模仿学习:一种新人工生命动画方法

Imitation Learning:A New Approach in Artificial Life Animation

提出一种新的人工生命动画方法—模仿学习.模仿是一种非常有效的掌握运动技能的学习方式.一项运动技能为无数个相关运动序列的集合.通过模仿代表性运动序列,将蕴含的局部运动技能泛化,可获得完整的运动技能.模仿学习以运动相似度匹配和简单-复杂行为方法论为核心,并以进化计算为优化方法.模仿学习降低进化计算对传统评价函数的依赖,减少评价函数设计时间,提高优化复杂目标的能力,因此提高了制作效率.基于PhysX仿真平台,本文以人工猫的着陆行为验证了本文方法的有效性,并取得了良好的效果.

This paper proposes a new artificial life animation approach—imitation learning.Imitation is a highly effective learning method for acquiring motion skill which can be regarded as a set of numerous motion sequences.Imitating representative motion sequences to acquire the local motion skill and generalizing them can achieve the entire motion skill.The cores of imitation learning are motion similarity and simple-compose behavior methodology,and evolutionary computation is used as an optimization method.Imitation learning decreases the dependence of evolutionary computation on traditional fitness function and the time spent on designing a suitable fitness function.It also increases the ability of optimizing complex goal.So it increases the effciency of producing animation.We verify our method by training an artificial cat robot to learn landing behavior based on PhysX simulation framework,which achieves a good result.