Task-oriented Sequential Pose Motion Primitives

Deconstructing a task into multiple phases and connecting each phase in a sequence to achieve complex motion planning is the prevailing approach in the field of robotics research. However, this paradigm also faces the problem of how to increase the generalization capacity of each stage and ensure effective smooth transitions between adjacent stages. To address this problem, we propose task-oriented sequential postural motion primitives. A number of phases are partitioned based on task characteristics, and the pose movement primitives for each phase are parameterized in a data -driven manner to facilitate the acquisition of specific skills from multiple demonstration trajectories. Besides, the problem of transitioning between temporally sequential motion primitives is modeled as a tracking problem of a moving target in order to achieve a seamless merging of movements. Finally, experiments on the Sawyer robot demonstrate the effectiveness and feasibility of the proposed method.