AADFNet: An adaptive asymmetric dual-branch fusion network for background-robust grasping

Robotic grasping, particularly for mobile manipulators, often suffers from performance degradation in diverse, unpredictable backgrounds. Existing RGB-D grasp detection methods frequently struggle to maintain accuracy and efficiency on resource-constrained platforms due to insufficient background robustness. To address this challenge, we propose AADFNet: an Adaptive Asymmetric Dual-branch Fusion Network consisting of three coordinated components for background-robust grasping: 1) an Asymmetric Dual-Branch Encoder (ADE) that processes RGB and depth modalities with specialized backbones to decouple object features from background noise while efficiently capturing modality-specific characteristics; 2) a Cross-Modal Coordinate Attention (CM-CA) module that facilitates deep, cross-modal fusion by generating a unified attention map guided jointly by RGB and depth; and 3) an Adaptive Multi-scale Feature Decoder (AMFD) that dynamically adjusts receptive fields for precise grasp localization amid complex background textures. We also introduce a synthetic dataset, GAA, for systematic training and evaluation. Extensive experiments demonstrate that AADFNet delivers competitive performance with significantly reduced model size. Real-world experiments on a mobile manipulator also validate its practicality and effectiveness for mobile grasping tasks.