Two Million Viewers Watch Six Robots Compete Overnight in a Grueling Head-to-Head Test

  It's the first time in the history of the world.

These two days, the robotics world's "heartbreaking offer" is completely out of the loop!

Over the past 24 hours, more than 2 million people have been on-line watching over the real-world convenience stores surrounding the six G2 geniuses. At 10 a.m. this morning, the robot "new employees" officially ended the "conversion challenge" at the 24-hour convenience store!

The entire 24-hour live broadcast, we followed it all the way down.

This live broadcast, just to answer a question and face a real-world messy dilemma, is it possible for robots to work?

On November 2, Xiang decided to fight this tough battle.

Six Genius G2 robots, equipped with the latest Agent Framework Genie Master, launched a 24-hour series of "transfer checks" as "temporary employees of convenience stores".

It's all live, it's all network surveillance. No filters, no clips.

Just now, four successful "reversals" have been released to critical acclaim!

"Robots Can't Stop Working" was broadcast live at convenience stores 24 hours a day, setting a record.

In total, the six G2 robots completed 1,730 tasks, including receiving customers 683 times, delivering 369 items, making coffee 147 times, serving breakfast 126 times, inspecting supplies 277 times, and cleaning tables 128 times.

This live broadcast, which marks the emergence of a water-sharing device that jumps out of a simulated environment, a laboratory, and can work in real-life situations.

Putting robots in convenience stores is a dangerous decision.

Because in less than 100 square meters, all the physical chaos in the real world: the mechanical arm can grab the hard bottles, pinch the soft bags, wear the holes; the system must simultaneously fill up, make coffee, drag tasks; customers can move around, talk, change orders at any time.

In this case, more deadly is multi-plane collision.

The filling machine stopped in front of the shelves, stopping customers for drinks; the cleaning machine curved into the narrow lane, blocking the outlet lines.

That's why hardness requires "24 hours of continuous operation". It doesn't measure the battery, only to amplify the hazard with time: how to roll a random cartoon that can't be seen for 10 minutes into a mission snowball in a day?

In order to receive this copy, Chiu sent six G2 "interns" with different personalities.

They are not only explicitly assigned jobs, they are even given real jobs:

It's not that bad.In the morning, eight people were rescued by the manager.

What is it?The company's main responsibilities are language interaction, welcoming customers and providing emotional value.

TreasureIt is responsible for the collection and packaging of goods, efficiency and efficiency.

Tired KingHe is responsible for patrolling the entire store and filling the shelves, silently filling the gaps.

The cold one.Exclusive skills are to clean tables and keep dining areas clean.

Oh, my God.Heating area manager: responsible for heating cooked food, precise control of heating and cooking.

The six trainees made a timely card call at 10 a.m. on November 2.

They were greeted not by a one-time demo, but by a 24-hour, four-hour demon-dragging exercise.

The peak at midday (13:00-14:25)Around: white collar flows, orders are high at the same time. Picking up, packing, eating must be done.

It's the night peak (19:00-20:00)After the day's consumption, hardware and computing began a period of fatigue.

It's a nighttime experiment.There are no human shopkeepers around, and robots need to guard the entire store independently.

It's been a long time coming.This is the ultimate battlefield between "Carrie" and "Carrie". The demand for coffee and hot food will peak, and the precision and rhythm of execution on the mechanical arm will be an extreme strain.

In order to make this examination absolutely real, the live stage set up a rigorous "examiner's dilemma".

Broadcast as "HR Director", Deadpool transforms the four key KPIs: continuous running time, the number of real-life services, four times resistance to stress, and customer reviews.

Technical experts sit in town "leading Mentor", and the routers play difficult customers, throwing out difficult drilling questions at any time.

It's good, it's applauded on the spot; it's bad, and it's on the big screen.

From "one-offs in the lab" to "continuous operations of real convenience stores", in the face of continuous customer flows, sudden demands and long-term physical wear and tear, can they not turn off the line and touch the fish?

When this unique convenience store first opened its doors, it was visited by a lot of people, and many of the sisters took pictures.

After all, the image we've been dreaming of for a long time: "Man enjoys, robots work" has finally arrived, and it looks incredibly comfortable.

Over the course of the 24-hour live broadcast, we witnessed many stunning scenes of robots driving around.

It's good that most of them have done their job well!

For example, when picking up goods, the rope recognizes the angle of attachment itself, raises the tools from a clever angle, makes it easy to grab them, and finally puts the goods in the basket.

Then it slithered to the human being.

Interestingly, the technicians exploded: they discovered that the coffin was stolen while it was in operation, so they added the function of timely patrol.

And then there's Cold Sister.

He recognized that the table next to him was clean, so he went over it. When he came to the table full of human trash, he picked out the trash and started cleaning it.

Unfortunately, the fridge was cut a few points in the process because it mistakenly threw human food away as garbage.

After broadcasting live for most of the day, in the early evening, the Vice President of the Center for Physical Research in Chiang Kai-shek visited the store with foreign experience officers.

He gave her a bottle of Coke, but it wasn't light enough for her to move a little bit.

What is the most difficult task for a treasure?

Human children may not open the cabinet until the age of three or four, while robots, like small children, will only grasp the cabinet door at the beginning and learn to open it after training.

When I was a teenager, I was a teenager, and I was a teenager, and I was a teenager, and I was a teenager.

The hard-working king is all around the place, and if the goods are not right, it will be corrected immediately.

The foreign boy cowardly changed the position of the gum in his mouth, tired of the king's eyes burning, immediately found the wrong thing, re-placed the goods with his claws, behind, all the genius master's brain is in command.

After that, my younger brother came to the cafe and for the first time in his life tried coffee made by a robot.

The younger brother was handed a hot bag by the side of the car.

Out to see the hot tub, in to see the door.

During the live broadcast, the viewers saw a robot making a cup of coffee and a bag of plastic bags.

The action seems simple, but in the tech world, it's called "crossing the gulf".

That's the famous one."The Moravian Paradox"Logical reasoning, which is extremely difficult for humans, is easy for computers to master; but unconscious skills that are very easy for humans, such as grabbing software objects, sensing space, are difficult for robots to climb.

For example, "supply to shelves".

When The Sleeping King approaches the shelf, it has at least three tasks to complete.

First of all, it needsSpatial understanding and environmental perception。

The lights of the convenience store are changing, customers are walking around at will, and the goods on the shelves may be distorted. It must construct 3D space in real time through visual models, recognizing that "it's a film, it's a hanging".

Second, it needsMission PlanningThe agent's brain decides which pack to take first, which angle to take, and how to spin the wrist afterwards.

Finally, it needs to beAction execution and feedbackLarge models at the bottom (such as GO-3 Preview) will convert the plan into the torque and joint angles of the mechanical arm, and real-time microtransactions based on the tactile and visual feedback when captured.

Without full collaboration, the robot will only crawl like a blind man, or directly overturn a shelf.

The real addiction of robots is that it can't be done with just a few pieces of hardware and an open-source model. It requires underlying computing power, a general model, a body-intelligent framework.

The usual action in a convenience store, behind a whole set of full-featured technology systems.

Completing the task requires model and body collaboration.

Stable operation, need to adapt to the environment, deal with abnormalities;

Scalable applications will also reduce training costs, accelerate deployment and reuse capabilities.

A real addiction is not a piece of hardware.

The core value of AI WAVE's overall deployment is this: using convenience stores as an application gateway to showcase huge systems under the iceberg: data, platforms, entities, agents, simulations, models.

The fourth edition of AGIBOT WORLD 2026, the first real-world Physical Agent data set, focuses on decision-making and covers nine categories of logical reasoning scenarios;

The Genie Studio 2.0 operating platform covers the entire lifecycle of the "training-testing-computing-tube", with deployment verification of all new operating scenarios from 0 to 1 being completed within 7 days.

G2 Max, the heavy-duty industrial robot genius, is the industry leader in long-range, multi-tasking tasks.

Genie Master allows robots to continue to evolve in real life scenarios;

The Genie Sim simulation platform undertakes training and pre-testing prior to deployment;

GO-3 Preview is a next-generation base model from Xiao, which uses the native WRAM (World Reasoning Action Model) architecture to connect task reasoning, future prediction, and action generation, allowing for further learning and long-range precision tasks using a small number of examples.

The group experience learning (CEL) paradigm, which opens up the experience acquisition, experience utilization, and scale up the entire online learning chain, turns every success or failure in a real job into a shared learning material for the group.

In the real world, mistakes cannot be avoided, and there are no robots that can't make mistakes.

If you deploy 100 convenience stores and you need a team of 100 engineers to be on site as a "nursery", what is the commercial value of a robot?

How to break this spell?

The killer code given by Xiao was: Genie Master with Agent Framework.

On October 27, Xiao officially released the Genie Master framework. Its core mission is only one: to enable robots to achieve "side-up, side-unlock new skills".

It is not only able to organize long-range tasks, maintain progress, but also to continue tasks autonomously after abnormal processing.

When G2 encounters unprecedented situations, it doesn't need humans to come to its rescue.

How is it done?

At its core, the Genie Master framework builds a bridge between a robot's "brain" and "body".Embodied Execution Protocol (EEP)

The results and feedback from the execution of the limbs will directly fuel subsequent decision-making and capacity optimization.

In the past, AI brains and mechanical limbs broke apart.

The brain receives the instruction to "hang up" and the body does it. But without it, the body cannot tell the brain "why it is not hanging up", nor can the brain give more subtle instructions.

The execution was completely blindfolded.

The EEP protocol opens a dual interface that unifies the four categories of interfaces: capability, guidance, feedback, and continuity.

Now, the execution results of the limbs are backed up to the brain in real time, accurately.

This involves one of the hardest technicalities in live broadcasting: Physical RSI, a continuous evolution from deployment to iteration.

With the support of the Genie Master framework and the underlying model GO-3 Preview, Xiao created a unique two-line evolutionary mechanism for robots.

One route allows the robot to "remember the experience"; another route allows the robot to "grow muscle memory".

The first evolutionary path: Harness RSI

You can think of it as a "working diary" of a robot and a "memory library of the brain".

During the daytime work, G2 will keep track of successes and failures.

When the late-night convenience store knocks, the big model behind the system begins to work. It will unify the data of the day, extract replicable laws from it, and immerse it in long-term memory.

Following a nightly summary, the system generated a replicable tactical tip: "Meeting small targets with visual obstruction, prioritize the use of wrist cameras for auxiliary positioning".

The next day, when another robot hangs a towel, it doesn't have to go through the "fail-analyze-adjust-re-test" trial and error process.

The second evolutionary route: Policy RSI (for capacity consolidation)

In the process of execution and recovery, the system generates a large amount of real feedback data containing "problems, adjustments, results".

These high-quality "error-correcting data" are then processed and sent directly back to re-train the action strategy model.

Effective improvement is no longer just a "guide word" hanging in the brain, but is deeply embedded in the underlying weight of the model, translating into "basic operational capabilities" that can be directly summoned for follow-up.

The restorative movement that used to take the brain "thought analysis" to complete has now become the "muscle memory" of the underlying model.

Two paths, two swords combined.

Yesterday's failed record is the fuel for today's evolution; today's work experience is the starting point for tomorrow's evolution.

It demonstrates that robots are not only better at running their predefined processes, but that they have actually learned "how to improve themselves".

The landing value of physical intelligence does not depend on how well the capital market PPT is painted, nor does it depend on how well the promotional video is cut.

It depends on only three things:

Will real tasks be stably completed, will the cost of human intervention be lowered, and will the capacity to improve be transferred to new tasks?

The 24-hour live broadcast, which opened on November 2, is about tearing these three questions open to everyone without reservation.

From "How to get addicted" to "How to keep improving after addiction".

Put the goods on the shelves, put a cup of hot coffee on the shelf, and clean up the scraps.

These actions, which are insignificant for humans, are the craze of robotics over half a century of technology.

This live broadcast demonstrates how robots have turned mistakes into stairs of evolution, and we will witness the dawn of a truly "working age" of physically intelligent, hardcore.