An Encoder-Decoder Architecture for Smooth Motion Generation

Trajectory generation for a dynamic task, where the outcome of the task is not ensured by simple repetition of a motion, is a complex problem. In this paper we explore a methodology for motion generation that retains the correspondence of the executed dynamic task. Throwing, which is not explored in this paper, is a very illustrative example. If we just imitate human throwing motion with a robot, the outcome of the throw with a robot will most likely not be very similar to the demonstrated one. In this paper we explore a deep encode-decode architecture, combined with ProDMP trajectory encoding in order to actively predict the behavior of the dynamic task and execute the motion such that the task is observed. Our example is based on the task of dragging a box across a surface. Guided by future work on transferability, in this paper we explore the parameters of the approach and the requirements for effective task transfer to a new domain.