Robots that grasp and move objects need to predict exactly where to place their gripper and how to move there. Current methods treat this as a language problem: they turn coordinates into words, then generate one word at a time. This ignores the spatial nature of the task and loses precision with every step.
The authors instead attach a small, lightweight module to an existing 2-billion-parameter . The module reads the model's internal spatial representations directly, without generating any language. It predicts the grasp location and movement path as continuous numbers, solving the problem as geometry rather than grammar.
On standard benchmarks, this approach cut trajectory error by over 60% and improved grasp prediction accuracy by 85% compared to the language-based baseline. It also ran nearly four times faster than a larger competitive system, showing that a better interface to existing models matters more than scaling up the model itself.