Strip the demo lighting and you get a robot that has to know where its own body is before it can move it. US11731277B2, granted to Boston Dynamics in August 2023, addresses exactly that with "Generalized coordinate surrogates for integrated estimation and control."
The B25J filing tells a different story than the keynote. Classified under B25J 9/1682, B62D 57/032 (legged locomotion) and B25J 13/089 (force/position sensing), the patent unifies two things usually built separately: estimation (figuring out the true state of the robot's body from noisy sensors) and control (deciding what the actuators should do). The 'surrogate' coordinates are a representation that lets both share the same math.
Here is why integrating them is the hard, valuable move. A legged robot's controller is only as good as its belief about where its limbs and center of mass actually are. If estimation and control are designed in isolation, errors in one corrupt the other, and the robot oscillates or falls. Coupling them — control that is aware of estimation uncertainty — is what makes balance robust on real ground.
This is a more sophisticated version of the slip-recovery and sensor-noise work Boston Dynamics patented years earlier. Where those addressed specific failures, this addresses the underlying framework: how a legged machine maintains a coherent, controllable model of its own dynamic body in real time.
The honest limit is that none of this is about what the robot does — picking, carrying, dancing. It is about staying upright while doing anything at all. That ordering, foundation before application, is the correct one, and it is the part humanoid startups most often skip on their way to a manipulation demo.
For anyone auditing the humanoid race, this patent marks the depth of the incumbent's lead. The flashy newcomers show hands; the company that has been at this longest is patenting the integrated estimation-and-control core that decides whether any of those hands ever do useful work without falling over.