The operating system that teaches robots K-pop
The training data is already out there
The data, the big data — it's all sitting on YouTube already. If there's a song or a dance video you like, we build automated training out of it and put it on a robot.— Choi Yong-ho, CEO of Galaxy Corporation · press conference, August 21, 2026
The usual way to teach a humanoid to dance is motion capture: put markers on a dancer, shoot in a dedicated studio, extract the coordinates. It is accurate, but slow. Every song needs its own shoot. And it leaves decades of existing choreography footage untouched.
Galaxy Corporation starts from the opposite end. The public dance video is the training data. Reconstruction instead of capture, collection instead of production. The moment that swap works, the economics of training change character: adding a new song stops being a matter of booking a shoot and becomes a matter of dropping a new video into the pipeline.
K-pop is unusually well suited to this. The same routine is filmed and published from multiple angles in multiple versions; the accuracy of the movement has already been proven on human bodies; and because the form is group choreography, multi-body data comes bundled in. From a robot-training standpoint, the K-pop video archive is a dataset in its own right.
From video to stage, in four steps
Korean outlets describe Soulware's process in four steps.
| Step | What happens |
|---|---|
| 1. Input | A dance video goes into the operating system. No separate motion-capture rig is used. |
| 2. Reconstruction | Inside a 3D virtual simulation, the human movement in the footage is rebuilt as volumetric motion and the routine is replicated. |
| 3. Correction | Foot traction is controlled and joint tremor and hyperextension are corrected. This is the stretch where a human range of motion is translated into something a robot's body can actually carry. |
| 4. Training | Thousands of rounds of reinforcement learning, cross-validated against simulation, until the robot holds its own formation and spacing while performing. |
Step three is the hard one. Transfer a human dance verbatim and the robot falls over. People correct their center of gravity in real time through the soles of their feet; a robot has to be handed that correction as arithmetic. Traction control and joint correction are not what make the choreography beautiful — they are what make it possible.
Step four rewrites the industry's clock. A human idol needs months of rehearsal to learn a routine, and that learning stays inside one body. Reinforcement learning inside a simulation is not bound to physical time, and what comes out of it is a file, so it copies. That is where the company's claim comes from: a routine learned in one place can be performed the same day by robots in another city.
One person, twelve robots
The measure of how mature this technology is isn't how flashy the moves look — it is how many robots one person can run at once. According to Korean press reports, as of December 2025 a single operator handled a single robot. As of August 2026, one operator runs twelve from one console, and on a stage shared with human dancers the robots carried off moves as demanding as a taekwondo spinning kick. The company has set out a target of scaling to as many as 100.
We started with one, then five at once, then twelve — and one person does all of it. We've got the capability to train 30, 50, 100 at the same time.— Choi Yong-ho, CEO of Galaxy Corporation
A curve that runs from one to twelve in eight months says something about the unit economics of live performance. Putting twelve people on a stage means aligning twelve schedules, twelve people's condition, twelve sets of travel. Putting twelve robots on a stage takes one operator and one console.
One further condition is worth noting. The operating system takes as its design goal the ability to control robots from different manufacturers in the same way. The aim is that the machines on stage will not have to come from the same manufacturer or share the same spec — which is to say, to prise the content away from the constraints of hardware procurement. Galaxy Robot Park's permanent show schedule and how it is run are covered in the Galaxy Robot Park FAQ.
Building the operating system, not the hardware
American and Chinese companies build the robot hardware. We're building the OS — think Google's Android.— Choi Yong-ho, CEO of Galaxy Corporation
The analogy the company keeps returning to is the smartphone industry. The company that makes the handset and the company that makes the operating system running on it do not replace each other; if anything, the OS side gets to treat every manufacturer's hardware as its market. Choi has said of Soulware that the aim is for "anyone to be able to buy a robot and run this OS on it," and at the August 2026 press conference the company set out a plan to ship the operating system as a product during 2027.
That coordinate clears up a common misreading of what Galaxy Corporation actually does. It is not a robot manufacturer competing on hardware performance, and it is not a virtual idol company running characters on a screen. Its center of gravity is the layer that sits on top of robot bodies that already exist — choreography, persona, content. That said, at its August 2026 press conference the company disclosed that it is separately preparing its own robot hardware dressed in character IP. The argument is developed further in Humanoid as platform and the Enter-Tech category map.
Calling it an operating system carries straight through to the business model. Ticket sales and theme park admissions are revenue from the content layer; the moment another operator puts Soulware on its own robots, revenue appears at the software layer. That is why the company puts its weight on the operating system layer rather than on selling hardware.
Why call it 'Soul'
What matters is that a robot has a heart and a mind too. Amid the global rush into Physical AI chips and robot hardware, we kept asking ourselves what comes next — and Soulware started there.— Choi Yong-ho, CEO of Galaxy Corporation
The name reads like marketing, but it lines up with the development direction the company describes. Soulware is not meant to stop at replaying pre-loaded movements: it is being built to carry an artist's movement and persona onto a robot that then interacts with the audience. The grand opening announcement was themed 'From Code to Soul.'
The company says it is deliberately slowing down on anything that touches personal data. A robot that remembers a visitor and greets them by name on their next visit will be deployed only once risk management and security frameworks are in place. "Privacy is the thing we take most seriously," Choi said, adding that "we can only collect personal information once a consent framework is there."
As a technical document, Soulware's coordinate is clear enough. It is not the technology that builds a robot's body; it is the technology that decides what goes inside a body someone else has already built. The concept and the industrial definition of Physical AI are covered in the Physical AI whitepaper.