Zhiyuan Genie G2 Works 24-Hour Consecutive Shifts at a Convenience Store; Two Robots Fail to Pass Probation

Add crowdsourced testing promotional copy Add crowdsourced testing promotional copy end Add App-exclusive traffic routing

Published by Machine Heart

Starting at 10 a.m. on November 2, in a normally operating convenience store in Xuhui District, Shanghai, the store's employees were replaced by six AgiBot Genie G2 humanoid robots.

Each had its own role: Ahao handled greeting and customer inquiries, Huobao took charge of fetching and restocking goods, Juanwang patrolled the shelves, Kashao was the dedicated barista, Dingjiang handled checkout and prepared food service, and Lengjie was responsible for table cleaning.

During the livestream, the store operated as usual and customers entered as usual, while the robots worked continuously for 24 hours, covering everything from the lunch peak to the evening customer flow, and from late-night duty to the next morning's peak, with all four shifts covered without exception.

The assessment criteria were fully modeled on a real store: on-duty hours, business completion volume, customer evaluations, plus task success rate, response time, and exception-handling capability together formed the conversion-to-regular-employee score.

Overall, the robots were still able to handle tricky tasks with ease, though there were occasional slip-ups. In the end, four robots successfully converted to regular status, while Lengjie and Juanwang were not yet converted due to minor flaws in on-duty enthusiasm and job stability.

In an industry context accustomed to "perfect endings," this unvarnished result is somewhat rare. The gap between robot demonstrations and deployment has rarely been publicly measured before. The convenience store's 24-hour livestream was precisely a public stress test in which AgiBot placed its complete technology stack in a real operating environment and subjected it to supervision by the entire internet.

Perception, manipulation, and scheduling challenges in a real store

Evaluation of embodied intelligence has long faced a structural dilemma: laboratories are controllable, the real world is not; and most evaluation environments are "semi-real" — lighting is controllable, positions can be calibrated, and processes can be preset, creating systematic deviations from deployment environments.

The convenience store happens to be one of the most open and severely underestimated scenarios in urban commerce. From a technical perspective, a real store "turns up the intensity" on robots in at least three dimensions.

First, fine manipulation of flexible objects. Soft-packaged goods deform during grasping, requiring the system to adjust control parameters in real time based on force and tactile feedback; gripping easily damaged objects such as golden cake places even higher demands on force-control precision — too little force and it slips, too much and it crumbles.

Second, planning and coupling of long-horizon tasks. Take restocking as an example: the complete process involves multiple subtasks such as target recognition, inventory judgment, path planning, and state maintenance, with strict dependencies among subtasks: before fetching goods, operational feasibility must be assessed, and if the target is occluded, obstacle clearing must be performed first. When multiple tasks run in parallel, the system must also maintain long-term state memory.

Third, obstacle avoidance in open dynamic environments. In the store, the movement of customers and pets is unpredictable, and the system must complete real-time perception and obstacle-avoidance decisions under strict safety constraints, actively slowing down or even halting actions when necessary.

In contrast, lighting, object poses, and work processes in laboratory environments can all be pre-calibrated and belong to in-distribution conditions; real stores continuously generate out-of-distribution samples — one successful grasp does not mean the same action can be reproduced under new spatial conditions.

The livestream also featured a patrol segment with specially invited experience officers, whose on-site task was just one thing: to turn up the intensity on the robots. For example, having the robots identify and pick out specified products from among sugary and sugar-free cola; temporarily placing people and boxes to block their routes; and deliberately shaking boxes while the robots were gripping prepared food.

During the livestream, the robots complained by voice that "work is so hard" while completing a series of operations: identifying and fetching goods, detouring around obstacles, and adjusting movements amid shaking to complete grasps.

Building a capability iteration loop after deployment

During the 24-hour operation, the robots were not always on target, and there were occasional on-site slip-ups.

When evaluating embodied intelligence systems, the industry's customary metrics are hard indicators such as success rate and cycle time. But in open environments, the more important metric is resilience — after an error occurs, can the system autonomously attribute the cause, autonomously adjust, and turn the lesson into reusable experience?

AgiBot's exploration is an embodied Agent framework called Genie Master. According to information previously released by AgiBot, it is the shared "brain" of the six robots: it is responsible for task scheduling and also connects execution results with subsequent learning. Its core embodied execution protocol links "judgment — execution — feedback" into a closed loop — the Agent knows each robot's capabilities and provides execution guidance, and execution results are fed back in real time to become the basis for the next round of judgment.

This allows robots to locate the causes of failure and adjust strategies themselves without relying on manual demonstration and debugging. In preliminary tests, the Genie G2 equipped with Genie Master improved process success rates on tasks such as prepared-food packaging, fetching, and restocking from 28.6% on the first day to 85.7% by the fifth day, and trended toward stability.

English text is being prepared.

Working together, these two paths make each round of operation provide feedback for the next round of learning — that is, "working on the job while unlocking new skills."

One direct change this brings is: the more robots there are and the longer they are on duty, the more experience accumulates, and the faster improvement becomes.

Raise the perspective a bit: for the large-scale deployment of embodied intelligence, this solves one of the most practical bottlenecks of embodied intelligence. Relying on manual teleoperation and engineer debugging as a fallback means that the more robots are deployed, the higher the labor cost, and the business model becomes hard to sustain; the approach represented by Genie Master is to hand these tasks to the robots themselves — execution results automatically flow back, and strategies adjust automatically.

At least judging from the convenience store data, this path has already initially closed the loop.

Full-stack coordination is the prerequisite for Physical AI deployment

If the tasks in the convenience store are broken down layer by layer, one finds that no single problem can be solved by relying on just one strong suit.

Accurate grasping depends on the robot's own performance — joint precision, force control capability, hardware consistency; smooth motion depends on the coordination of strategy and control; reusable experience depends on a complete data recording and evaluation system.

Breakthroughs in single-point technologies cannot replace full-stack coordination — this is one of the industry characteristics of embodied intelligence.

This is also the main line of AgiBot's releases over the past ten days. Since the launch of AI WAVE on October 22, AgiBot has successively rolled out seven achievements: the fourth phase of the AGIBOT WORLD 2026 dataset, the operation platform Genie Studio 2.0, the heavy-load robot G2 MAX, the embodied Agent framework Genie Master, the simulation platform Genie Sim 3.5, the foundation model GO-3 Preview, and collective experience learning (CEL).

The various links of embodied intelligence are highly coupled: the stability of the robot itself determines data quality; data runs through pre-training, post-training, and continuous learning after deployment; and model deployment also depends on system integration, process adaptation, and equipment coordination — any weak link will hold back the entire loop.

The 24-hour live broadcast from the convenience store, throwing the complete technology stack into a real store and subjecting it to supervision by the entire internet, is itself a public stress test of this choice.

Two robots failed to become regular employees, which also allowed this test to retain the complexity that a real scenario should have. Only when successes and shortcomings are presented together is there a chance to see clearly how far the robots still are from the job requirements.

The significance of public testing lies precisely in letting technology undergo the test of actual business and further improve.

The robots' 24-hour live broadcast in the convenience store has ended, but the robots driven by real scenarios have only just begun.article_adlist[

For reprinting, please contact this official account for authorization

Submissions or reporting requests: liyazhou@jiqizhixin.com ]article_adlist Statement: This is an exclusive article of Sina.com; unauthorized reproduction is prohibited. tech_web_article_v2015_block_bottom non-targeted 300*250 button 17/09 wenjing begin

non-targeted 300*250 button end