基於動作序列學習的感知關聯性角色動畫

Abstract

物理模擬、資料驅動或是兩者混合的方法是近年來最普遍被應用來產生角色動畫的方 法。一個常見於基於物理模擬的動畫的問題是它通常需要輸入參考動作來決定模擬角 色的動作類型或形式。雖然資料驅動技術可以被應用來解決這類問題，此類技術的能 力仍嚴重受限於所用資料庫內運動的變化性。其主因在於資料驅動技術大多僅由資料 庫中合成動作，實際上並未對運動建模。即使目前有一些統計或是物理模型被提出來 協助資料驅動技術，這些模型並未具有物理上或是生物力學上的意義。現有電腦動生 成方法的另一關鍵問題是運動的產生與模擬角色的感知是無關的。然而，在現實世界 裡，人們通常會感受外界環境並調整自己的動作，這些調整有些甚至是無意識的預期 或預防動作。本計劃的目的即是為了要解決以上動作類型產生及缺乏感知關連的這兩 個問題。我們將利用生物力學分析設備蒐集人類運動資料，並研究應用動作技能學習 來產生角色動畫，以發展出與感知關聯的動作生成方法。此計劃所發展的方法，不僅 可產生物理上正確以及真實自然的人體動作。更重要的突破是我們的方法將不只是模 仿既有動作，更能藉由動作技能學習產生新的動作類型，並能利用模擬知覺產生更自 然的反應動作，例如平衡保持或是需高度協調性的運動動作，如揮棒、揮拍或是揮桿 的擊球動作等。我們相信，這些突破在未來將可對角色動畫的研究帶來重大貢獻。

Physics-based simulation, data-driven techniques, and their hybrid approaches have been the most popular way to generate character animation in recent years. A problem of physics-based animation is that reference motions are usually needed to determine the type and style of the motion of the simulated character. Although data-driven techniques have been applied to solve this problem, their capability is seriously limited by the varieties of motions in the database since they do not actually model motions but simply synthesize new motions from existing motion database. Even though there are some statistical models or physical models proposed recently to assist the data-driven techniques, these models do not represent motions in a physically or biologically meaningful way. Another critical problem of existing animation approaches is that motions are generated independent of perception; however, people perceive and adjust their motion in reality. This project aims to solve these two critical problems in character animation. We will investigate to develop character animation approaches that utilize motor skill learning and respect the effects of perception to motions. Our approaches will generate physically correct as well as realistically natural motions. More importantly, our approach can not only mimic a motion type but also create new motion types. By exploiting simulated perception, more natural responses can be generated, e.g., balance maintenance or highly coordinated sport motions. We believe our project will have great potential contributions to the research fields of character animation and humanoid robotics.