Trajectory generation with multi-stage cost functions learned from demonstrations

Abstract

Abstract Learning from demonstration provides an effective method to resolve the problem of teaching robot to execute complex motions without expert knowledge about the robot system. In this paper, we present a novel learning from demonstration method based on multi-stage cost learning. The recorded demonstrations are assumed to be composed of several substages chained together. Leveraging this assumption, a segmentation and cost learning framework is proposed to search for the cutting points that divide the unsegmented demonstrations into multiple substages and retrieve cost function for each substage. To the best of our knowledge, it is the first solution to learn multi-stage cost functions in continuous domain without restricting the possible cost functions into limited numbers or simple forms. To generate new trajectory, a complete objective functional is constructed based on the learned multi-stage cost functions plus other constraints like obstacle avoidance and is optimized with functional gradient method. The generated trajectory can adapt to new environments while maintain the specific properties of each substage as demonstrations. The effectiveness of the proposed method is verified through simulation study and experiments conducted on a real robot manipulator.