GAMES Webinar 2017-16期(Siggraph 2017论文报告)| 刘利斌(DeepMotion Inc),刘天添(宾夕法尼亚大学)

【GAMES Webinar 2017-16期(Siggraph 2017论文报告)】

报告嘉宾1:刘利斌, DeepMotion Inc.





Libin Liu and Jessica Hodgins. 2017. Learning to Schedule Control Fragments for Physics-Based Characters Using Deep Q-Learning. ACM Trans. Graph. 36, 3, Article 29 (June 2017), 14 pages. DOI:


通过设计动作控制器并利用物理模拟,我们可以很容易的得到高交互性、高物理真实感的计算机角色动画。然而在实际应用中针对特定的动作设计控制器往往非常困难,这也使得基于物理模拟的方法在现有的游戏以及虚拟现实产品中较少得到应用。利用跟踪控制结合反馈来复现参考动作数据是一类能够有效降低控制器设计难度的方法。在本次报告中, 我将介绍一种利用Guided Policy Search为复杂动作学习跟踪控制器的方法,并在其基础上介绍一种利用Deep Q-learning算法学习的控制规划策略,从而实现对动态场景中的复杂人体动作的有效控制。


刘利斌,2009年本科毕业于清华大学基础科学班,2014于清华大学高等研究院获得计算机博士学位,2015-2017年分别于UBC及Disney Research任博士后研究员。现于初创公司DeepMotion Inc任首席科学家。研究兴趣包括基于物理模拟的计算机角色动画,以及与此相关的最优化控制、增强学习、深度学习等方向。主页:


报告嘉宾2:刘天添, 宾夕法尼亚大学





[1] Tiantian Liu, Adam W. Bargteil, James F. O’Brien, Ladislav Kavan. 2013. Fast Simulation of Mass-Spring Systems. ACM Trans. Graph. 32, 214:1–214:7.
[2] Sofien Bouaziz, Sebastian Martin, Tiantian Liu, Ladislav Kavan, Mark Pauly. 2014. Projective dynamics: Fusing constraint projections for fast simulation. ACM Trans. Graph. 33, 154:1–154:11.
[3] Tiantian Liu, Sofien Bouaziz, Ladislav Kavan. 2017. Quasi-Newton Methods for Real-time Simulation of Hyperelastic Materials. ACM Trans. Graph. 36, 23:1–23:16.

Traditional FEM based deformation methods are usually considered slow, because simulating those types of FEM materials relies on Newton’s method, where even one Newton iteration is too expensive to fit in the limited time budget for real-time applications. Fast methods such as Position Based Dynamics support only a limited selection of materials, and often converge to a different solution compared to the ground truth. From mass-spring system to general hyper-elastic materials, we proposed a series of methods, trying to fill the gap between slow but accurate physics-based animation and fast but ad-hoc real-time simulations. Our final method is typically more than 10 times faster than one iteration of Newton’s method without compromising quality.
In this talk, I will not only cover the idea of our SIGGRAPH 2017 paper but explain how we come up with that idea as well, which can be tracked back to our previous publication of fast mass-spring system simulation and projective dynamics.

刘天添,宾夕法尼亚大学计算机与信息科学系在读博士研究生,本科毕业于浙江大学竺可桢学院,硕士毕业于宾夕法尼亚大学,现为宾夕法尼亚大学五年级博士生,师从Ladislav Kavan教授。主要研究方向是实时软体仿真和快速几何处理。 个人主页:



You may also like...