A Constrained DMPs Framework for Robot Skills Learning and Generalization from Human Demonstrations

Dynamical movement primitives (DMPs) model is a useful tool for efficiently robotic learning manipulation skills from human demonstrations and then generalizing these skills to fulfill new tasks. It is improved and applied for the cases with multiple constraints such as having obstacles or relative distance limitation for multi-agent formation. However, the improved DMPs should change additional terms according to the specified constraints of different tasks. In this paper, we will propose a novel DMPs framework facing the constrained conditions for robotic skills generalization. First, we conclude the common characteristics of previous modified DMPs with constraints and propose a general DMPs framework with various classified constraints. Inspired by barrier Lyapunov functions (BLFs), an additional acceleration term of the general model is deduced to compensate tracking errors between the real and desired trajectories with constraints. Furthermore, we prove convergence of the generated path and makes a discussion about advantages of the proposed method compared with existing literature. Finally, we instantiate the novel framework through three experiments: obstacle avoidance in the static and dynamic environment and human-like cooperative manipulation, to certify its effectiveness.