Robust trajectory segmentation for programming by demonstration

Abstract

A novel trajectory segmentation and modeling approach is presented. Trajectory segmentation and matching is an important step in the programming by demonstration (PbD) process to extract the user's intentions from multiple trajectories. To match multiple trajectories, the segmentation and modeling approach must be consistent and robust to disparities caused by robot dynamics and human imperfections. Several curve segmentation approaches have demonstrated substantial potential in the field of image processing and gesture recognition. They emphasize reduction of the degree of mismatch between given and model curves. However they fail to reduce mismatch between models of multiple trajectories recorded to demonstrate the same intention.We propose an M-estimator for trajectory modeling and set up a new segmentation criterion to address the issue. The proposed approach is better suited for PbD of mobile robots. The approach is evaluated for real robot trajectories.