一种基于姿态关系特征的机器人舞蹈生成方法

A robot dance generation method using gesture relation features

如何使机器人伴随着音乐舞蹈是一个有趣又有挑战性的课题,为此提出了一种自动生成机器人舞蹈动作序列的方法.采用门控循环单元(GRU)网络分别学习音乐的全局特征与舞蹈姿态关系特征之间的相关性、音乐局部特征与舞蹈动作密度特征之间的相关性,再结合舞蹈动作图,采样并规划出与节拍同步的机器人舞蹈动作.该方法适用于目前商业娱乐机器人平台上提供的规模小、风格多样的机器人舞蹈数据集.将其在优必选Alpha1S机器人平台上进行实验后发现,机器人能够根据算法生成的动作序列演绎出稳定、流畅的舞蹈;调查问卷表明,人们很难区分舞蹈片段是由该算法生成的还是由人类设计的.

To maneuver the robot to dance in the accompaniment of the music poses an interesting and challenging problem.This paper presents a method of automatically generating robot dance sequence,which uses gated recurrent units(GRUs)respectively to learn correlations between the global features of music and the gesture relation features,and those between the local features of music and the motion density features.Base on outputs of GRUs,robot dance motion sequences are sampled from a motion graph and synchronized with beats of the music.This method can be applied to the small-scale data set,which contains various styles of dances and is provided on the current commercial entertainment robot platform.The experiment is carried out on the Alpha1S robot,which can perform stable and smooth dance movements in synchronization with motion sequences.The result of survey shows that it is difficult to distinguish whether a dance fragment is generated by the algorithm or designed by human beings.