Physically based simulation: learning from and interacting with reality

在过去的10年里，我一直致力于基于物理的计算机动画模拟。 这包括模拟控制人类的运动和抓取，物理启发的运动编辑工具，碰撞，接触，变形模型，以及最近的3D打印界面和机制。 虽然在过去十年中取得了令人难以置信的进展，但同样存在许多新的机会和重要的挑战。 该研究计划的目标是开发新的模型，配方和仿真技术，更好地匹配真实的世界的例子的行为，与使用这些模型在计算设计制造，控制和与真实的世界的互动的最终目标。 准确有效的柔顺接触，以及学习摩擦和变形的数据驱动模型，可以使3D打印机构的计算设计受益。 此外，从捕获中学习手部操作策略以创建灵巧的控制器是仿真和机器人技术中的一个持续挑战。 作为一个长期目标，该研究计划旨在帮助弥合计算机模拟与生活世界中材料，机制和机器人的实际行为之间存在的现实差距。我在研究中采取的一般方法是：（1）尽可能地识别和捕捉真实的世界的微妙之处;（2）设计有效地近似相关细节的模型;（3）开发有效的方法来模拟与这些自然现象的复杂相互作用。 在接下来几年的研究计划中，我还计划利用机器学习领域的最新进展。 函数逼近技术可以帮助捕捉材料的真实行为，深度网络似乎很有希望用于灵巧的操作，因为它们成功地推进了运动控制领域的最新技术。我的研究项目的产品将提供微小的调整以及基于物理的模拟的根本不同的方法。 结果导致在娱乐中的应用，例如，下一代视频游戏、电影、电视和新媒体的角色动画。在训练模拟、人体工程学、机器人、电子商务和康复方面也有重要的应用。

Over the last 10 years I have worked on physically based simulation for computer animation. This includes simulated control of humans for locomotion and grasping, physically inspired motion editing tools, models for collision, contact, deformation, and more recently 3D printed interfaces and mechanisms. While incredible progress has been made in the last decade, there are likewise many new opportunities and important challenges. The objective of this research program is the development of new models, formulations, and simulation techniques that better match the behaviour of real world examples, with the ultimate goal of using these models in computational design for fabrication, control, and interaction with the real world. Accurate and efficient compliant contact, and learning data driven models for friction and deformation can benefit computational design of 3D printed mechanisms. Furthermore, learning hand manipulation strategies from capture for the creation of dexterous controllers is an ongoing challenge in both simulation and robotics. As a long term objective, this research program aims to help bridge the reality gap that exists between computer simulation and the actual behaviour of materials, mechanisms, and robots in the living world. The general approach I take in my research is to (1) identify and capture the subtleties of the real world where possible, (2) design models that effectively approximate the relevant details, and (3) develop efficient methods for simulating complex interactions with these natural phenomena. In the next years of my research program, I also plan to exploit recent advances that have been made in machine learning. Function approximation techniques can help capture the true behaviour of materials, and deep networks appear to be promising for dexterous manipulation given how they have successfully advanced the state of the art in the domain of locomotion control. The products of my research program will offer both minor adjustments as well as fundamentally different approaches to physics based simulation. The results lead to applications in entertainment, for instance, character animation for next generation video games, movies, television, and new media. There are also important applications in training simulations, ergonomics, robotics, e-commerce, and rehabilitation.