The problem
Almost every robot controller today is trained for a single body. Lengthen a leg, strap a payload on one side so the centre of gravity moves, or swap a sensor, and the policy that worked yesterday stumbles. The fix is to collect data and retrain, often for weeks. For a team with five variants of one robot, or a customer who bolts on their own gear, that cost repeats every time. The body is treated as a fixed fact when it is really an input.
Why now
Generalist whole-body controllers have just become real: NVIDIA's SONIC showed one policy driving a humanoid across many kinds of input. The open half is generalising across bodies rather than motions. Morphology-conditioned policy work, and startups pitching an omni-bodied robot brain, are all working on it. The idea is easy to state and hard to feel, which is why a demo that lets a layperson feel it is useful now.
What it is
A controller that is handed a description of whatever body it is in: joints and torque limits, sensors, mass distribution, contact points. It reads that, takes over, and carries out the task. This demo is a mockup of that moment. You pick a biped, quadruped, hexapod or wheeled base, drag the payload or the leg length, and a spirit pours in as the descriptor fills. The controller is a few dozen hand-written rules, not a learned model, and every panel says so. It does not learn, and its physics is a simplified kinematic model.
Who it's for
Robotics integrators and product teams with a family of bodies: warehouse platforms with swappable payloads, inspection robots in several sizes. It is not for single-robot research labs that already own a tuned policy, and it is not for consumers who just want a toy robot. It is also not a safety case: nothing here should be read as evidence that a real controller can do this.
How it makes money
We do not charge for the demo, and that is correct for now: the claim is unproven until a model trained across many bodies holds up on a body it has never seen. If it does, the honest model is per-robot-variant licensing or a per-seat integration tool, sold to the team that currently pays for retraining. Selling a hosted policy is rejected, because latency and safety push control onto the robot.
Go to market
Start with one integrator that already maintains three or more variants of a single platform, where retraining cost is visible on their own invoices. Offer a descriptor format and a conformance test they can run on a new body in a day. The reason they would forward it: the failure table of what breaks when a sensor drops out, which engineers share with each other because it names problems they already recognise.