainable fulfillment and profitability. This means that robots entering stores is by no means a simple hardware installation, but the embedding of a systematic business capability.
Prospects
The situations of various enterprises vary, but taken together, robot retail is currently facing a reassessment of value: Can it truly be jointly regarded by capital and the retail industry as a long-term scenario for sustainable innovation, to be verified and continuously developed?
The influencing factors here are complex. Only by penetrating the surface can we better feel the more realistic business purposes of robot enterprises.
In the early stages of industry development, people's enthusiasm for robot retail stems more from curiosity and the impulse to check in. Eye-catching features will invisibly amplify consumers' novelty-seeking mentality, inevitably harvesting a wave of traffic in the short term, but this is just the starting point.
From the practice of manufacturers such as Galbot, Sharpa, and SenseTime Shanhui entering robot retail, the core driving force for robots entering stores at this stage is mainly the hope to rely on standard SKUs, standard SOPs, and real interaction data in retail scenarios to achieve large-scale real data collection, thereby feeding back into subsequent robot training.
When basic perception and interaction capabilities mature, the improvement in model performance from new data will significantly slow down. As the inflection point of diminishing marginal data value arrives, stores will no longer pursue the 'quantity' of data, but rather more challenging long-tail cases. Scarce data such as outdoor pickup, multi-language mixed commands, and non-standard interactions will become the focus of model breakthroughs in the next stage.
At this time, robot retail needs to maintain its 'retail essence' while evolving into a more future-oriented new business model. For example, the RaaS (Robot as a Service) model will be highly valued. Robot stores like Shaomai Go can become new touchpoints for brands to reach consumers under this model, and reversely promote the brand's product R&D and precision marketing.
To transform the novelty effect into long-term value, the key lies in creating unique consumer experiences or completing tasks that are difficult for humans.
This logic has been verified in practice. Retail is becoming the 'mother scenario' for robot capabilities. SenseTime's SenseMart OS is migrating from convenience stores to scenarios such as supermarkets and parks; Galbot is extending from retail to multiple fields such as pharmacies and healthcare.
On the other hand, the race among manufacturers is also showing that the endgame of 'robot retail' and 'unmanned services' is not to 'replace humans,' but to 'activate humans.' This is also a reflection of the sustainability of their business models.
Robot retail needs to maintain the flexibility and sufficient resilience of human-machine collaboration. Robots ensure the stable output of basic services, while human labor is used to refine scenarios, solve problems, maintain emotional connections with consumers, and so on.
Therefore, measuring the maturity of robot retail cannot be limited to the single dimension of 'replacing human labor.' Compared to simple human replacement, the precise coupling and value symbiosis of human and machine capabilities may become a more valuable evaluation metric.
From the overall market status, humans and 'human-like' employees still need to go through a relatively long run-in period. Robot retail enterprises that completely aim to replace humans may encounter varying degrees of difficulties and pressures.
Conclusion
The evolution and upgrade of robot retail is also a necessary path for the industry to return from technological imagination to business common sense.
The traffic brought by novelty will eventually fade. Only the human-machine collaboration model verified in real operations, replicable scenario capabilities, and more sustainable profitability paths can support participating enterprises in crossing the cycle.
Currently, whether it is the breakthrough of the critical point of data accumulation, the exploration of the RaaS model, or the formation of the consensus of 'activating humans' rather than 'replacing humans,' all point to the same direction:
Robot retail is shedding its performance attributes and striving to become a part of the retail industry and a new direction.
It may not completely reshape the industry as a disruptor, but as the most pragmatic testing ground and amplifier for the deployment of embodied AI, its significance has long gone beyond just selling bottled water or ice cream.
Only when robot employees truly learn to think, collaborate, and grow like store clerks, and even complete business and profitable work like store managers, will this revolution reach its true starting point.