Shanghai International Embodied Intelligence Skills Competition kicks off, with robots competing in real-world scenarios of inspection, quality control, retail, and elderly care

Pills must be weighed one by one, test tubes must be placed into the centrifuge in order; on supermarket shelves, products with similar packaging must not be picked up by mistake, and when restocking, robots must also avoid customers; in nursing homes, robots not only have to deliver items and patrol rooms, but also be able to communicate with the elderly... Walking into the competition venue is like walking into a laboratory, a supermarket, and a nursing home. Over the past two days, the Zhangjiang Science Hall has been transformed into a "super exam room" for robot applications across the country. The test questions for robots competing in skills are no longer just action demonstrations on exhibition stands, but rather a series of difficult problems drawn from real work.

From October 23 to 25, the 2026 Global Developer Pioneer Conference and International Embodied Intelligent Skills Competition is being held in Shanghai. During an early visit to the venue, reporters saw that real-scenario tracks such as substation inspection, quality testing, supermarket retail, emergency search and rescue, and elderly care and rehabilitation had already been set up, awaiting robot candidates from all over the country to "enter the field."

From "looking real" to "being able to do the job," what counts as a real scenario?

Why does a robot competition move its venue into laboratories, supermarkets, and nursing homes?

"The key to a so-called 'real scenario' does not lie in how realistically the venue is set up, but in whether the tasks the robots face come from real business operations and whether the assessment standards correspond to actual work," said Zhong Junhao, secretary-general of the Shanghai Artificial Intelligence Industry Association. Shanghai is hosting the embodied intelligent skills competition for the second year, he noted. Compared with last year, this year's events have been further refined: robots must not only demonstrate "what they can do," but also prove whether they can complete specific tasks, adapt to real environments, and meet the requirements of actual work.

The changes are reflected in the design of the test questions. In the past, a robot might only need to complete a relatively independent action, and the scenario was imagined by designers; today, a task often includes multiple consecutive steps, testing not only whether an individual action can be performed correctly, but also whether the steps before and after can connect and whether the entire process can be completed smoothly.

Take the inspection track as an example. The venue presents a real inspection scenario from the State Grid. Robots must navigate complex road conditions and sequentially complete tasks such as reading and reporting meters, infrared temperature measurement, identifying equipment abnormalities, checking switch status, and detecting personnel and environmental safety. Inspection is no longer simply "walking around and taking a photo," but rather consistently carrying out work such as discovering problems, identifying abnormalities, and recording information.

Among the more than 1,200 real business scenario demands collected nationwide, inspection-related needs account for about a quarter, involving multiple fields such as electric power, security, factory management, and nighttime cultural tourism duty. Participating teams first refine their algorithms in digital simulation scenarios, and after passing screening, they enter the offline venue to be tested in a real environment. Moving from virtual scenarios to physical operations tests whether robots can cross the threshold from "being able to calculate" to "being able to do."

A single pill, a single item of goods, a single care task—what is tested is not just dexterity

In the quality inspection track, a laboratory scenario from the Shanghai Institute of Quality Inspection and Technical Research has been brought into the venue.

Robots need to sequentially complete weighing pills, unscrewing bottle caps, dropping in pills, loading test tubes, operating a centrifuge, and generating an inspection report based on the test results. The actions may seem simple, but they are not easy to carry out in practice: unscrewing a bottle cap requires force control, and opening the centrifuge chamber door and taking out test tubes requires stable movements. A slight deviation may knock over containers, affect operations, or even damage equipment. Liquid dispensing is listed as a high-difficulty bonus item: robots must use a syringe to extract liquid from a large test tube and then precisely transfer it to a small bottle.

"This is a real paid order scenario in a laboratory," said Li Yuanjia, manager of the event. This tests not only whether the algorithm can accurately identify targets, but also the stability of the robotic arm, the control capability of the dexterous hand, and whether visual judgment and actual movements can cooperate precisely.

Turning to the exhibition area, a fully replicated Lianhua supermarket is likewise full of challenges. Shelves and aisle spacing are arranged according to the real environment of an offline store. Canned, boxed, and bagged goods come in different forms, and products with similar packaging are more likely to make robots "misidentify goods."

Robots need to simulate nighttime store receiving, picking goods from turnover boxes, taking inventory, scanning barcodes, and checking whether products are damaged; when facing concentrated restocking before and after business hours and dynamic restocking triggered by the system, they must identify products of different specifications and place goods accurately back on shelves. The venue also features mannequins simulating customers, testing whether robots can proactively avoid them and ensure safety while moving. In addition, patrolling the floor and arranging goods, returning products misplaced by customers, and checking misplaced magnetic price tags are all included in the test questions.

From receiving and restocking to shelf organization, a supermarket employee's entire set of daily work has been broken down into practical test questions. Robots must see accurately, grasp steadily, understand where products should be placed, and complete work in an environment with people, obstacles, and tasks that change at any time. The complexity of real scenarios is hidden in these seemingly inconspicuous details.

The elderly care and rehabilitation track brings the test to the service stage. Robots need to sequentially complete greetings and communication, item delivery, room patrol and inspection, rehabilitation training, and vital sign monitoring, running through a coherent service process.

Pre-competition debugging is also not always smooth. Participating team members said that during debugging, robots occasionally wandered around on their own before completing tasks. Team members could only repeatedly adjust behavior programming to help them better understand task order and adapt to specific environments.

This small episode reveals a key challenge in implementing embodied intelligence: robots must not only know how to perform an action, but also know when to do it, where to do it, and what to do next after finishing. Especially in elderly care service scenarios, item delivery, communication, rehabilitation, and inspection are connected to one another, and a breakdown in any step may affect the entire service. What is being compared is not just rankings, but the "connection" between technology and demand. From the relatively broad action demonstrations last year to this year's more detailed task flows and assessment requirements, the embodied intelligence skills competition is pushing "capability demonstration" toward "practical work capability testing." Zhong Junhao introduced that this year, the national level further clarified robot application scenarios and competition requirements, forming 9 major scenarios and 23 competition events. More than 1,200 real business scenario demands converge here, giving robot R&D teams the opportunity to discover problems and refine products in an environment close to actual work, and also enabling scenario demand-side parties to more intuitively judge which technologies are truly suitable for their business processes. For robot companies, the competition is not just about vying for rankings; it is more like a concentrated "health check." Problems such as insufficiently accurate recognition, insufficiently stable grasping, and insufficiently smooth task transitions may all be exposed in practical operations and become the direction for the next round of R&D iteration. For scenario-side parties, real tasks also provide an opportunity to test technologies and help them judge how far robots are from actual deployment. After all, real production lines, laboratories, and elderly care service sites are often difficult to open to R&D teams for repeated trial and error over the long term—it may affect normal operations and also involves the safety of equipment and personnel. Letting robots first undergo testing in a controllable real-scenario competition venue is precisely a way to connect technology R&D with practical application. Whether a robot can work ultimately does not depend on how nimble it is on an exhibition stand, but on whether, after entering the real world, it can do things correctly, do them well, and do them stably and continuously.