Learning From Demonstration For Redundant Manipulator

Motion planning for redundant manipulator is one of the central topic in robotics communi-ty. In recent years, algorithms on learning from demonstration(LfD) have been widely researched, which could reduce the complexity of robot motion programming. Therefore, LfD appeared as a promising route to improve the intelligence of robot motion and as a way to promote the popularity of robots in human society. This thesis first reviews the current process of humanoid redundant manipulator and summarizes the achievements in motion planning and LfD have been made. Sec-ond, based on traditional algorithm framework, human motor information is fused to guide the manipulator’s trajectory planning, which is implemented by the service robot. Then, This thesis describes a novel method based on affine trajectory deformation to refine human demonstration trajectory and optimize the executive time within manipulator’s physical constraints. The feasibil-ity of our approach is demonstrated on the humanoid robot arm, which learned table tennis strikes and reproduced same motion as fast as possible.The research content of the thesis includes two parts:1. Instead of current point-to-point trajectory planning, the thesis employs cubic-splines based multi-points trajectory planing to optimize motion duration within joint constraints. Further-more, in the inverse kinematics of redundant manipulator, human demonstration keyframes are introduced as initial value in each part of manipulator’s work area, in order to obtain trajectory plan with human motion information.2. The thesis proposes an optimal trajectory learning based on affine deformation. This ap-proach simultaneously optimizes trajectory similarity and executive time within manipula-tor’s physical constraints in order to obtain continuous motion trajectory close to human demonstration and give full play to manipulator’s performance.