Research on Robot Motion Teaching Based on Human Pose Reconstruction Models

The objective of this research is to explore a humanoid dual-arm robot motion teaching method that integrates human pose reconstruction technology, additionally, the study assesses the accuracy of the robot in replicating actions. Firstly, we employed an improved Denavit-Hartenberg (D-H) parameter method for kinematic modeling of the robot's configuration. Subsequently, we extracted motion data from various video sources using a human pose reconstruction model. After preprocessing the data, we optimized redundant degrees of freedom and applied the Jacobian iterative method for kinematic solving to replicate human actions observed in videos. Furthermore, we assessed the results of robotic action reproduction through two dimensions: positional error and action similarity. Experimental results indicate that the robot successfully reproduces various actions from different videos. However, factors such as video quality and background can affect model capture precision, thereby influencing the reconstruction effectiveness of certain movements. For humanoid-configured robots, after appropriate data processing, employing deep learning-driven human pose reconstruction algorithms enables rapid and cost-effective construction of a robotic motion library.