Realism in animation sequences requires movements to be adapted to changing environments within the virtual world. To enhance visual experiences from animated characters, research is being focused on recreating realistic character movement adapted to surrounding environment within the character's world. Existing methods as applied to the problem of controlling character animations are often poorly suited to the problem as they focus on modifying and adapting static sequences, favoring responsiveness and reaching the motion objective rather than realism in characters movements.

Algorithms for synthesizing motion sequences can then bridge the gap between motion quality and responsiveness, and recent methods have shown to open the possibility to recreate specific motions and movement patterns. Effectiveness of proposed methods to synthesize motion can however be questioned, particularly due to the sparsity and quality of evaluations between methods. An issue which is further complicated by variations in learning tasks and motion data used to train models.

Rather than directly propose a new synthesis method, focus is put on refuting existing methods by applying them to the task of synthesizing objective-oriented motion involving the action of throwing a ball. To achieve this goal, two experiments are designed. The first experiment evaluates if a phase-functioned neural network (PFNN) model based on absolute joint configurations can generate objective oriented motion.

To achieve this objective, a separate approach utilizing a hierarchy of phase-function networks is designed and implemented. By comparing application of the two methods on the learning task, the proposed hierarchy model showed significant improvement regarding the ability to fit generated motion to intended end effector trajectories.

To be able to refute the idea of using dense feed-forward neural networks, a second experiment is performed comparing PFNN and feed-forward based network hierarchies. Outcome from the experiment show significant differences in favor for the hierarchy model utilizing phase-function networks.

To facilitate experimentation, objective oriented motion data for training network models are obtained by researching and implementing methods for processing optical motion capture data over repeated practices of over-arm ball throws. Contribution is then threefold: creation of a dataset containing motion sequences of ball throwing actions, evaluation of PFNN on the task of learning sequences of objective oriented motion, and definition of a hierarchy based neural network model applicable to the motion synthesis task.