Latent Dynamics for Artefact-Free Character Animation via Data-Driven Reinforcement Learning

In the field of character animation, recent work has shown that data-driven reinforcement learning (RL) methods can address issues such as the difficulty of crafting reward functions, and train agents that can portray generalisable social behaviours. However, particularly when portraying subtle movements, these agents have shown a propensity for noticeable artefacts, that may have an adverse perceptual effect. Thus, for these agents to be effectively used in applications where they would interact with humans, the likelihood of these artefacts need to be minimised. In this paper, we present a novel architecture for agents to learn latent dynamics in a more efficient manner, while maintaining modelling flexibility and performance, and reduce the occurrence of noticeable artefacts when generating animation. Furthermore, we introduce a mean-sampling technique when applying learned latent stochastic dynamics to improve the stability of trained model-based RL agents.