Browsing by Author "Pan, Zherong"

Now showing 1 - 2 of 2
  • Thumbnail Image
    Item
    GPU-Based Contact-Aware Trajectory Optimization Using A Smooth Force Model
    (ACM, 2019) Pan, Zherong; Ren, Bo; Manocha, Dinesh; Batty, Christopher and Huang, Jin
    We present a new formulation of trajectory optimization for articulated bodies. Our approach uses a fully differentiable dynamic model of the articulated body, and a smooth force model that approximates all kinds of internal/external forces as a smooth function of the articulated body's kinematic state. Our formulation is contact-aware and its complexity is not dependent on the contact positions or the number of contacts. Furthermore, we exploit the block-tridiagonal structure of the Hessian matrix and present a highly parallel Newton-type trajectory optimizer that maps well to GPU architectures. Moreover, we use a Markovian regularization term to overcome the local minima problems in the optimization formulation. We highlight the performance of our approach using a set of locomotion tasks performed by characters with 15 − 35 DOFs. In practice, our GPU-based algorithm running on a NVIDIA TITAN-X GPU provides more than 30× speedup over a multi-core CPU-based implementation running on Intel Xeon E5-1620 CPU. In addition, we demonstrate applications of our method on various applications such as contact-rich motion planning, receding-horizon control, and motion graph construction.
  • Thumbnail Image
    Item
    Occluder Generation for Buildings in Digital Games
    (The Eurographics Association and John Wiley & Sons Ltd., 2022) Wu, Kui; He, Xu; Pan, Zherong; Gao, Xifeng; Umetani, Nobuyuki; Wojtan, Chris; Vouga, Etienne
    Occlusion culling has become a prevalent method in modern game engines. It can significantly reduce the rendering cost by using an approximate coarse mesh (occluder) for culling hidden objects. An ideal occluder should use as few faces as possible to represent the high-resolution input mesh with a high culling accuracy. We address the open problem of automatic occluder generation for 3D building models with complex topology and interior structures. Our method first generates two coarse sets of faces via patch-based and voxel-based mesh simplification techniques. A metric-guided selection algorithm chooses the best subset of faces to form the occluder, achieving a high occlusion rate and accuracy. Over an evaluation of 77 building models, our method compares favorably against state-of-the-arts in terms of occlusion accuracy, occlusion rate, and face number.