Six AgiBot Genie G2 robots enter a Shanghai convenience store, taking on a 24-hour real-world work challenge

English text is being prepared.

At 10:00 on November 2, in an operating convenience store in Xuhui, Shanghai, six special "new employees" clocked in for work.

They are six AgiBot Genie G2 humanoid robots, working in this 24-hour convenience store for a full day in roles such as greeter, barista, stocker, and cleaner—taking orders, making drinks, restocking, and cleaning—carrying out the "Robots Can't Stop Working" challenge and accepting live online supervision from viewers across the internet.

Since AI WAVE launched on October 22, AgiBot has successively released multiple technology results in areas including data, simulation, models, bodies, and operating systems. Today, these results converge into a single test: in a real operating convenience store, they support the robots in completing work and maintaining stable output over 24 consecutive hours. The large-scale application of embodied intelligence begins precisely with practice in one role and the accumulation of experience again and again. Six "new employees" work four shifts around the clock as tricky tests come one after another.

The six robot "new employees" wear badges showing their respective roles: Ah Hao stands at the entrance, responsible for greeting and customer inquiries; Huobao and Juanwang are a stocking pair, one fetching and restocking goods while the other patrols the shelves, sustaining the store's product flow; Kashao is the dedicated "barista" at the coffee station; Dingjiang guards the checkout counter and also handles prepared food service; Lengjie is responsible for cleaning tables throughout the store. The six roles are interlinked, covering the complete service flow of a convenience store from opening and welcoming customers to daily cleaning.

Genie Master summarizes successful and failed work trajectories every day and stores the relevant experience in long-term memory, directly calling on it when encountering similar tasks; effective adjustment processes then form training data that feeds back into the continuous optimization of action policies. The two paths of "remembering experience" and "training skills" work together, so that each round of work provides feedback for the next round of learning, and the robots thereby "learn new skills while on the job."

Today, two "examiners" have already arrived with test questions. Special experience officers Xiaoyang biubiu and Xiao Xiao Nathan completed two rounds of store patrols, specifically setting "questions" in creative ways from the customer's perspective: regular and sugar-free cola look almost identical, so with one sentence, "get the sugar-free one," they see whether it picks the wrong one; they mess up the entire chewing gum shelf to see whether it can put each box back in place; they stand in the middle of the aisle as human roadblocks and then throw a box down to block the way, to see whether it goes around or crashes straight into it. When pushed hard, the robot even shot back, "You've got some nerve, chubby."

However, while voicing complaints like "work is so hard," the robots do not forget to keep taking on the challenges: for cola, they act only after dual verification of "ears understanding and eyes confirming"; for messed-up shelves, they do not rely on labels but use visual inventory to put items back cell by cell; when encountering people, they stop first and then go around, and if a box blocks the way, they go around again. A cup of coffee, a restock: small actions, but highly difficult.

The area is not large and the processes are standardized, making the convenience store one of the most common commercial scenarios in the city. But for robots, being competent in this work is not easy.

For example, to hang a bag of Skittles on a hook, if the angle is slightly off, the robot may clamp other products; to pick up a piece of golden cake, a tiny difference in force is the difference between "dropped it" and "crushed it."

In the laboratory, lighting, item positions, and operating procedures can all be set in advance; in a real store, customers move around at any time, and the positions, orientations, and packaging states of products change constantly. The same "take an item" task may present different conditions each time; one successful grasp does not mean the next one can be repeated.

Facing these test questions, the robots' overall performance did not falter: for products customers casually put in the wrong place, they do not rely on label coordinates but use real-time visual inventory to find where items belong and return them cell by cell; when the restocking robot finds a gap on the shelf, it automatically triggers the restocking process without human reminders; for stains on the counter, it can distinguish object patterns from spill marks and prioritizes using tools to clean them. Failures are recorded in an "error notebook," allowing the robots to continuously self-evolve.

During real store operations, the robots are not perfect either and occasionally slip up while working. But a slip does not mean failure—in real scenarios, humans make mistakes too, and the key is what happens after a mistake: whether the problem can be identified, corrected, and the lesson turned into experience.

Therefore, the standard for judging whether these "new employees" are qualified is not a single success or failure, but whether they can "correct their mistakes when they know them." This is precisely the starting point of Genie Master, the embodied Agent framework AgiBot previously launched: as the shared Agent "brain" of the six robots, it is responsible both for scheduling tasks and for connecting execution results with subsequent learning.

“Correcting mistakes once recognized” first depends on a working execution loop. Through the Embodied Execution Protocol (EEP), Genie Master connects task judgment, action execution, and result feedback. The Agent organizes tasks based on the robot's capabilities and on-site conditions; the action policy combines real-time observations to complete operations; and the execution results are returned to the Agent, providing the basis for subsequent judgment and exception handling.

But on-the-spot correction is not enough; more critical is turning newly acquired skills into reusable experience. Genie Master summarizes successful and failed operation trajectories every day and stores the relevant experience in long-term memory, calling on it directly when similar tasks arise; effective adjustment processes then become training data that feeds back into the continuous optimization of the action policy. The two paths of “remembering experience” and “practicing skills” work together, so that each round of work provides feedback for the next round of learning, and the robot thereby “learns new skills on the job while working.”

This mechanism has already begun to show results in stores: over the past period, Genie G2 has continued to improve on multiple complex and delicate tasks; late at night when idle, the robot even actively puts in extra practice. The next stop for AI WAVE is more real-world scenarios.

Pulling the camera back, the robot's real-world work in the convenience store is a microcosm of the full-stack technology release of AgiBot AI WAVE.

From the series of technology releases to six Genie G2 robots entering the convenience store, AgiBot is putting the capabilities of full-stack embodied AI into specific business tasks for testing. The value of data, models, the simulation platform, the robot body, and the operation platform thus comes down to practical questions: whether service can be completed, whether exceptions can be handled, and whether experience can be reused.

The next stop for AI WAVE is more real-world scenarios. Next, AgiBot will continue to connect technology with real applications, so that robots can be tested in their posts and continuously improve through work. As this set of capabilities is validated and accumulated across more tasks, the same experience will be replicated to more positions and more scenarios.

At present, in this exam at the convenience store, the robot “new employees” have not yet handed in their papers. After the late-night shift and the morning rush shift are over, the final results of the six AgiBot Genie G2 robots will be officially revealed. Can they successfully become regular employees? Please keep following the livestream and witness this on-the-job exam together.

Source: contributed article

The images in this article are from: 123RF licensed image library. Infringement report Return to sharing Related articles All comments