Seamless Human Motion Composition with Blended Positional Encodings Seems very similar to the MotionVAE paper for the figure 8 motion.