FLEXIBLE AND REALISTIC CHARACTER ANIMATIONS IN COMPLEX AND DYNAMIC ENVIRONMENTS

现代游戏越来越注重超现实主义和沉浸感，以更好地吸引玩家的注意力。游戏打破这种沉浸感的方法之一是，使用预先定义好的动画来打破移动或行动的流程。运动匹配是一种通过观察姿态和用户轨迹来预测动画最佳下一帧的解决方案。然而，缺点是当你增加数据库中可能的动画数量时，运行时成本也会增加。提出了一种称为“学习运动匹配”的解决方案（Holden等人，2020），该解决方案采用了运动匹配的积极特性，但也实现了基于神经网络的生成模型的可扩展性。本项目将通过实现记忆层来探索和改进学习运动匹配方法，在不牺牲增加运行时成本的情况下提高准确性。

Modern games have an increasing focus on hyper-realism and immersion to better capture the attention of players. One of the ways that games can break this immersion is by having animations that break the flow of movement or actions through the use of predefined animations. Motion matching is a solution for predicting the best next frame of an animation by looking at the pose and user trajectory. The downside however, is that when you increase the amount of possible animations in the database the runtime cost also increases. A solution was proposed known as 'learned motion matching' (Holden et al., 2020) which takes the positive properties of motion matching but also achieves the scalability of neural-network-based generative models. This project will explore and improve the learned motion matching method through implementation of memory layers to improve accuracy without the sacrifice of increasing runtime costs.