Abstract: Soft enveloping grippers have emerged as a promising solution for grasping complex objects due to their superior adaptive performance and inherent safety.However,the diverse morphological structures of such grippers and the scarcity of relevant datasets impede the development of autonomous grasping strategies.To address these challenges,we propose an autonomous grasping method for soft enveloping grippers that integrates physical simulation with deep learning.Central to our approach is a physical simulator developed to parametrically model the gripper and construct diverse grasping scenarios.Through subregion analysis and a mechanism based on normal consistency contribution,we establish a quantitative evaluation system for grasping stability,while providing pixel-level estimation of grasping depth and surface normal vectors.Utilizing this pipeline,we generated a synthetic dataset comprising approximately 9 000 grasping samples across diverse viewpoints and objects.Based on this dataset,we developed SEP-GraspNet,an end-to-end deep learning network designed to predict optimal grasping strategies.Experimental results from both simulations and real-world trials demonstrate the feasibility and effectiveness of the proposed method,while achieving excellent grasping performance for objects in specific scenarios and poses.This work provides a feasible technical route toward realizing autonomous grasping with similar soft grippers.