TSSG: A Two-Stage Segmentation and Grasping Model for Enhancing Grasping Precision

Language-guided image segmentation and grasping represent fundamental tasks in vision-language-action (VLA) systems. However, current single-stage models often suffer from task interference and incomplete feature fusion, which hinder the effective integration of segmentation and grasping in multi-task scenarios. To address this, we propose TSSG, a Two-Stage Segmentation and Grasping model that decouples segmentation and grasping into sequential processes. First, a segmentation network trained with a composite loss function (Dice Loss + Focal Loss) generates high quality probability object masks. These masks then guide a lightweight Mini U-Net grasp decoder to predict antipodal grasp configurations. Evaluated on the OCID-VLG dataset, TSSG achieves state-of-the-art performance, with a 1% improvement in mIoU (segmentation) and significant gains of 4.7% in J@1 and 1.2% in J@5 (grasping) compared to prior single-stage approaches. Real-world deployment on a Franka Emika robot further validates its robustness, demonstrating reliable grasping across diverse objects. TSSG’s modular design enhances precision, offering a scalable solution for vision-guided grasping.