Human motion retargeting to generate robot trajectory-replication of human-guided path

Abstract

In this project we propose and study a methodology to generate a robot motion by the retargeting of human motion captured using an HTC Vive system. As a way to test the outcome, the human motion was obtained by the performer trying to mimic or follow a robot end effector in real time. Our task is to take the human motion as an input, use several techniques to process it and then reconstruct from it the motion that the performer was trying to achieve. We then compare this motion to the original motion recorded from the robot itself. The processing of the data consists of applying PCA to obtain a two-dimensional projection, computing the approximate period of the movements by the location of points, sampling average points at points evenly spaced in time along the cycle, and then using a Bspline to reconstruct a continuous, smooth and closed curve. A generic method is developed utilizing Python script, except for a first preprocessing step realized with Blender. The results of our study show that from a dataset with variation in the order of 10 to 20cm (the human motion) we obtain a result with an error in respect to the motion recorded from the robot in the order of 1 to 5cm.