基于坐标注意力的杂乱环境中机器人推抓协同学习

Cooperative Learning of Robot Pushing and Grasping Based on Coordinate Attention in Cluttered Environment

为提升机器人在杂乱环境中推抓协同性能、增强网络感知物体位置和物体间的位置信息的能力,提出一种基于物体位置信息的推动与抓取协同网络来解决机器人在杂乱环境中的抓取问题。该网络使用2个全卷积网络分别从视觉观察中推断出抓取和推动操作的位置与方向。使用坐标注意力模块分别沿着二维空间的2个方向聚合特征,即在水平空间方向上捕获长距离依赖关系的同时在垂直空间方向上保持物体的位置信息。然后生成推动和抓取的位置特征的注意力图,以提升网络推断操作位置的准确性。提出物体分散度从全局角度衡量环境中物体间的分散程度,并设计基于物体分散度的推动奖励函数来提升推动动作的质量。在仿真实验中,该网络的抓取成功率和动作效率分别为75.1%和73.2%。在现实世界中,该网络的抓取成功率和动作效率分别为80.1%和76.2%。

To enhance the collaborative performance of pushing and grasping in cluttered environments and to strengthen the network's ability to perceive the location of objects and the positional information between objects,a collaborative network based on object positional information for pushing and grasping was proposed to address the problem of robotic grasping in cluttered environments.This network employed two fully convolutional networks to infer the locations and directions of grasping and pushing actions respectively from visual observations.A coordinate attention module was utilized to aggregate features along two directions in the two-dimensional space,capturing long-distance dependencies in the horizontal space direction while preserving object positional information in the vertical space direction.Attention maps of the pushing and grasping location features were then generated to enhance the accuracy of the network's inference on the positions of actions.Object dispersion was introduced to measure the dispersion degree of objects in the environment from a global perspective,and a pushing reward function based on object dispersion was designed to enhance the quality of pushing actions.In simulation experiments,the network achieved a grasping success rate and action efficiency of 75.1%and 73.2%,respectively.In the real world,the network achieved a grasping success rate and action efficiency of 80.1%and 76.2%,respectively.