Dynamic Movement Primitives (DMPs) enable robots to learn and generalise motion skills from demonstration, but their performance depends on both trajectory characteristics and data preprocessing. This paper investigates the effect of preprocessing choices and trajectory properties on the reproduction and execution accuracy of DMP-based learning from IMU demonstrations.
A complete pipeline is developed to capture human arm motions, preprocess the data, learn DMP models, and execute the resulting trajectories on a simulated Franka Emika Panda robot. Multiple trajectories of varying complexity are evaluated, including a sine wave, a circle, a letter “R”, a pick-and-place task, and a jerk-optimised trajectory. The impact of outlier removal, resampling, smoothing, and the number of basis functions is analysed using quantitative error metrics.
Results show that preprocessing significantly affects performance, particularly in the presence of noise. High-curvature trajectories are highly sensitive to outliers, while smoother trajectories are more robust. Additionally, complex tasks require higher model capacity for accurate reproduction. The findings highlight the importance of adapting preprocessing and model parameters to the specific trajectory, providing practical guidelines for deploying DMP-based learning in robotic applications.