SANet: lightweight geometry fusion for adaptive robotic grasping

Abstract

Real-time object perception with geometric awareness is a critical enabling technology for intelligent visual perception in robotics. However, deploying robust visual recognition on resource-constrained edge devices remains challenging due to computational limitations and the lack of geometric priors. This paper presents a lightweight technology framework that integrates an improved AlexNet-based dual-branch network (SANet) with dilated convolutions and a geometry-informed loss function, enabling simultaneous object classification and geometric property inference from a single forward propagation. Unlike conventional end-to-end grasping methods, the proposed framework emphasizes system-level technology integration: it fuses RGB appearance with noise-free STL geometric priors during training, then discards the auxiliary branch at inference to maintain a low computational footprint suitable for embedded platforms. Experimental validation on a humanoid robot for automated manipulation tasks demonstrates high recognition accuracy (93–97%) and stable operation across varying object appearances, with an end-to-end processing time under 0.31 s on a consumer-grade GPU. The work provides an efficient, interpretable, and geometry-aware solution that advances intelligent perception technologies in robotics and automation.