A sixteen-legged robot navigating rough terrain has, in principle, an enormous coordination problem. Each leg must decide when to lift, where to place, how long to push. The combinatorics of gait planning scale brutally with leg count. Centralized controllers that try to solve the whole problem at once become intractable.
Chen, Wang, and Revzen didn't solve the coordination problem. They dissolved it. Each body segment runs an identical state machine. The state machine takes input from exactly one source: the segment in front of it. When a leg is in contact with the ground, the controller couples tightly to surface events — an event cascade, responsive to terrain. When contact is lost, the same state machine produces rhythmic fictive locomotion — a central pattern generator, maintaining cadence without feedback.
The two modes are not separate controllers. They are the same state machine operating in different regimes, selected by a single binary signal: ground contact, yes or no. Contact present means respond to the world. Contact absent means generate your own rhythm. The transition is automatic.
This produces adaptive gaits. On flat ground, the cascade propagates cleanly from front to rear, and the robot walks in regular waves. On rough terrain, where individual legs lose contact at irregular intervals, each segment independently switches between cascade and CPG modes. The global gait pattern — the thing a centralized planner would try to compute — emerges without anyone computing it.
The through-claim is about where coordination lives. It doesn't live in a planner that sees the whole body and the whole terrain. It lives in the propagation rule: each segment copies the segment ahead of it, with a delay set by its own ground contact. The global pattern is a shadow of local propagation.
This is coordination without a coordinator. The regularity of flat-ground gaits and the adaptiveness of rough-terrain gaits come from the same rule. The difference is not in the controller but in the terrain.