6 AgiBot Genie G2 units work back-to-back 24 hours at a convenience store, 4 pass the retention assessment

The robotic arm steadily lifted the freshly brewed hot coffee, picked up the just-steamed buns and smoothly placed them into a paper bag, then turned and moved through the aisles to check for gaps on the shelves and wipe up water stains on the tables... At 10 a.m. yesterday, the 24-hour convenience store livestream challenge "Robot Workers Can't Stop" by AgiBot came to a close. In this assessment, broadcast publicly to the entire internet, six AgiBot Genie G2 humanoid robots were assigned to the daily shift of a normally operating convenience store in Xuhui District, staying on duty from the early morning rush hour all the way to running the store independently late at night, working around the clock for 24 hours. Over the full day they handled a total of 1,730 miscellaneous tasks, including 683 customer inquiries, 369 item deliveries and pickups, 147 freshly made coffees, 126 servings of hot breakfast items packed, as well as 277 shelf inspection and restocking rounds and 128 counter cleanings. After all-weather, multi-process, high-intensity operation, four robots ultimately passed the real-world test and the online evaluation by public judges, successfully earning retention qualifications. Photo by Xinmin Evening News reporter Ma Dan. What was tested was on-site comprehensive adaptability. Although an ordinary convenience store occupies little space and its processes seem routine, for humanoid equipment the uncontrollable factors here are far more complex than in an orderly factory workshop. The retention assessment indicators included how long the robot stayed on duty, how much business it completed, how customers evaluated it, and whether tasks succeeded, responses were timely, and abnormalities were handled properly. Two specially invited experience officers also came to test them with "tricky questions": placing regular and sugar-free cola with nearly identical appearances in front of the robot and issuing the instruction "bring the sugar-free cola" to see whether it would pick the wrong one; they also completely disordered a chewing gum shelf and had the robot put each box back in place... "The assembly line in a factory has a clear rhythm, but a convenience store is a small society that changes all the time," Ren Guanghui, vice president of AgiBot's Embodied Research Center, told the Xinmin Evening News reporter. Product packaging in the store ranges from soft to hard and varies in weight, and items may be misplaced or disordered by customers at any time; there are also always people walking and lingering in the aisles. Several devices squeeze into a small space and divide the work, appearing to be busy separately while in fact each link is tied to the next. As long as one machine pauses too long in the aisle because of obstacle avoidance, not only will the cleaning route behind it be cut off, but the delivery of goods ahead will also be delayed accordingly, placing extremely high demands on the agility with which the devices coordinate with one another. Many "work" steps test the robots' "fine movements." When the mechanical gripper picks up soft food, a slightly stronger grip can deform it, while a slightly lighter grip can cause it to slip and fall midway; the device must sense the rebound force at the moment of contact and then steadily place it into the packaging bag. When encountering drinks or snacks that have been casually moved to another position, the robot cannot rigidly apply the original shelf coordinates; it must rely on visual scanning to recognize the items again and put them back one by one in the correct slots. Once a customer suddenly walks toward it or the aisle is temporarily blocked by boxes, the robot's underlying logic is to slow down and give priority to avoidance, then detour after confirming the passage is safe. For the first time, robot workers have a "persona." They can chat and also complain; in this real-world assessment, the distinctive human-robot interactions also broke people's traditional impression of machines as cold and rigid. The reporter found on site that AgiBot for the first time set clear personality labels and MBTI personas for the six Genie G2 robots participating in convenience store work. "Ahao," responsible for greeting, was set as an ENTP with a "social butterfly" attribute; upon entering, it would deliver a skilled set of warm-up moves and even actively ask customers for "votes for regularization." "Kashao," serving as "barista," combines ESFP traits and does not forget to offer a greeting when serving drinks. "Lengjie," responsible for cleaning, is a rigorous and earnest ISTJ "cyber hygiene commissioner," muttering softly while wiping the table. There is also "Huobao," an ESTJ responsible for picking up goods and specializing in "ruthless brick-moving," the ISFP "Juanwang," who slowly and methodically patrols the shelves, and the ESTP "Dingjiang," who manages the breakfast steamer. The originally monotonous mechanical work was endowed with a service experience of distinct personalities. Ren Guanghui introduced that injecting personality into robots is not simply for entertainment gimmicks, but a new attempt at the boundaries of embodied intelligence interaction. In the future, such personality settings and interaction modalities will be precisely adapted according to specific work scenarios. For example, in serious industrial production and hospital guide scenarios, robots will maintain a calm, rigorous, and concise interaction style; while in consumer service scenarios such as supermarkets and convenience stores, cultural tourism parks, and elderly care companionship, they can autonomously switch to humorous, patient, or more approachable communication modes, allowing embodied intelligence to truly integrate into complex daily interpersonal environments while actually "doing work." On-site failures were included in a learning "error collection." In a real operating store, robots have to face a wide variety of emergencies and inevitably have moments of slipping or getting stuck. These failure scenes were also shown in the 24-hour livestream. Under the lenses of netizens across the internet acting as real-time cloud supervisors, the most representative failure scenes were mainly concentrated in flexible force control, visual anti-shake, and human-robot interaction logic, which was also the direct reason why only four of the six robots ultimately passed the regularization assessment. In front of the cooked food counter, soft food first posed a problem for the mechanical hand: the skin of freshly steamed pastries is soft, and a slightly stronger gripping force would pierce the bun skin and crush the dim sum; with slightly reduced force, it would slip from the hand midway through bagging. The cleaning post also showed judgment deviations. Under overhead lighting, the robot mistook the wood-grain shadow on the table for spilled coffee stains and wiped the same clean table back and forth, which netizens described as "slacking off." While on night duty, the cleaning "Lengjie" was single-mindedly looking for the trash can, and while searching walked out of the store door; it was brought back by a technician, and it had also thrown a customer's food away as trash. In the end, "Lengjie" failed to be regularized. Photo by Xinmin Evening News reporter Ma Dan. Ren Guanghui said that this is precisely one of the most valuable data collected under real operating conditions. These failure cases will be automatically imported into the backend "error memo" as key error samples for algorithm iteration. The reporter learned that all on-site operating models are mounted under AgiBot's "Genie Master" agent framework. This central network is responsible for issuing instructions to each workstation and closely monitoring feedback from every action to help the algorithm locate problem nodes. Ren Guanghui said: "With this adaptive mechanism of learning while working, fine-tuning in process scheduling can usually be adjusted in place the next day; deeper optimization involving underlying fine force control and perception algorithms can also be completed within one to two weeks. At the same time, swarm learning technology allows experience to flow and be shared among machines; a grasping trick figured out by one device can be synchronously pushed to other models of the same type." However, Ren Guanghui also said that going from a single store to hundreds or thousands of stores still requires solving issues such as standardization of models and tools, integration of merchandise and business systems, rapid deployment, remote operations and maintenance, and how experience can be reused across different stores. Since the beginning of this year, AgiBot has had robots densely deployed in factory workshops, cultural tourism parks, and street-front shops, and at the World Internet Conference Wuzhen Summit it won the "Leading Science and Technology Award" for technological breakthroughs in the field of embodied intelligence, pushing embodied intelligence to continue maturing through real work and trial and error.