The following Fig. \ref{964077} and \ref{253451} show the test of the prediction algorithm: after being fed the first 8 steps of the sequence, the remaining 12 steps are predicted by sampling the future distribution at each step. Obviously, errors tend to accumulate and throw the trajectory far from the ground truth, but as the learning proceeds, each step becomes more accurate (means are more spot on and the covariances become smaller) and the overall trajectory is preserved.