Next generation motion controller and synthesis for game characters

自从电脑游戏诞生以来，动画就是游戏体验的精华部分。随着近年来动画获取和合成的进步，游戏中的计算机动画变得越来越复杂，越来越逼真。这种现实主义是有代价的：新游戏的计算和内存要求非常高。本项目的目标是提出一种新型的游戏动画控制器，在增加动画真实感的同时，显著减少计算和内存占用。我们利用人工智能和神经网络的最新进展来实现这一点。当前最先进的动画控制器将预先录制的动画存储在内存中，并在运行时将它们缝合在一起，以合成符合角色所具有的约束的新运动，如地形、与其他对象的碰撞等。在这项工作中，我们将使用它们来学习给定角色的运动，而不是将所有这些预先录制的动画保存在内存中，并使用它们来合成新的运动。一旦学习了运动，合成就是轻量级的，并且不会占用大量的内存。通过对神经网络的查询，实现了基于神经网络的运动合成。这一操作可以很容易地转移到云上，允许游戏机将大量处理时间释放给其他任务。另一方面，学习阶段是困难的和计算昂贵的，但它只需要以离线的方式完成一次，并且它可以由游戏公司使用集中的计算机集群来完成。使用这一新的游戏控制器概念，有效地将过程的复杂性转移到离线过程，只留下轻松的任务需要实时执行。我们正在与育碧在这个项目上密切合作，育碧是一家领先的电脑游戏公司，在蒙特利尔有一个大型研发工作室。使用这一技术将实现计算机动画的重大范式转变，这将有可能给计算机游戏领域带来革命性的变化，从而使公司、计算机动漫界以及整个加拿大游戏行业受益。

Animation is a quintessential part of game playing experience since the dawn of computer gaming. With recent advances in animation acquisition and synthesis, computer animation in games is getting more and more complex and more and more realistic. This realism comes at a cost: the computational and memory requirements for new games are very high. The goal of this project is to propose a new type of animation controller for games that increases the realism of the animation while, at the same time, significantly reduces the computational and memory footprints. We achieve this by using recent advances in artificial intelligence and neural networks. Current state of the art animation controllers store pre-recorded animations in memory, and at run-time they stitch them together to synthesize a new motion that conforms to the constraints that the character has such as terrain, collision with other objects, etc. In this work, instead of carrying all these pre-recorded animation in memory and using them to synthesize new motion, we will use them to learn the motion of a given character using state of the art deep learning techniques. Once the motion is learned, the synthesis is light-weight, and does not have a large memory footprint. Furthermore, the motion synthesis using neural networks is achieved by querying the neural network. This operation can be easily moved onto the cloud allowing the gaming console to release a lot of the processing time to other tasks. The learning stage, on the other hand, is difficult and computationally expensive, but it needs to be done only once in an off-line manner and it can be done by the gaming company using a centralized cluster of computers. Using this new game controller concept effectively shifts the complexity of the process to an off-line process, leaving only light tasks to be performed in real-time. We are collaborating closely on this project with UBISOFT, a leading company in computer games that has a large R&D studio in Montreal. Using this technique will achieve a significant paradigm shift in computer animation that has the potential of revolutionizing the computer gaming field thus benefiting the company, the computer animation community as well as the Canadian gaming industry as a whole.