GraspNet Baseline

This repository contains code for training and running a machine learning model that teaches robots how to pick up objects. The model looks at 3D camera images of a pile or shelf of items and figures out the best places to grab each object—essentially solving the problem of "where should a robotic hand grip this thing so it doesn't drop it?"

The model works by analyzing depth and color images captured from cameras (like those used in Kinect or RealSense devices). It processes this visual information to generate a ranked list of possible grasp points on each object, scored by how likely each grip is to succeed. The system can then filter out grips that would collide with other objects or the table, leaving only safe, executable picks. The code includes everything needed: the training pipeline to teach the model on a large dataset, testing tools to evaluate accuracy, and a demo script you can run on your own images.

This is most useful for roboticists, researchers, and companies building picking robots—think warehouse automation, bin-picking systems, or robotic arms in manufacturing. Instead of hand-coding rules about how to grip different shapes, teams can use this pre-trained model or fine-tune it on their own data. The repo provides pretrained weights already optimized for two common camera types (RealSense and Kinect), so you can start with working models rather than training from scratch.

The project is grounded in academic research (published at CVPR 2020) and comes with a large standardized dataset of over a billion grasp annotations to validate results. The README includes straightforward setup instructions, links to download both the trained models and any data you might need, and clear examples of how to train, test, and demo the system. It's designed as a baseline for others to build on—a solid starting point rather than a finished product.