Motion Capture-Based Robotic Imitation: A Keyframeless Implementation Method Using Multivariate Empirical Mode Decomposition

Robotic imitation faces challenges due to the lack of nuanced movements when employing keyframe methods, which can potentially lead to the uncanny valley effect due to constraints in fitting data within motor speed capacities. This research proposes a keyframeless motion-transferring method for robotic imitation using motion capture data. Initially, we implement motion capture data into the NAO 6 robot, retargeting the joint angles from a hierarchical human body structure to the motor rotations. Second, to biomechanically optimize robotic imitation, we adopt multivariate empirical mode decomposition (MEMD) to decompose and analyze the motion capture data in the frequency domain. Third, we demonstrate that MEMD outperforms the Fourier transform (FT) in motion-capture-based robotic imitation and introduce an optimization algorithm. Finally, we evaluate four types of robotic motion imitations (picking-up, walking, punching, and Bunraku puppet motion) across five implementation methods (original data implementation, Laban keyframe method, FT, convolutional neural network autoencoder, and our method) using both NAO 6 robot sensors and a motion capture system. The results indicated that our proposed keyframeless motion-transferring method outperforms others in applying and controlling complex, nuanced nonlinear motion capture data for robotic imitation, offering a potential approach to studying the uncanny valley issue.