China's physicist RSI kills the black horse

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

In the morning, you don't even close your eyes, and there's a slight clash in the kitchen.

When you push the door out, the robot operator X1 stands at the table, and you don't talk, and it pulls the chair out a little bit.

When you close the door, it turns to the living room and picks up the cat's claw.

The cat bows down and strikes the ground with its tail. As soon as the feathers on the cat's stalk reach half-height, the cat looks up and runs away.

At the last moment, X1 lifted his arm, and the cat's claw scratched through the air. His wrist was lighter, his feathers moved back to the tip of the cat's nose, the cat turned around again, and it slipped away again. A few laps down, it had touched the cat's breath.

When you open the door at night, before the words "I'll be back" hit, X1 comes up with a pair of sneakers and sends a special greeting: "Princess, please change your shoes".

This is the personalized family service that everyone dreams of, a robot manager who can adapt to your habits, understand your family environment, and customize services for you.

But the way to make fun of cats, the "greeting ceremony" of returning home is different in each household. These things can't be pre-packed or finished when they leave the factory.

The easiest way is to show it yourself again and again.

To this end, Knowin has released the generative learning framework GLOW (Generative Learning of World).

They believe that if robots can get into thousands of homes, the key is to "understand people's skills, generalize execution".

This is what is behind their unique judgment:

Commercialization of home scenes does not begin with generic capabilities, but with limited, high-frequency, verifiable task sets.

And there are elderly people, children, pets, fireworks and chemicals in the house, and security is the gateway to the home, not the dividing line.

In other words, let's see what the effect of teaching is.

Knoen published four groups of real-world controls, each group on the left of which is a human-operated process, and on the right of which is a self-executed robot after learning.

The first group is open box receivers.

When humans demonstrate, they collect the same thing, but when the robot executes, the object to be collected is replaced by something else.

It first opens the lid and can be generalized to identify changes in objects and locations, then puts things in the box, and finally fuses the lid together to form an air.

Left: people showing the box receiver open; right: robots switching objects to do so (accelerated 4.5 times)

Note that the action is only taught once, and robots need to use on-site feedback to complete each stage, successfully embracing the action into multiple objects and locations.

The second group is the pomegranates.

People use a black mouthpiece, but there is only one green mouthpiece in the robot's hand.

The size, color, and location of the shell varies, but it is capable of identifying and applying the grip, direction, and slope relationships learned from human operations to the shell, and then shelling and releasing it.

Left: Humans pump water with their black mouths; right: robots replace green pump lights with 3.5 times the acceleration)

In other words, robots learn to perform this task, rather than simply repeated teaching actions that can adapt to changes in the landscape.

The third group is table scrubbing, which is also the most interesting.

The robot learns to erase a heart shape by rubbing a blanket along a heart-shaped path.

Left: People scrub the table along the heart-shaped trajectory; right: Robot retrieval (accelerated 3.5 times)

It combines the goal of "screwing the table" with the habit of "screwing the table".

The fourth group is the drinkers.

There are several cups on the table, and people pour them out in order, so that the robot can see which cup to pour first and which to pour next.

The robot learns how to identify objects accurately, picking up, pouring, rearranging, and throwing back, taking several steps into a complete process.

On the left: people take turns drinking; on the right: robots restore the entire process (accelerated 12 times)

There are four groups of tasks, the model weight has not moved a single parameter. Seeing the human operation process, the robot can do it well, learn it, and can be generalized to more locations, objects, and scenes.

In tasks, AI must synchronize multiple capabilities: identify objects and actions, locate operational targets, judge interaction outcomes, and plan for the next step.

In addition, models must also have the ability to generalize to objects and locations.

In the Knox test, GLOW's capability was ranked first in three lists, and the numbers proved its ability to generalize.

GLOW achieved an average success rate of 62.2% over 18 complex tasks in RoboDojo, while GPT-6 (Robocurve) achieved an average success rate of 22.2%, up 40 percentage points.

The GLOW score was 86.7%, the GPT-6 Astra score was 58%, and the single model ranked first with an average score of 86.7.

In Embodied Arena's list of 2D-Embodied QA Benchmarks, GLOW was ranked No. 1 out of 20 body models with an overall score of 65.62.

When robots watch human-operated videos, they use a skill encoder to compress continuous visual experiences into a replicable skill context.

It extracts the structure of the task, i.e. what the object is to be operated on, in which region to be placed, what the sequence of actions is, and how to do it.

There are six stages of boiling water. The top row is filled by humans using their black mouths, and the bottom row is mapped by robots to grasp, orientate, and incline the relationship to the green boiling water.

When executed, the context is combined with the current image, language instructions, robot status, and each token generates judgments and actions.

GLOW also combines current real-time observation to integrate your "teaching" of yesterday into current self-generation.

It automatically plans the route, avoids the cups of water in the teapot and opens the boxes. You can also adjust your goals and practices with language at any time.

It is possible to do this if the brain and hands are in the same model.

A common VLA route in the industry is to use a large visual model as an encoder, extract an action module, and perform a complex representational conversion in the middle.

The WAM route is designed to produce a future "real-life video image" and then output the moves, calculating the expense.

At the core of GLOW, KnowinGlow is a native unified model of self-return that weaves vision, language, and action into the same sequence.

BrainSkill is responsible for spatial understanding and semantic landscaping, while ActSkill is responsible for generating closed-loop motion, originally separating "brain" and "hands", which are now two skills in the same model.

It's a real-time RGB-D image of the camera in the background, the instructions you gave, the current angle of the joint and the clavicle of the mechanical arm, and what moves you've made and what feedback you've received in just one second.

GLOW combines information from the physical world into a "long article" like text, and then sequentially continues the next move.

Physical reasoning tests. The color lines in the diagram are the candidate's movement trajectory, asking "which group to pick in order to cook the tomatoes", and the model answers: cut the tomatoes, open the stove fire, uncover the lid, pour the oil, and put the tomatoes in.

This is not the same as the vague "move_to cup" command.

The model directly generates the three-dimensional position of the end-executor, the relative parallelism with the increase in rotation, and the claw joint control, accurate to millimeters.

Target point positioning test. Left: "Put the orange blocks on the green blocks" where the model gives the target point [[368,664]]; right: "Location of the blank area between the blue cup and the blue bowl" where the model gives [[534,800]]

Knowles' choice of this path is not only about "learning over and over again", but also about the long-term value of unified modeling and continuing learning.

GPT-6 Astra's latest advances in operating robots further confirm the long-standing path of self-return technology that Knox has followed.

In the real physical world, the drawer may stick, the cup may slip out of the claw, the light is not planned enough.

At this point, Knoen directly pulled the GPT-6 Astra from what Brockman called the AGI into the same simulated kitchen.

The result, called AGI's GPT-6, was a card on a drawer.

But GLOW knows -- it has a mission operating system called Harness, which gets an execution feedback for every move it makes.

For example, when the system comes down to "going 24.3 millimeters to the cabinet", reversals show that it actually went only 1.6 millimeters.

Once the displacement remains, the system immediately realizes that it encounters physical resistance. This "obstruction" results in the context of the next round, along with the clawed state and the new image.

The brain decides whether to continue using the force or to change the posture.

The same mission, the same starting point, for GLOW and GPT-6 Astra to do it all over again, it's different.

The bottom drawer of the closet, the GPT-6 Astra hand extended past the top drawer, corrected 1200 steps, and the bottom drawer was still closed.

GLOW retracely detects the obstacle, switches to a level near the claw, closes the opposite corner, 1138 steps, and opens.

Open the bottom drawer. Left: GPT-6 Astra, 1200 steps not open; right: GLOW, 1138 steps open for 10x video acceleration)

GPT-6 Astra folded 863 steps, the bottle and bowl weighed 3.4 cm, and the threshold value was 3 cm.

GLOW 157 steps, with a distance of 6.9 mm.

Grab the bottle and put the bowl. Left: GPT-6 Astra, 863 steps ((7 times faster); right: GLOW, 157 steps

When approaching the bottle, the bowl or pulling the hand, the GLOW switches to a small boundary movement of only 5 mm each time, re-examined between the two tones.

After grabbing the cup, it does not move directly, but first asks for a lift of 30 millimeters to see if the object moves with the claw, confirming that the claw has entered the carrier.

They also "try to pull" before opening the drawer, confirming that the gravity has been established.

This kind of physical test is very much like a human instinct.

The GPT-6 Astra control robot is not weak. Robocurve connects it to its two arms, blocks it into the bowl 20 times to 19 times, and brushes the entire network.

But in the second half of the same test, the millimeter-scale puzzle enters the slot 20 times, only two times.

To understand the task, and to be able to complete the task, there is only a "pain" hand in the middle.

The general route of the big model is to collect, train, release, and retrieve massive amounts of human data.

It's difficult to collect massive amounts of data in each and every complex, unmarked user household.

GLOW's solution is to create your own data.

The engine that generates the data is known as the KnowinDream. It generates a virtual family by type, furniture and objects, and then arranges various events, such as a family taking away a cup or passing by while the robot moves something.

KnowinDream keeps changing the scene: changing the lights, changing the cyborg punk neon, and changing the Chinese restaurant.

The real experience of running out of the machine will give it a boost. If X1 makes a cat succeed, or grabs a cup slide, the processes of success, failure and recovery will flow back to the KnowinDream.

It reconstructed the scene at the time, expanded it into more similar situations, and handed it over to Knowin World for verification.

KnowinWorld · Two-arms simulation interaction. Scenes, missions, multiple viewpoints and actions result in a continuous body experience.

So far, KnowinDream has generated about 200M pieces of data like this.

KnowinWorld doesn't require huge computing power to generate video images, it only performs pure physical state projections.

In this way, it can make millions of tries and mistakes in the brain at a very low cost.

Folding bowl assignments

The most critical step is to calibrate the "evolutionary mechanism" itself.

The actual results of robotic execution are compared to previous KnowinWorld predictions and biases used to continually calibrate models. These real feedback also flows back to data generation links, modifying synthetic data, labeling and screening standards.

Stronger base models will generate more accurate data and labels; higher-quality data will train the next generation of stronger base models. The new base models will in turn improve the quality of the next round of data generation, labeling and verification.

This resulted in a recurring reinforcement of the "Model Strength → Data Improvement → Next Generation Model Stronger → Data Further Improvement" loop.

This is called the RSI, which refers to self-improvement in the physical world: an improved model will in turn enhance the ability of the next round of self-improvement.

The safety valve is also located in this loop. Each capacity upgrade must be verified in the world model and implemented through the three-way inspection of task completion rate, recovery rate and safety return.

As soon as they find a step back, they immediately roll back.

Norin has no intention of letting robots walk through the door of the house with "anything they can do" - they'll be cut off from limited, high-frequency, security-verifying tasks.

There are elderly people, children, pets (such as the first cat) in the house, fire and chemicals, and security is at the doorstep.

At the top of the threshold, deciding who can really get into a million households is a "one-size-fits-all" method of delivery.

The engineer doesn't have to go upstairs to collect the data, refine it, and you do it all over again.

In September 2026, GPT-6 Astra demonstrated that the cognitive ability of the general large model has reached the edge of the physical world.

GLOW took a different path.

The architecture is unified into regression, and the world model does not pursue image realism; it merely seeks to make decisions that are useful, and the data from passive collection is transformed into real RSI circuits that execute regression flows.

From bringing new toys to your cat, to collecting your unique personal items, "Once Upon a Time" takes you through the last mile from running water tools to personalized housekeepers.

The X1 folded to accommodate a height of less than 40 cm and is scheduled for release next year.

At that point, you want it to say what, how to wipe the table and teach it again.