The State Machine Just Disappeared

In humanoid control, the layer that used to be hand-written is now being learned. The Light-Loco-Parkour preprint shows what that looks like in practice: a single policy, conditioned on depth and a velocity command, fires the right whole-body skill on the fly, with no state machine in the loop.

For years, humanoid demos split the work in two. A learned controller did the movement. A hand-coded state machine chose the moment. The split was comfortable. It was also a tell. Every clean, scripted-looking parkour clip was, on inspection, a stack of if-then rules telling the robot which controller to load. The advance in the Light-Loco-Parkour paper is that the if-then rules are gone. Reward alone picks the transition.

That changes the test a reader can apply to the next demo. Does the policy actually decide when to switch, on the same depth stream it walks on, or does a state machine hide somewhere in the pipeline, dressed up as a learned module? If the first, the field has moved. If the second, the demo is a costume.

The limit sits in plain sight: one humanoid, controlled obstacle layouts, indoors and out. The state machine just disappeared, but only inside the controlled course. Outside it, the question stays open.