Gasgoo Munich- At 10 a.m. on November 2, a convenience store in Shanghai's Xuhui district welcomed several unusual "employees"—six AgiBot Genie G2 humanoid robots.
Over the next 24 hours, the six robots would work continuous shifts as greeters, baristas, clerks, and cleaners. They handled ordering, preparation, restocking, and cleaning—streamed live and scrutinized by the entire internet. AgiBot claims this marks the industry's first validation of Physical RSI in a real-world business scenario.
In embodied AI, putting robots in convenience stores isn't new. What's novel is the test itself: six robots dropped into a fully operational store, with real customers, real orders, and unpredictable surprises—no dodging any of it.
Past industry demos mostly asked "can it?": grasping a cup, walking, folding clothes. The tasks were pre-set, the environments staged.
A real store tests something else entirely: can a robot sustain a full workflow in a shifting environment, recover from glitches, and transfer its skills to a different location?
That is the question AgiBot set out to answer.
Image source: AgiBot
Why a Convenience Store?
Ren Guanghui, vice president of AgiBot's Embodied Intelligence Research Center, summed it up: "Real enough, and hard enough."
Over the past two years, companies have been pushing humanoid robots into factories and logistics sorting. In Ren's view, these settings are relatively fixed—testing rhythm, precision, and reliability. A robot that perfects a single repetitive motion can mostly meet the bar.
A convenience store is different.
"Stores demand adaptability: product variations, flexible handling, human-robot coexistence, customer interaction, multi-robot collaboration—all crammed into a small space where tasks collide," Ren said. "Restocking might block a pickup spot; cleaning could clog an aisle. One delay cascades, pushing back a whole string of deliveries."
This is a world away from a lab demo.
In a demo, lighting, object placement, and workflows are pre-arranged; the robot follows a script. A real store has no script. A customer walking past the camera invalidates prior judgments. One successful grasp doesn't guarantee the next.
Zoom out, and this test matters because it hits on the industry's most awkward reality.
A Counterpoint report shows global humanoid robot shipments topped 22,000 units in the first half of 2026, up nearly 300% year-on-year. Chinese firms swept the top five, claiming a combined 86% share. AgiBot led with roughly 9,700 units—over 43%—while Unitree took second with over 7,000 units, or about 31%.
On shipments alone, Chinese companies have seized a commanding position.
Yet soaring shipments don't mean robots are truly ready for work.
The same report breaks down where those robots went: entertainment and commercial performances, plus education and data collection, accounted for over 60%. Guided tours and retail greeting took about 19%. Those actually entering smart manufacturing and logistics made up just 13% and 5%, respectively.
In other words, most robots sold today are still performing on stage, being studied in classrooms and labs, or greeting visitors at entrances. The share doing heavy lifting on production lines and in warehouses remains low.
Building and selling a robot is one thing; having it reliably create value in a real job is quite another. The gap between the two remains wide.
By making the convenience store the "first exam room" for its full-stack technology, AgiBot is taking aim at that very gap.
"We want to validate not just isolated skills, but whether the system can keep pushing forward amid changes, interruptions, and anomalies—and whether those experiences become reusable knowledge to make future operations more stable," Ren said.
Image source: AgiBot
Specifically, AgiBot deployed six robots across four grueling shifts: the lunch rush, evening peak stress test, overnight watch, and morning sprint. They hit every beat of store operations, and special inspectors were invited to drop by and deliberately make things difficult.
The yardstick was the store's own. "We tracked task success rate, first-attempt success rate, response time, end-to-end latency, and qualified deliveries per unit of time—alongside human takeover frequency, anomaly recovery time, and system availability," Ren said.
After 24 hours of non-stop work, AgiBot disclosed that the six robots completed 1,730 tasks, served 683 customers, and drew over 2 million live viewers.
The breakdown tells a clearer story: "Clerk" Huobao handled 369 pickup-and-delivery orders; "Barista" Kasha made 147 cups of coffee; "Hot Food Manager" Dingjiang served 126 hot meals; Juanwang patrolled and restocked 277 times; and Lengjie cleared tables 128 times.
Ultimately, four robots "passed probation." "Greeter" Ahao proactively welcomed customers with steady, friendly service, earning the top customer rating. Huobao stayed aggressive and fast during order surges, weathering spikes and sudden cancellations without stalling—a stress-test pass. Kasha delivered consistent quality with a pleasant voice. Dingjiang met overall benchmarks, though the occasional dropped corn cob in practice showed its stability still needs work.
Two failed to make the cut. Juanwang lacked initiative, seldom taking proactive action while on duty and failing to match the store's pace. Lengjie's issue was more specific: it drifted off its planned route and wandered out of the store, falling short on positional stability.
Juanwang and Lengjie exposed a common shortcoming in today's humanoid robots: the drive and rhythm to initiate tasks, and the ability to respect boundaries when rules don't explicitly cover a situation.
These two factors sit much closer to the true threshold of deployment than simply knowing "how to make coffee."
How Did AgiBot Pull It Off?
Convenience store tasks sound trivial. For a robot, each one is high-difficulty.
Ren gave two examples. First, picking up a golden sponge cake: the robot must grip the tongs firmly, then control the opening, contact point, and force—too light and it slips, too hard and it crushes. After lifting, it has to align with the bag, drop the cake in, withdraw the tongs, and switch to a different grip to lift the bag.
Second, hooked items: removing a bag of Skittles requires sliding the packaging hole off the hook without snagging the wrapper or knocking over adjacent products.
"Once a soft package deforms under force, the conditions for all subsequent actions change entirely," Ren said.
A human clerk handles these by feel. For robots, it's where they most easily falter, demanding a full stack of coordination: vision identifies the object and position, a motion policy dictates the approach, dexterous hands or grippers execute, and low-level motion control delivers the right force and movement.
If any link in that chain fails, the object slips.
The Genie G2's ability to handle these tough tasks doesn't rely on any single "black tech" breakthrough, but on the synergy of an entire system.
First, the task foundation model GO-3 Preview, which answers where the robot's capabilities come from.
Image source: AgiBot
If every task requires step-by-step human teaching, robots will never escape the lab. GO-3 Preview exists to solve that.
According to Ren, GO-3 Preview uses AgiBot's self-developed WRAM architecture—short for World Reasoning Action Model—which jointly models understanding, prediction, and action. The model learns operational experience across different tasks, forming capabilities that transfer across tasks rather than starting from scratch each time.
"When picking something up, whether it's a cup, a soft package, or a hooked item, the robot chooses its grip and movement path based on the task requirements and the live situation," Ren said.
Around GO-3 Preview, AgiBot also built Genie Evolver—a distributed, hybrid-environment post-training system combining real hardware, physical simulation, and world models.
In practice, each environment plays a role: real hardware provides physical feedback, simulation expands the training scope, and world models deduce potential action outcomes—all working together to hone operational skills. After deployment, Genie Master orchestrates real tasks, feeding execution feedback into subsequent improvements.
Image source: AgiBot
Second, the embodied Agent framework Genie Master, which acts as the "brain" for the entire crew.
Individual competence doesn't guarantee multi-robot coordination. To bridge this, AgiBot built the Genie Master embodied Agent framework, using the Embodied Execution Protocol (EEP) to connect robot capabilities, execution guidance, result feedback, and task continuity.
"In a store, scheduling must juggle order priority, robot status, and workspace," Ren explained. "If two tasks need the same shelf spot, they're staggered; if they don't clash, they run in parallel. If one robot fails or delays, whether another can step in depends on whether it has the right end-effector, the operational skill, and idle capacity—if so, it takes over; if not, the system re-plans, or requests human help if necessary."
Throughout, the system continuously logs successful and failed trajectories to iterate on motion policies.
Ren admitted the early ordering system hit plenty of snags—failed orders, status updates missed after completion, and positional conflicts between robots.
AgiBot tackled this with a tiered mechanism: Harness RSI-level experiences are summarized the same day and take effect in execution the next, consuming almost no training resources. Policy RSI-level updates depend on data accumulation and training schedules, typically iterating within a few days.
For instance, data from this livestream—after filtering, de-identification, and quality checks—enters the training queue and typically shows up in subsequent releases within one to two weeks.
Beyond the tech, all actions must stay within clear safety boundaries—a non-negotiable principle.
In human-robot shared spaces, safety is the inviolable baseline. Translated into action: "Better slow than sorry—never collide."
"In narrow aisles, if a customer suddenly enters the work zone or a child approaches, the system reacts based on distance, motion trends, and current risk—slowing down, pausing, yielding, or replanning," Ren said. "If perception uncertainty spikes, it proactively dials back intensity, stopping and requesting help if needed."
For hot food and coffee—where spill and scald risks loom, and handoff stability is critical—formal deployment designs workflows around the container, grip, travel path, and a fixed handover spot. The robot only releases the item once it confirms the customer is ready to receive it.
In noisy settings, the robot gauges intent by synthesizing speech content, customer orientation, the current person being served, and order context. If speech is unclear, the target ambiguous, or the request would alter items, quantities, or prices, the robot briefly confirms—cutting down on mistaken or duplicate orders.
It was this system that held together 24 hours of continuous operation for six robots.
Still, Ren conceded that robots remain most proficient at operations with standardized objects, defined work zones, and easily verifiable results. Their weak spots are deformable objects, complex contacts, and continuous multi-step operations.
"Compared to a seasoned clerk, robots still lag noticeably—in how fast they adapt to on-site changes, the rhythm of fine manipulation, and anomaly handling."
From One Store to a Thousand: How Far?
Humanoids "working" in convenience stores isn't a new idea, but progress has been slow.
Ren traces the root cause to the gap between a robot's isolated skills and a store's actual operational demands.
"Grasping certain products doesn't equal stably handling diverse packaging, complex orders, customer disruptions, and anomaly recovery. If on-site adaptation and maintenance always rely on engineers, replication costs won't come down."
Never mind scaling from a single store—which brings even more concrete challenges: standardizing hardware and tools, integrating products and business systems, rapid deployment, remote operations, and transferring experience across locations. "Moving into food service and industry also means meeting separate hygiene, process, rhythm, and safety requirements."
That's why AgiBot chose to integrate data, models, hardware, and deployment across the full stack, letting capabilities solidify as a system: a skill learned once can be called up and reused across stores, rather than sitting scattered in isolated functions.
Image source: AgiBot
"Embodied AI components are highly coupled; a short board anywhere can bottleneck the final deployment loop," Ren said. "AgiBot's full-stack commitment is designed to connect these tightly linked pieces, turning technical breakthroughs into complete, stable, replicable operational capability."
Take hooked items again. If a price tag blocks the hook, the issue might start with observation: the system must decide whether to proceed or shift its perspective. The hardware needs to reach the spot, perception must lock onto the target, and the motion policy executes alignment and hanging.
After execution, it checks whether the item is securely hung; if not, how to safely back out and recover. If the same issue recurs, the system loops back to adjust observation rules or supplement motion training.
"These steps involve hardware, perception, models, control, the task system, and data feedback. If any link is broken, a seemingly simple task fails," Ren said.
Full-stack integration sounds simple. Executing it is another test entirely.
Ren noted that priorities must be set around real deployment hurdles—first, issues threatening safety and sustained operation, then those most frequently causing failures, human takeovers, or rhythm slowdowns. "The ultimate metrics are simple: is the deployment cycle shrinking? Is manual adaptation dropping? Is effective output rising?"
Alongside this, AgiBot introduced an evaluation framework for robot labor: Worken (Effective Intelligent Work Hours)—calculated as "active robot count × autonomous work duration × output efficiency ratio"—to measure the actual value robots create.
"Worken answers a very practical question: over how much autonomous work time, how many qualified deliveries did the robot complete?"
The math is blunt: the longer a robot runs autonomously and the less human supervision it needs, the lower the cost per qualified delivery. No flashy demo can replace this calculation.
So, which roles will be the first to scale stably?
Ren offered no timeline. Instead, he laid out three criteria: is the task boundary clear? Is the result easily verifiable? Is the anomaly rate low enough for the system to self-recover?
By these standards, roles with structured environments, standardized objects, and fixed workflows—like sorting, restocking, and cleaning assistance—are most likely to achieve stable deployment first. Jobs involving complex interaction, identity verification, or open-ended operations will take far longer to validate.
"The distance depends on the task, not a timeline." It's a measured judgment—no storytelling that paints every job as just around the corner.
Moving forward, he sees the key technical hurdles as covering long-tail anomalies and cross-store few-shot adaptation. On the commercial side, it's getting the operations human-efficiency ratio and full-lifecycle cost to truly work.
Ultimately, how do you measure the success of a full-stack strategy?
Ren pointed to four outcomes: higher qualified output and lower retry/takeover time under the same task and service standards; less debugging effort when items, layouts, or stores change; post-deployment improvements that hold up in independent evaluation without degrading existing skills; and costs that steadily decline with maturity over long operational cycles.
Conclusion
Hardware and motion can be rapidly advanced by throwing resources at them. Verifiable evolution, replicable systems, and viable unit economics—these cannot be rushed.
Six Genie G2s: four passed, two held back. It's not a perfect score, but it's an honest one.
The race from "can demo" to "can work" in embodied AI has begun.
