andyzeng/visual-pushing-grasping

Implements dual fully convolutional Q-networks that map RGB-D observations to pixel-wise action utilities for pushing and grasping, jointly trained via self-supervised Q-learning with rewards only from successful grasps. The approach integrates with V-REP/CoppeliaSim for simulation and runs on UR5 hardware, supporting both CPU and GPU acceleration (CUDA/cuDNN) with PyTorch backends. Learn synergies between complementary non-prehensile and prehensile actions through model-free trial-and-error, achieving generalization to novel objects within hours of training.

1,087 stars. No commits in the last 6 months.