CAREER: Precomputing Data-driven Deformable Systems for Multimodal Interactive Simulation

点击翻译按钮获取中文摘要

James, DougCarnegie Mellon UniversityCCF-0347740Deformation phenomena are a rich and integral part of our lives; it affects our appearance (e.g., skin, hair), the sounds we make (e.g., clothing, vocal cords), beauty in nature (e.g., a forest blowing in the wind), and important decisions (e.g., medical planning). Deformation modeling has made enormous progress, but we still desire increasingly realistic levels of geometric and dynamic complexity. Deformation also plays an important role in multimodal feedback for virtual environments: feeling contact forces using haptic interfaces, seeing realistic deformable appearances with shadows and interreflections, and hearing contact sounds. Not only are accurate deformations required at high rates, but the phenomena can also be extremely complex.Given the conflicting demands of real-time interactivity and complex large-scale simulation, one logical algorithmic strategy is to do as much work ahead of time as possible. Exploiting massive precomputation and data-driven tabulation is appealing, but how this can be done, and to what extent, remains a mystery for most nonlinear systems. Nevertheless, preliminary evidence suggests that potential precomputation speedups are easily a million-fold. By researching precomputation techniques today, we stand to exploit million-fold speedups for multimodal simulation on supercomputers of tomorrow.Our scientific objective is to understand how to systematically precompute, simulate and experience data-driven models of large-scale nonlinear deformable systems. We address these goals in three ways: (1) precomputation techniques for data-driven deformable models based on reduced coordinate representations; (2) deformable motion databases of complex motions for real-time ``playback;'' (3) multimodal aspects of interactive simulation, including haptic rendering of contact forces, data-driven deformable sound models, and precomputed deformable object appearance models. Our research is driven by real-world applications, and will be explored in the broad context of interactive computer animation, virtual medicine, robotics and manufacturing, and simulation of large-scale natural environments.