Goal-conditioned manipulation policies rely on a goal image—a picture of the desired post-manipulation scene—to steer actions. These images are typically synthesized from the robot's current view and a natural-language instruction, but their role as control targets has been taken for granted.
The new paper reframes this setup by treating the synthesized goal image itself as the control target, rather than as an intermediate representation. The authors propose a method for spatially-grounded goal-image synthesis, aiming to make the generated images more directly usable for policy execution.
The abstract hints that the work challenges a prevailing assumption about how such goal images should be produced or evaluated, though the full argument is not detailed in the available text. The approach could simplify the pipeline from language to action in robotic manipulation.