受脑启发的机器人认知抓取决策模型

Brain-inspired Decision-making Model for Robot Cognitive Grasping

为了让机器人获得更加通用的能力,抓取是机器人必要掌握的技能.针对目前大多数机器人抓取决策方法存在物品特征理解浅显,缺乏抓取先验知识,导致任务兼容性较差的问题,同时受大脑中分区分块功能结构的启发,提出了将物品感知、先验知识和抓取任务融合的认知决策模型.该模型包含卷积感知网络、记忆图网络和贝叶斯决策网络三部分,分别实现了物品能供性(affordance)提取、抓取先验知识推理和联想,以及信息融合编码决策,三部分之间的信息流以语义向量的形式传递.利用UMD part affordance数据集、该文构建的抓取常识图和决策数据集对3个网络分别进行训练,认知决策模型的测试准确率达到99.8%,并且抓取位置可视化结果展示了决策的正确性.该模型还能判断物品是否属于当前任务场景,以决策是否抓取以及选择什么部位抓取物品,有助于提高机器人实际场景的应用能力.

To obtain general purpose ability in human’s life and work,robots first need to master the skill of grasping objects.However,most current robot grasping decision-making methods have many problems such as simple understanding of object features,lack of grasping prior knowledge and poor task compatibility.Inspired by the functional structure of partitions and blocks in the brain,this paper proposed a decision model that integrates object perception,prior knowledge and grasping task.The model consists of three parts:convolutional perception network,memory graph network and Bayesian decision network,which realize the functional affordance extraction of objects,grasping prior knowledge reasoning and association,and decision-making with information fusion,respectively.Three networks were respectively trained on the UMD part affordance dataset,self-built common-sense graph,and self-built decision dataset.Test on the cognitive model verified its good performance with the accuracy of 99.8%.Results show that it can make reasonable decisions,including the ability whether the object belongs to the current task scene and the ability whether and where to grasp,which can help improve the robot’s availability in real applications.