Abstract—To increase Non-Player Characters’ (NPCs) realism,
we trained recurrent neural networks modeling human emotion
patterns. We generate training data by putting acting footage
through Facial Expression Recognition (FER) software. We used
4 different actors reacting to each other in a simulated police
investigation setting. Using Blender, we created 3D models of the
actors’ heads and animations of Paul Ekman’s 6 basic emotions.
In Unity, two head models face each other, with cameras coming
from their eyes to view the other. We allow the heads to receive
an emotion vector and animate them changing emotions. The
neural network can eventually be attached to the Unity scene, to
make the faces react to each other intelligently.
