neka-nat/cupoch

What it solves

Cupoch is designed to accelerate 3D data processing for robotics, moving computationally heavy tasks from the CPU to the GPU. It provides a high-performance alternative to libraries like Open3D for real-time robot systems that require rapid 3D computation for tasks such as SLAM, collision avoidance, and path planning.

How it works

The library leverages CUDA (and ROCm for AMD GPUs) to implement parallel processing of 3D data. It provides a suite of GPU-accelerated algorithms for point cloud registration, clustering, and filtering, as well as tools for visual odometry and occupancy grids. It also supports interoperability with DLPack, allowing it to work seamlessly with data structures from PyTorch and CuPy.

Who it’s for

Robotics engineers and researchers who need real-time 3D perception and navigation capabilities on NVIDIA or AMD GPUs, including those using Jetson Nano devices.

Highlights

• GPU-Accelerated 3D Processing: Implements KNN, ICP, and G-DBSCAN on the GPU for significant speedups over CPU-based libraries.
• Robotics-Specific Tools: Includes built-in support for collision checking, occupancy grids, and path planning on graph structures.
• Comprehensive 3D IO: Supports multiple file formats (pcd, ply, stl, obj, urdf) and ROS messages.
• Broad Hardware Support: Compatible with CUDA and ROCm (AMD), and provides specific installation paths for Jetson Nano.
• Interactive Visualization: Features an OpenGL CUDA interop GUI for rendering point clouds, meshes, and voxel grids.