AGIBOT Sends Six G2 Robots Through a 24-Hour Convenience Store Stress Test

Robots That Don't Clock Out

Over the first weekend of November 2026, six AGIBOT G2 wheeled humanoid robots reported for duty at a 24-hour convenience store in Shanghai's Xuhui district. They worked four continuous shifts — handling customer greetings, preparing coffee, retrieving products, restocking shelves, and cleaning tables — without clocking out. When the 24-hour shift ended, four robots had their employment contracts extended. Two did not.

AGIBOT called it a probation test. The setup closely resembled something any retail manager would recognize: new hires assigned to real tasks, evaluated on performance, and sorted into those who cut it and those who don't. The difference is that the two dismissed workers were robots, and the AI model running their decision-making had been unveiled just four days earlier.

The event marked the first extended public deployment of GO-3 Preview, AGIBOT's new embodied AI model announced October 29, 2026 — and an early test of whether Chinese humanoid robots can hold up under the conditions of commercial retail, not just laboratory demonstrations.

A Six-Robot Crew, Four Shifts

The six robots completed 1,730 discrete tasks across four shifts: 683 customer greetings, 369 product retrievals and deliveries, 147 coffee and hot meal preparations, 277 shelf inspections and restocking rounds, and 128 table cleanings. Each robot was assigned a specific role and given a name matching its function.

Yingjie managed customer greetings. Wenwen handled coffee preparation and hot food. Duoduo focused on product retrieval. Tangzai was assigned restocking. Juanwang covered shelf patrol. Lengjie was responsible for floor and table cleaning.

At the end of the deployment, AGIBOT evaluated each robot against a defined set of probation criteria: task completion rate, response time, and the ability to handle unfamiliar situations without human intervention.

Why Two Robots Failed

Four of the six passed. The two that did not were Juanwang and Lengjie — and their failures pointed to different gaps in autonomous robot behavior.

Juanwang, the shelf patrol robot, was dismissed for a lack of initiative. It completed tasks when directed but did not proactively identify and begin shelf inspections or restocking without prompting. In a retail environment designed to run with minimal human oversight, waiting for explicit instruction was disqualifying.

Lengjie, the cleaning robot, failed for a different reason: it navigated off its assigned route and, on at least one occasion, wandered outside the store entirely. Spatial discipline — staying on assigned routes, maintaining store-boundary awareness — is a core operational requirement for autonomous floor robots.

The passing robots also faced challenges. Product retrieval presented difficulties when items were soft or irregularly packaged, and goods hanging from display hooks required precise manipulation the robots found demanding. Those are known manipulation challenges in humanoid robotics, not unique to this deployment. The difference between the four that passed and the two that did not came down to initiative and navigation, not dexterity.

The AI System Running the Store

The robots operated on GO-3 Preview, AGIBOT's latest embodied foundation model, announced four days before the retail event. Its underlying architecture is called WRAM — World Reasoning Action Model — and it integrates three components: CoAE (a control-oriented autoencoder that processes raw visual input into compact robot-ready feature representations), SVP (a single-step future-state prediction module that generates candidate actions), and KASO (a selection module that filters those predictions against trajectories observed during training).

The system is intended to give robots the ability to reason about the physical world before acting, rather than relying on pre-scripted task sequences.

Coordinating the robots' behavior inside the store was Genie Master, AGIBOT's embodied agent framework. It connects planning, execution, real-time feedback, and experience retention through a protocol called EEP — Embodied Execution Protocol — that enables robots to build on past task attempts rather than starting each new situation from scratch.

Underlying the retail event was a capability AGIBOT calls Genie Evolver — a distributed post-training system that uses approximately 100 physical robots alongside physics simulation and world models to continuously improve robot behavior. It runs a process called Collective Experience Learning, in which experience gathered across the robot fleet is shared system-wide. AGIBOT describes the result as Physical RSI — recursive self-improvement in the physical world — meaning robots can improve autonomously without requiring new human demonstrations or manual recalibration.

According to Interesting Engineering, GO-3 Preview's task success rate during the convenience store learning demonstration rose from 28.6 percent on the first day to 85.7 percent by the fifth day of deployment, without additional human training.

GO-3 Preview builds on design lineage established by GO-2, AGIBOT's April 2026 model, which introduced Action Chain-of-Thought reasoning and was accepted at CVPR 2026. AGIBOT said it plans to open-source GO-3 Preview by the end of 2026.

Hardware Purpose-Built for Retail

The AGIBOT G2 was launched in October 2025 as an industrial-grade humanoid designed for physically demanding environments. It weighs 185 kilograms, stands on a wheeled base, and uses seven degrees of freedom per arm in a force-controlled configuration — 26 active degrees of freedom in total. The system is IP42-rated for dust and moisture resistance, which matters for environments like convenience stores with wet spills and food preparation areas.

It runs on dual hot-swappable battery packs and is powered by an NVIDIA Jetson Thor T5000 computing module with 2,070 TFLOPS of FP4 compute — a processing capacity that enables the kind of real-time vision-based task reasoning required by WRAM. The list price is approximately $138,000 USD.

Commercial Scale and Competitive Position

The retail event did not take place in isolation. AGIBOT delivered its 10,000th G2 robot in March 2026, its 15,000th in June, and its 20,000th in September. The company's production partner Longcheer reported a yield rate of 99.99 percent at a volume of 64,828 units.

Counterpoint Research estimated that AGIBOT shipped approximately 9,700 humanoid robots in the first half of 2026 — representing more than 43 percent of a global market that totaled roughly 22,000 units during the period, a figure Counterpoint described as nearly 300 percent higher than the same period in 2025. IDC separately estimated approximately 8,600 AGIBOT units out of a global total exceeding 25,000, reflecting differences in analyst methodology and counting conventions. Both figures placed AGIBOT as the global market leader.

AGIBOT is a Shanghai-based company whose operations are subject to Chinese law, including regulations that may require disclosure of certain data to Chinese government authorities. The company has not publicly disclosed the specific data handling and privacy practices governing its robots' operational data.

A Competitive Retail Landscape in Shanghai

AGIBOT is not the only company deploying humanoid robots in Shanghai-area retail. Galbot's wheeled humanoid robot — known as "Xiao Gai" in mainland China deployments and "Little Gal" in Hong Kong — has been operating at a convenience store in Shanghai's Lujiazui district since January 2026. SenseTime's subsidiary SenseTime Shanhui has deployed its robot-staffed store brand, SenseMartGo, with plans to reach 100 locations by the end of 2026.

The competitive clustering is not accidental. Shanghai has prioritized humanoid robotics as a municipal industrial focus, and several companies are using its dense urban retail environment as a live deployment testbed — measuring real-world robot performance against metrics that factory floors don't provide.

What Probation Actually Tests

The framing of a "probation" event is deliberate. AGIBOT is competing not just on hardware specifications or benchmark scores but on a demonstration that its robots can be evaluated by the same criteria a retail employer would use for a human worker.

The 1,730-task completion record and the dismissal of two underperforming units are designed to communicate operational credibility — that robot deployment at commercial scale can be managed, measured, and improved through the same feedback loops that retail operators already understand.

Whether the approach translates into sustained commercial contracts, and whether four passing robots in one Xuhui store represents a threshold crossed or a carefully staged milestone, will become clearer as the deployment expands.