Robot Dance Generation System Using Deep Learning

To make robot dance according to music is an interesting and challenging problem.On the one hand,dance is an art form that can bring pleasure to both dancers and audiences.On the other hand,dance involves a variety of human advanced abilities,including sports ability,perception ability,emotional experience ability,etc.Therefore,the research of robot dance has a positive effect on enriching the spiritual civilization of society and promoting the development of artificial intelligence.Recent advances in deep learning have enabled the extraction of high-level features from raw sensor data and made it possible for complex sequence modeling.New breakthroughs have been made in many fields,such as music information retrieve,natural language processing(NLP).There are some similarities between robot dance sequence and natural language sequence.Therefore,it will be a meaningful attempt to view dance gesture as vocabulary and use tools in NLP to analyze dance gestures and create dance sequences.In this paper,we use Word2vec to get the semantic features of dance gesture,namely Gesture Embedding.Then the basic language model is used to generate the robot dance sequences to analyze the similarities and differences between the dance sequence and the language sequence.Considering the result of gesture analysis and the constraints of robot platform,this paper presents a robot dance generating system,which uses gated recurrent units(GRUs)respectively to learn correlations between the global features of music and the gesture relation features,the local features of music and the motion density.Base on the outputs of the GRUs,robot dance motion sequences are sampled from a motion graph and synchronized with the beats of music.This method can be applied to the small-scale data set with various styles of dances in it,provided on the current commercial entertainment robot platform.The experiment is carried out on the AlphalS robot,it can perform stable and smooth dance movements according to the 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.