Abstract: To address the issue of outlier trajectories in the demonstration learning process for robots with a wide operating range, a method for robust modeling of robot demonstration trajectories based on mixture probability motion primitives is proposed. The method first preprocesses the collected demonstration trajectories using dynamic time warping (DTW) to optimize the distribution of trajectory points by optimizing the mapping matrix. This resolves the problem of varying trajectory lengths caused by jerky demonstration movements and uneven speeds. After obtaining trajectory sets with equal time lengths, a trajectory point threshold based on confidence intervals in the probabilistic motion primitive model is used to cluster and assign weights to samples within the trajectory set. Finally, a mixture probability motion primitive model is constructed based on the clustering results and weight parameters to suppress the parameter deviation caused by outlier trajectories and improve model robustness. To validate the effectiveness of the proposed method, experiments were conducted using handwritten letter trajectory datasets and real robotic arm feeding trajectory demonstrations. Metrics such as average Fréchet distance and average Euclidean distance are introduced for quantitative evaluation of the model’s reliability. The experimental results demonstrate that the proposed method effectively mitigates the parameter deviation caused by outlier trajectories. It exhibits stronger robustness compared to traditional probabilistic motion primitive models and Gaussian mixture models, highlighting its promising application prospects in the field of industrial robots with large workspace radii.

Key words: demonstration learning, dynamic time warping, probabilistic movement primitives, trajectory modeling