In manual assembly environment, human motion generation enables a new paradigm of human centered production planning and human interaction with machine systems. Therefore, this paper presents a new method for constrained human motion generation using low-dimensional latent spaces to optimize manual assembly tasks, such as reaching tasks. A space partitioning method is proposed to project human motions to a specific point in the latent space. Furthermore, the probabilistic models are extended further by mapping the reaching right-hand position in Cartesian space with their latent space representation using Multivariate Polynomial Regression (MPR). The obtained results are promising compared to regular human motion-capture-based analyses because the proposed method generates human motions faster, and it has the potential to enable cobots to anticipate human motion during human-robot collaboration.