README: re-correct the terrain line (random rough patches, not flat)

README.md

@@ -11,7 +11,7 @@ pipeline_tag: robotics
 
 # walk-rough
 
-Walking gait trained in the rough-terrain task (Mjlab-Velocity-Rough-MicroDuck), simulation only: 8,000 iterations, 4096 envs, 324 min on one RTX 5090. The stock terrain curriculum kept almost every environment on the easiest tier for the whole run (steps under 1.5 mm, slopes under 4%), so treat this as a flat-ground walker with light terrain noise. Upright for a 12 s flat-ground take; no rough-ground take recorded yet. Not yet tested on a real Microduck.
+Walking gait trained in the rough-terrain task (Mjlab-Velocity-Rough-MicroDuck), simulation only: 8,000 iterations, 4096 envs, 324 min on one RTX 5090. Each terrain patch in this task gets a random type (flat, stairs up to 1.5 cm, an uneven grid, a slope) and a random difficulty, so it trained on mixed rough ground. An earlier version of this card called it a flat-ground walker; that was our misreading of the terrain setup. Upright for a 12 s flat-ground take; no rough-ground take recorded yet. Not yet tested on a real Microduck.
 
 A **perpetual** policy for the [microduck](https://github.com/pollen-robotics/microduck) (61-D observation, 14 actions, 50 Hz). Runs until told otherwise — a gait for the `walk` slot.