Everyone wants to be the layer above the robot

Between August 21 and September 2, 2026, three organisations described a product that runs on hardware they did not build. Galaxy Corporation set out Soulware, a reinforcement-learning control layer its chief executive Choi Yong-ho likened to an Android-style operating system for robots, currently teaching K-pop choreography on Unitree machines and planned as a standalone product in 2027. ABEJA and Murata Manufacturing demonstrated a dual-arm manipulation policy built on NVIDIA’s Isaac GR00T N1.7 running on Murata’s own hardware. And Generalist was valued at $3 billion on a cross-embodiment foundation model intended to drive different bodies from one policy.

All three are betting the margin sits above the machine. They are betting on different things transferring.

Three claims about what transfers

Behaviour. Soulware’s current demonstration is choreography — timed, full-body, repeatable motion performed to an external clock. Choreography is unusually portable because it is specified in advance and does not have to react to a changing object; what has to transfer is a mapping from an intended pose sequence onto a given set of joints. That is a real engineering problem and a much narrower one than manipulation. It also has an immediate buyer, which manipulation research mostly does not.

Manipulation policies. GR00T N1.7 is a base model others fine-tune with their own teleoperation data. What transfers is representation and pretraining, not the finished skill; ABEJA and Murata still had to collect demonstrations of their own task. This is the most conservative of the three claims and the one with the most working examples behind it.

Embodiment itself. Generalist’s claim is the strongest: one policy, many bodies, new tasks from three to twelve seconds of video. If it holds, hardware differentiation collapses. It is also the least demonstrated in public, which is what the valuation is actually pricing.

The wall all three run into

Kinematics. A policy learned on one arm’s link lengths, joint limits and dynamics does not move to another for free, and the failure is not graceful — it is a trajectory that was correct on the training body and collides on the target one. That is the same wall sim-to-real transfer has been hitting, arriving from a different direction. It is why the shallow claim (pretraining transfers) has more evidence behind it than the deep one (skills transfer).

There is also an unglamorous version of this business that nobody is valuing at $3 billion: tooling. Evaluation, data management and simulation harnesses are what a fleet operator buys before it buys a policy, and they are indifferent to which model wins.

Why it matters

The layer question determines who keeps the margin as humanoid unit costs fall. Every hardware maker in the sector is currently arguing that its integration is the differentiator; every software layer is arguing that integration is a transitional cost. Historically the software side of that argument has won when the hardware became interchangeable and lost when it did not, and robot bodies are not interchangeable yet.

What makes Soulware the odd one out is that it is the only one of the three with revenue attached today — not from licensing the layer, but from tickets sold to watch what the layer does. That is a weaker claim on the platform position and a stronger claim on the near-term cash, and it is worth keeping the two apart when the 2027 product launch arrives.

Disclosure: Embodied Wire has a commercial relationship with Galaxy Corporation. Coverage follows the same sourcing standard as all other companies, and self-reported figures are labeled as such.