visual-pushing-grasping

Why we included this project

For roboticists working on pick-and-place, this is the PyTorch reference implementation behind an IROS 2018 paper on learning pushing and grasping together. Instead of hand-coded control logic, two fully convolutional networks take RGB-D images and learn through trial and error in a Q-learning framework, so a UR5 arm gradually figures out how to push cluttered objects into clearer positions and then grasp them. The repo includes training and testing scripts that run both in simulation and on real hardware, which makes it a practical entry point for experimenting with self-supervised manipulation. Because the published paper includes its results and the code matches it, the project also works as a reproducible baseline for comparing a new learning-based manipulation approach.