An Empirical Study of Attention Mechanisms for Lightweight Robotic Grasp Detection

Attention mechanisms have been widely adopted in robotic grasp detection, yet which components actually contribute to performance remains unclear. We present a systematic empirical study isolating the effects of channel attention, spatial attention, and multi-scale feature fusion on a lightweight grasp detection network (∼1.1M parameters, 400+ FPS). Through controlled ablation experiments on the Cornell Grasp Dataset with five-seed evaluation, we find that: (1) spatial attention is the most effective and stable single component (84.7±1.2% IoU accuracy, 70.6±4.6% with full rectangle metric), while channel attention achieves the highest mean IoU (0.425±0.017, +3.9% over baseline); (2) combining channel and spatial attention (CBAM) degrades performance due to cascaded multiplicative over-suppression; and (3) multi-scale fusion increases training variance without consistent benefit on small datasets. We also introduce a quality-weighted loss with sin/cos angle representation that resolves angle prediction collapse. These findings provide practical guidance for attention design in grasp detection.