graspnet/graspnet-baseline

Implements 6-DOF grasp detection using PointNet2 backbone with point cloud processing from RGB-D inputs, incorporating single-view collision detection as post-processing. The architecture uses custom CUDA operators (PointNet2 and KNN) for efficient 3D feature extraction and supports multi-camera evaluation across RealSense and Kinect sensors with tolerance-based quality metrics. Training and inference integrate TensorBoard logging and the GraspNetAPI for standardized benchmark evaluation.

893 stars. No commits in the last 6 months.