[source] siggraph
[year] 2007
模仿人的动作。
修改动作数据,PDs驱动
Imitation-based learning appoach使可以迁移运动风格,但需要精细的控制。因此有本文的思路。
得到的控制器还可以在几个间转移。
相关工作:
1. Data-Driven using Mocap data. Motion Graph
2. dynamic control systems
3. Dynamic motion synthesis as variational optimization problem
4. ML techniques
Data Collection and Processing
Motion Rectification
因为表演人物与模型不一致,且为平面模型
Formulation
\hat m = m + d
d = \sum_{j=1}^{m}{h_{ij}B_j(t;c_j,w_j)}
B = 1/2(1+cos(\pi/w(t-cj))) c_j – w_j < t < c_j + w_j时
每个node恰为swing takeoff and kickdown of a stance foot or with the halfway of a swing phase
因此,要确定 h 及 w
目标:使PDs模拟的动作与原始动作接近
Optimization Method
各种优化算法的选择 : randomly choose initial parameters and run a downhill simplex method reapeatedly with diff. init. param. to find a local extremum.
Behavior Control
构造stationary和transition controller以使状态空间可以延续
方法: regression techniques that selects nearby samples in the state space and combines the output poses at those samples to produce a desired output.
近邻的计算,提出考量几个要素间距离
个人以为重要的是思路,对动作数据作修改而非制作精细的控制器算法。