Transferring skills from humans to robots is an appealing way for teaching artificial systems to perform a variety of different tasks. In this context, imitation learning appears as an important approach for teaching robots due to the generation of human-like movements and the ease of teaching new tasks. This paper addresses the use of a Kinect-based human motion capture system and the reproduction of arm movements in an upper-body humanoid robot. The objectives are threefold: (1) to deal with the lack of a kinematics model that assure coherence in the recorded 3D human poses; (2) to explore the inclusion of a shoulder complex based on a parallel mechanism; and (3) to demonstrate and evaluate how two robot models can be used to reproduce the human demonstrations. Several experimental results are included showing the upper-limb reproduction of human arm movements.

Functional locomotion requires continuous modulation of coordination within and between legs to flexibly accommodate demands of real-word environments. In this context, dynamic movement primitives (DMP) is a powerful tool for motion planning based on demonstrations, being used as a compact policy representation well-suited for robot learning. In this work, we study on-line adaptation of robot biped locomotion patterns when employing DMP as trajectory representations. Here, the adaptation of learned walking movements is obtained from a single demonstration. The goal is to demonstrate and evaluate how new movements can be generated by simply modifying the parameters of rhythmic DMP learned in task space. The formulation in task space allows recreating new movements such that the DMP's parameters directly relate to task variables, such as step length, hip height, foot clearance and forward velocity. Several experiments are conducted using the V-REP robotics simulator, including the adaptation of the robot's gait pattern to irregularities on the ground surface and stepping over obstacles.

In this paper, we compare two approaches for the reproduction and generalization of human-like movements by an upper-body humanoid robot. The specific purpose is to gain new insights into a key component of imitation learning: the ability to generalize from very few examples. In line with this, we aim to accurately portray the characteristics and generalization performance of those approaches based on the paradigm of motion primitives. Specifically, we focus on two established techniques, the dynamic motion primitives (DMP) and the principal component analysis (PCA). First, we recorded demonstrations from human subjects executing a reaching task in order to create a database of reaching motions. Then, the multiple demonstrations are mapped to an upper-body humanoid robot through an inverse kinematics algorithm that attempts to track the kinematics of the movement of human arms. Finally, the generalization performance is evaluated and the feasibility of the two approaches is discussed for this study of reaching movements