Kaolin is a PyTorch library that provides GPU-optimized modules for accelerating 3D deep learning research.
The library addresses the fragmentation of 3D deep learning by packaging reusable building blocks from NVIDIA research into a cohesive API. It offers representation-agnostic physics simulation through Simplicits that works across meshes, splats, and point clouds with collision support. The tool includes differentiable rendering implementations, native support for 3D Gaussian splats with PLY and USD I/O, an octree acceleration structure called Structured Point Clouds, and utilities for converting between 3D representations, performing quaternion operations, and handling USD file I/O.
Kaolin suits researchers and practitioners building 3D deep learning systems who need production-grade, GPU-accelerated components rather than implementing these primitives from scratch. It is particularly valuable for projects involving neural rendering, 3D scene understanding, physics-informed learning, or interactive 3D applications. The library's representation-agnostic approach to physics means code written for meshes can work with point clouds or splats with minimal changes, reducing friction when experimenting with different 3D representations.
The project maintains active development with ongoing feature additions. Recent work includes a web client-server framework for rapid prototyping of interactive browser interfaces over 3D research, available on a prerelease branch. The team is developing capture-to-simulation pipelines that combine 3D Gaussian splats with physics prediction and mixed splat-mesh simulation, along with custom USD physics schemas for sharing results. The project demonstrates responsiveness to emerging research directions, particularly around Gaussian splatting and interactive visualization patterns distilled from production tools.
