Over 1,200 organizations gather in Zhangjiang, Shanghai as embodied intelligence skills competition kicks off in real-world scenarios

Walking into the Zhangjiang Science Hall, a 1:1 replica of a State Grid ultra-high-voltage substation, supermarket shelves built to real store standards, and a strictly replicated quality inspection laboratory are lined up in a row. There is no flashy light show, only the busy figures of robots climbing over complex terrain, performing delicate grasps in front of test tube racks, and sorting goods among cluttered shelves. The three-day 2026 Global Developer Pioneer Conference and International Embodied Intelligent Skills Competition opens tomorrow. Reporters visited in advance today to experience the training atmosphere of robots in one real scenario after another. After this competition concludes, the 2,000-square-meter venue of the Smart Bay Exhibition Hall at the Zhangjiang Science Hall will be converted into a permanent embodied intelligence scenario training space, where robots will accelerate their "evolution." A 1:1 restored "high-voltage" test Since the beginning of this year, major embodied intelligence leading companies have successively announced entry into a "deployment state," and real scenarios have become the most core resource element in this competition. In June this year, the Ministry of Industry and Information Technology and the State-owned Assets Supervision and Administration Commission of the State Council jointly launched a special action for real-scenario training of humanoid robots and embodied intelligence, using real scenario units such as production workstations and service operation points as training carriers to accumulate high-quality real-machine data and build spaces that can be trained, tested, and verified. This competition is one such test. "The biggest difference between this year's competition and last year's is that we no longer artificially set up 'taken-for-granted' scenarios, but put robots directly into real working-condition scenarios and assembly lines, down to specific questions such as how much work efficiency is improved, how stable operation is, and whether input-output is cost-effective. The measured data can be viewed intuitively on site," said Zhong Junhao, secretary-general of the Shanghai Artificial Intelligence Industry Association. After preliminary solicitation, the competition site gathered more than 1,200 body units and scenario units from across the country. Complete-machine and humanoid robot companies such as AgiBot, Unitree, the National and Local Center, Fourier, Lenovo, Keenon, JAKA, Kepler, Robo Party, Xuqing Intelligent, and Chenhunxian, together with core component manufacturers and scenario application enterprises, made a concentrated appearance. The competition has set up 9 tracks and 23 events across the three major fields of industry, service, and special operations. Energy inspection is the most popular scenario, with more than 10 teams having obtained competition qualifications; they will compete in the real operating environment of the Ultra-High Voltage Branch of State Grid Shanghai Municipal Electric Power Company. Zhong Junhao revealed that last year's competition inspection scenario also included roadblocks, climbing, and other elements, but they were all set arbitrarily; this year, a 1:1 scenario restoration has been achieved. According to the competition rules, robots must complete tasks within 30 minutes in sequence, including passing through complex terrain, reading and reporting instruments, infrared temperature measurement, equipment anomaly identification, switch status verification, assisting simulated switching operations, and monitoring personnel and environmental safety. They must also challenge optional actions such as passing through a 50-centimeter narrow passage and crawling through a 60-centimeter low area. One detail is that the robot needs to identify whether workers are wearing work clothes and safety helmets. If it were not real-machine training, this point would not be listed as an assessment item. A "pick-and-place" contest within a square inch In the "Testing and Analysis - Quality Inspection" event, the full-process testing of milk tablet samples by the Shanghai Institute of Quality Inspection and Technical Research is condensed into a small space. Event manager Li Jiayuan introduced that this scenario highly restores a real quality inspection laboratory. The competition is limited to 30 minutes, and the robot needs to continuously complete six basic actions: "receiving and sampling, weighing and preparation, centrifuge loading, removal and return, status inspection, and judgment and sorting." "Manual picking is very accurate, but robots are not accurate enough," Li Jiayuan said, noting that the biggest challenge lies in fine movements. For example, the centrifuge needs to operate under real-environment settings (1,000 revolutions per minute for 50 minutes), but during the competition the robot is required to open it and take items out after 5 minutes; the lid is very heavy and very easy to knock over. The optional innovative action segment even requires the robot to use a 3-milliliter syringe to extract liquid from the test tube after centrifugation, filter it, put it into a 5-milliliter small test tube, and put on the cap. These operations will test the model algorithm and the software-hardware integration capability of dexterous hands. This also tests the dexterity of robot hands. Last year it was "flower arranging," and this year clear action indicators have been broken down, which shows the speed of embodied intelligence development over the past year. However, what outsiders see as rapid development has actually been slowly "polished" bit by bit with "clumsy efforts." Zhu Mianrui, data collection team leader at Mifeng Technology, said: "Even for simple small actions like picking up a bottle or placing an object, we need to repeatedly 'teach' hundreds of times before the robot can gradually become proficient." The "generalization" test of complex scenarios Today, training grounds for embodied intelligence are no longer a scarce resource. A report by the China Academy of Information and Communications Technology shows that as of the end of June this year, more than 70 embodied intelligence training grounds have been built and put into use nationwide, with another 46 under construction or planned. However, the difficulty lies between "training" and "use": scenario tasks are homogeneous, data production capacity is limited, circulation is insufficient, and a sustainable business model has not yet taken shape. The industry is moving from a construction boom toward a period of capability accumulation. In the retail warehousing event of the competition, Bailian Group brought its real stores and warehouses to the site. Guo Yongping, a staff member of Bailian Group, said that the shelf spacing, lighting interference, and product complexity here all benchmark real stores. The robot needs to complete five tasks within the specified time: identifying and taking inventory of turnover-box goods, shelf display replenishment, price tag and product verification and replacement, order picking for multiple stock-keeping units, and transfer of tasting service items. Only in real scenarios will problems be exposed. "Many teams had previously trained their own small models, but when they reached real supermarket scenarios, many of them could not be used," Guo Yongping said. This is also the pain point encountered after current robots "take up their posts"—weak generalization ability. This competition is dedicated to solving this common problem. Zhong Junhao said: "After the on-site competition ends, we will also move the real scenarios online and let more teams participate in simulation competitions, hoping to produce a batch of 'atomic-level skill packages.'"They will be put into different real scenarios again for iteration, striving to form 'plug-and-play' generalization capabilities before next year's World Artificial Intelligence Conference.