Machine Learning for intelligent avatars in Virtual Reality simulations: Analysis of Reinforcement Learning techniques

Recent advances in Deep Reinforcement Learning (RL) allowed driving the avatar in increasingly realistic and complex environments, even Virtual Reality (VR) simulations. The granted wide degrees of freedom and the immersivity of the human in the virtual space, require the interaction with believable Non-Player Characters (NPCs) with non-scripted behaviors and able to adapt to contingent changes. This work focuses on the creation of ML-driven intelligent avatars able to behave an interact with a human in Virtual Reality simulations. Their behavior is decided by a neural network able to perceive the surrounding environment and to perform an action accordingly. This approach allows completing the assigned procedure while reacting to external stimuli, adapting to the human choices, and increasing the variance of the NPC behaviors. In particular are analyzed and compared different RL techniques, such as Curriculum Learning (progressive task complexity) and Imitative Learning (mimicking the proposed human solutions).