Comparing Point Cloud Encoding Architectures for Language-Guided Semantic Grasping

Language-guided robotic grasping systems almost universally default to PointNet++ for 3D encoding and fuse language with geometry through simple concatenation. We show that how you fuse matters far more than what you encode. In a controlled comparison of three frozen backbones and two decoder architectures across 30 household objects, cross-attention fusion yields a +9.1 percentage point gain in Success@1 over standard concatenation, while switching backbones alone produces no significant difference. A cross-attention decoder that lets language queries attend to per-point geometry \emph{before} global pooling replaces the information bottleneck of late fusion with task-directed spatial focusing. The best-performing configuration, DGCNN with cross-attention, reaches 47.0% Success@1 while training 5-9 times faster than alternatives. An interaction trend suggests that DGCNN's edge convolution features may benefit disproportionately from cross-attention, though per-backbone comparisons remain underpowered and require further validation.