Directable Animation

动画是一项创造性的工作，它严重依赖于可用的工具和工作流程，关键帧是许多传统动画和计算机动画的核心。最近，越来越多的新方法利用机器学习和物理来通过预测建模来创建动画运动。然而，这些方法中的许多都来自机器学习社区，专注于自动化解决方案，很少关注让动画师对运动进行足够的控制。在这个项目中，我们寻求开发方法，利用物理学和数据驱动的预测方法的好处，同时也允许动画师的指导性。我们追求三个具体的方法。首先，我们将研究关键帧如何利用基于被动物理的模拟来轻松生成自然接触和碰撞行为，如应用于与其环境交互的角色。其次，我们探索和扩展了数据驱动的解决方案，以解决基于关键帧的运动完成问题。第三，我们研究了使动画师能够更好地利用深度强化学习来产生具有可控动作样式和动作变化的基于物理的角色的方法。综上所述，这三种方法很有希望将艺术家和动画师的一些创造性需求与预测模型所提供的能力结合起来。

Animation is a creative endeavor that relies heavily on available tools and workflows, with keyframing being at the heart of much of traditional animation and computer animation. Recently, new methods are increasingly leveraging machine learning and physics to create animated movement via predictive modeling. However, many of these approaches some from the machine learning community and focus on automated solutions, with little attention being paid to giving animators sufficient control over the motion. In this project, we seek to develop methods that exploit the benefits of physics and data-driven predictive methods, while also allowing for animator directability.We pursue three specific approaches. First, we will investigate how keyframing can leverage passive physics-based simulation in order to easily produce natural contact and collision behaviors, as applied to characters interacting with their environment. Second, we explore and extend data-driven solutions to keyframing-based motion completion problems. Third, we investigate methods that enable animators to better leverage deep reinforcement learning to produce physics-based characters with controllable motion styles and motion variations. Taken together, these three tackle promising way of integrating some of the creative demands of artists and animators with the capabilities afforded by predictive models.