物理即時模擬下的人體肢體平衡

Abstract

本論文提出了一個嶄新的動作合成問題，根據上半身的動作軌跡產生出對應的下半身運動軌跡，並且讓這個動作在物理模擬引擎運算的時候可以令整個身體維持平衡，而為了解決這個問題，我們利用增強式學習，讓電腦能夠在不斷的進行嘗試之後找到一個最好的策略來應付不同的上半身運動軌跡。當一個上半身的運動軌跡進入到我們系統，首先會對這個只有上半身的動作進行一次簡單的物理模擬，藉此從中擷取出這個動作的一些特徵，再輸入事前已經訓練好的類神經網路模型，而這個網路模型所得到的響應就會是我們下半身相對應動作的特徵，之後再經過一個轉換演算法將下半身動作的特徵轉換成動作軌跡，將這個動作軌跡與原本的上半身動作軌跡結合起來同時驅動的話，就是我們最後所得到的動畫。利用增強式學習讓電腦自行學習出平衡的策略方法，可以讓我們不需要過度的花費時間在尋找及調整以往利用最佳化方法時所需的目標函數以及限制函數，並且這種方法也比較符合人類在初學一個新的動作時大腦及身體在運作的方式。

In this thesis we propose a new motion synthesis problem. For an upper body movement as input, system generates a corresponding lower body movement. When they animate at the same time in a physical simulation software, the human model should maintain body balance. To solve this problem, we try to use reinforce learning to let computer find the best control policy during iteratively testing and improving to adjust different upper body movement. When set an upper body movement as system input, first we execute a physical simulation for the upper-body-only movement to extract features of the movement. Then we pass these to a learned neural network model. The responses of the model is the features of corresponding lower body movement. We use a decoding algorithm to transfer output features to the lower body movement trajectory. In the last, we animate upper and lower body movement at the same time, it will be the final animation. This method makes us not to cost too much time in searching or revising objective and constraint function in traditional optimization methods. Also, the method is much like a human start to learn a new movement or skill.