基于GraspNet的物體平鋪場景下類別導(dǎo)向抓取算法

打開文本圖片集

中圖分類號：TP391 文獻標(biāo)識碼：A

GraspNet-based Category-oriented Grasping Method for Object Planar Scenes

SONG Shimiao，GU Feifan，GE Jiashang，YANG Jie

Abstract： To solve the problem class-based grasping in multicategory tiled scenes，this paper adopts different feature fusion methods proposes a joint optimization algorithm MC-GSNet （Multi-Class GraspNet） that fuses category semantics grasping posture an optimization algorithm MT-GSNet （Multi-Task GraspNet） that builds a multitask learning model. The improved methods explicitly incorporate category information， optimize the generation logic grasp poses enhance the algorithm's adaptability success rate in multi-category object planar scenes. Experimental results on the public dataset GraspNet-lBillion demonstrate that the proposed methods significantly improve task adaptability grasping success rates in multi-category planar scenes. MC-GSNet MT-GSNet achieve 32.6% 43.9% average accuracy improvements in grasp detection， respectively； MT-GSNet exhibits superior adaptability to unseen objects due to its integration segmentation features. The experimental results in the simulation environment show that the grasp successful rates （GSR） MC-GSNet MT-GSNet reached 88.3% 95.0% respectively，which can meet the needs actual engineering deployment.

Keywords： grasping detection; category-oriented; feature fusion; GraspNet-lBillion

機器人抓取技術(shù)在結(jié)構(gòu)化場景中已取得顯著進展，但在復(fù)雜平鋪場景，如平面上隨機的多個類別物體中仍面臨諸多挑戰(zhàn)[1]。（剩余10151字）