Object manipulation often demands both non-prehensile actions—like pushing or sliding—and grasping. Yet most existing approaches treat these as separate capabilities, then integrate them by sequencing distinct primitives according to a prescribed plan. That separation can limit flexibility and generalization.

A new paper on arXiv proposes Uni-PNP, a unified policy for prehensile and non-prehensile learning. Instead of stitching together separate modules, Uni-PNP learns a single policy that can decide when to push, slide, or grasp as needed to reach a target object pose. The authors frame this as a step toward more generalizable object manipulation.

The abstract does not provide experimental details or quantitative results, so the claimed benefits remain to be verified. But the conceptual shift—from sequencing separate primitives to learning one integrated policy—addresses a known bottleneck in robotic manipulation research.