Source-demo success only requires the box to reach the wall and settle near a 90 degree tip. During PGDG sampling, generated rollouts must also stay collision-free with the center block. The mouse sphere is the only actuator; the rest of the page is generated from that single recorded demonstration.
This direct view overlays the dashed sampled demo command trajectory, the solid fitted Bezier decode, and the sampled action trajectories so the final dataset distribution can be compared directly to the source demo in the same space used for collection.