Research On HMM-based Humanoid Robot Motion Imitation Learning Method

The humanoid robot refers to a robot that has humanoid appearance characteristics such as limbs,head,and trunk,and corresponding humanoid motion characteristics.Compared with other types of robots,humanoid robots are easier to imitate human behaviors,that is,their actions are "like human".Human-like actions enable humanoid robots to adapt to the human environment and use human tools and do not need modification.They can complete various complex tasks with human-like actions,and are more emotionally acceptable to humans.Therefore,imitation learning of human actions has always been a research hotspot in the field of humanoid robot motion planning.However,due to the wide variety and complex representations of human movements,and the differences in the body structure of different types of humanoid robots,the humanoid robots' imitation learning methods for human movements face huge challenges.In this paper,we conducted research on the human movement imitation learning method of small humanoid robot(NAO).Taking the actions in human action open data set(MSR Action3 D dataset)as target imitation actions,an imitation learning framework is established.In this framework,human motion data processing method,motion data modeling method,motion recognition method,humanoid robot motion trajectory generation method and motion model storage method were introduced in detail.The specific research work is as follows:At first,considering NAO robot's mechanism parameters and degrees of freedom configuration,the calculation method of 4 joint angles in the simplified model of human right arm is proposed.And the Cartesian coordinate change sequences of the joint position in the data set is converted into the joint angle sequences.Further,based on the obtained angle sequences and the kinematic constraints of the NAO robot,the angular velocity sequences and angular acceleration sequences are calculated.The angular velocity sequences and angular acceleration sequences are formed as motion data together with the angle sequences.Secondly,the Gaussian mixture hidden Markov model(GMM-HMM)is used to model the action data to establish action models.According to the established action models,the recognition method of action data is proposed.According to the model parameters,a method for extracting discrete motion elements from continuous motion data is proposed.According to the extracted motion elements,the methods of motion sequence generation,sequence length normalization and sequence averaging are proposed to realize the generation of humanoid robot's motion trajectories.According to the relationship between action models,an action model storage method which organizes action models into a binary tree storage structure is proposed.Finally,based on the human motion data in the MSR Action3 D data set and NAO robot experimental platform,joint angle calculation verification experiments,motion modeling and motion recognition experiments,motion element extraction and humanoid robot motion generation experiments,and motion model storage experiments are designed.The experimental results prove the effectiveness of the imitation learning framework and the methods proposed in this paper.