Currently, contactless demonstration learning faces significant challenges in data acquisition, and its trajectory generalization capability is limited when adapting to new task parameters. To address these issues, this paper proposes a method for learning action primitive trajectories based on a dynamic parametric regression model. The proposed approach first integrates a Gaussian mixture regression model to reduce redundant features arising from individual demonstrations by synthesizing multiple demonstration trajectories. Subsequently, building on dynamic motion primitives (DMPs), the dynamic parametric regression model is used to capture the dynamic relationship between feature parameters and task parameters, thereby improving the model’s generalization performance. Additionally, this study introduces the Fréchet distance as a similarity measure to evaluate generalization ability. Finally, the effectiveness of the proposed method is demonstrated through pick-and-place experiments involving objects, using the Kinova collaborative robotic arm.