Differentiable Position-Based Simulation of Compliant Constraint Dynamics
本文提出了一种全面可微分的求解器，可以通过计算目标函数的梯度来获得与所需参数相关的梯度，从而实现目标函数的最小化。作者通过优化身体形状和姿势，仅通过观察服装，优化时间变化的外部力序列，以及弹性布料和体积材料参数估计等示例，展示了该方法的有效性和高效性。
Paper Reading List
Differentiable Physics Simulations with Contacts: Do They Have Correct Gradients w.r.t. Position, Velocity and Control?
本研究探讨了可微分物理模拟中不同接触模型对位置、速度和控制的梯度计算的正确性，并通过比较模拟计算的梯度与解析梯度的结果，揭示了当前可微分模拟器的局限性，并为开发更可靠的物理模拟器提供了新的研究机会。
Multi-Layer Thick Shells
实现了Thick Shell的模拟，实现了一定的褶皱，并且提出了一个交替方法来快速求解。
- Reduced Prism Elements：三棱柱+中轴面建模 Thick Shell ⇒ 解决 Shell Locking
- Complementary Wrinkle Coupling：褶皱耦合系统
- Alternating Minimization：交替最小化，先固定低频信息（PN+Cholesky）、再处理高频信息（PN+CG）。
例子：瑜伽垫、皮衣、记忆枕等
∇Sim: Differentiable simulation for system identification and visuomotor control
本文提出了一种名为 ∇Sim 的可微分仿真框架，通过结合可微分多物理仿真和可微分渲染，从图像序列中直接估计物体的物理属性，解决了系统识别问题和视觉运动控制问题。
ChainQueen: A Real-Time Differentiable Physical Simulator for Soft Robotics
本文提出了一种用于软体机器人的实时可微分物理模拟器ChainQueen，基于Moving Least Squares Material Point Method (MLS-MPM)，能够高精度地进行前向模拟和反向梯度计算，并成功应用于软体机器人的控制任务。
Single-Level Differentiable Contact Simulation
本文提出了一种单层可微分刚体接触动力学的表达方式，用于表示由凸多面体组成的物体和机器人。通过将接触仿真和碰撞检测结合为一个统一的单层优化问题，解决了传统物理引擎和最新的优化方法在现实接触仿真场景中的问题。通过在机器人操作任务中的应用，证明了该方法的可行性和优越性能。
NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis
Basic Information
- Title: NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis
- Authors: Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, Ren Ng
- Affiliations: University of California, Berkeley; Google Research
- Publication Date: 2020
- Publication Venue: ECCV 2020
Introduction
This paper introduces Neural Radiance Fields (NeRF), a novel approach to synthesizing photorealistic images from novel viewpoints by learning volumetric scene functions of complex 3D scenes. The method significantly advances the state-of-the-art in view synthesis, demonstrating the ability to produce highly detailed and coherent renderings from a sparse set of input photographs.
Details
- Methodology: NeRF models the scene as a continuous 5D function (spatial location and direction) using a deep neural network. It learns to map these 5D coordinates to color and density, which are then used to render images via differentiable volume rendering.
- Techniques: The paper details the use of positional encoding to allow the model to learn high-frequency details more effectively and a hierarchical sampling strategy to improve rendering efficiency.
- Experiments: The authors conducted extensive experiments on both synthetic and real-world datasets, comparing NeRF's performance against several baseline methods in terms of photorealism and accuracy.
Conclusion
NeRF represents a significant step forward in 3D scene representation and rendering, offering unprecedented detail and realism. The approach has potential applications in virtual and augmented reality, film production, and virtual tourism. The paper suggests areas for future research, including improving computational efficiency and extending the method to dynamic scenes.
This summary follows the guidelines specified, aiming to provide a clear overview of the paper's content and contributions.
DiffTaichi: Differentiable Programming For Physical Simulation
我们介绍了 DiffTaichi，这是一种新的可微编程语言，专为构建高性能可微物理模拟器而设计。基于命令式编程语言，DiffTaichi 使用源代码转换生成仿真步骤的梯度，从而保持算术强度和并行性。使用轻量级Tape记录整个仿真程序结构，并以相反的顺序回放梯度内核，以实现端到端的反向传播。我们在 10 个不同的物理模拟器上展示了我们的语言在基于梯度的学习和优化任务中的性能和生产力。例如，用我们的语言编写的可微弹性对象模拟器比手工设计的 CUDA 版本短 4.2×但运行速度却一样快，比 TensorFlow 实现快 188 ×。使用我们的可微分程序，神经网络控制器通常只需数十次迭代即可进行优化。
Efficient Differentiable Simulation of Articulated Bodies
本文提出了一种高效可微分的关节体模拟方法，通过使用空间代数和伴随方法推导出前向动力学的梯度，实现了对关节体动力学的深度学习框架集成和基于梯度的神经网络优化。
Differentiable Simulation of Soft Multi-body Systems
本文提出了一种可微分仿真软多体系统的方法，通过将物理动力学与基于梯度的流程相结合，实现了可微分仿真。通过引入新的矩阵分解策略，我们设计了一种自上而下的矩阵组装算法，并推导出了适用于软连续体的广义干摩擦模型。