This document summarizes an internship project to generate 3D avatar animations from latent representations. The intern aimed to manipulate robots through natural language by utilizing 3D avatar motion data provided by DeNA. Experiments using autoencoders and principal component analysis (PCA) found that PCA performed best, able to compress the high-dimensional motion data into 30 dimensions with less than 1% error. The conclusion is that linear dimension reduction methods like PCA can highly compress motion data, and each dimension of the resulting latent representations may correspond to meaningful motions like floating or squatting.