基于深度强化学习的混杂场景目标物体推抓协同策略  

A Cooperative Strategy for Pushing and Grasping Target Object in Cluttered Scenes Based on Deep Reinforcement Learning

作　　者：胡楷雄 宋远航 周勇[1] 李卫东 HU Kaixiong;SONG Yuanhang;ZHOU Yong;LI Weidong(School of Transportation and Logistics Engineering,Wuhan University of Technology,Wuhan,430063;School of Mechanical Engineering,Shanghai University of Technology,Shanghai,200093)

机构地区：[1]武汉理工大学交通与物流工程学院,武汉430063 [2]上海理工大学机械工程学院,上海200093

出　　处：《中国机械工程》2025年第1期133-140,共8页China Mechanical Engineering

基　　金：国家自然科学基金(51975444)。

摘　　要：为提高机器人在混杂场景中抓取被遮挡目标物体的成功率和效率,提出一种基于深度强化学习的“推动”和“抓取”协同推抓策略。该策略利用两个深度Q网络,以RGB-D图像为输入来确定推动或抓取动作,并通过推动改变物体排列以优化抓取条件。该网络使用“抓推抓”三阶段模型训练方法显著提高了抓取能力。基于图像形态处理的方法识别并过滤低质量抓取动作,从而提高成功率和效率。实验结果表明,该方法有效提高了目标物体的抓取成功率和效率。To improve the success rate and efficiency of robotic grasping for occluded target objects in cluttered scenes,a collaborative push-grasp strategy was proposed based on deep reinforcement learning.The strategy employed 2 deep Q networks and used RGB-D images as inputs to determine push or grasp actions,which optimized object arrangement for better grasping conditions.A“grab-push-grab”three-stage training method was introduced in the model to enhance grasping capabilities significantly.An image morphology-based assessment method effectively identified and filtered low-quality grasp actions to increase successful rates and efficiency.Experimental results confirm that this method significantly enhances the successful rate and efficiency of grasping target objects.

关 键 词：机器人抓取 混杂场景 深度强化学习 多动作协同 

分 类 号：TG54[金属学及工艺—金属切削加工及机床] TP242.2[自动化与计算机技术—检测技术与自动化装置]