Robots that imagine the future before acting propose both a move and a guess at how the scene will change. A convincing imagined future does not mean the paired move will actually produce it, and the mismatch surfaces only when the robot moves.
So: draw several candidate move-plus-future pairs from the same frozen model, score each before executing, and run only the best. The score blends three checks. Is the move smooth and within range? Does a learned inverse dynamics model, which infers motion from an observed change, recover the same move from the imagined future? And does that future show progress on the instruction?
On a simulated manipulation benchmark, success rose from 6% to 24%. The task-progress check carried the most weight.