Deep Reinforcement Learning Achieves 100% Grasp Success on Objects Previously Deemed Ungraspable

A new framework employing Deep Q-Networks (DQN) and keypoint-based object representations has achieved a remarkable 100% success rate in grasping objects previously considered ungraspable. This system operates using 2D overhead images in a simulated environment, where a geometric algorithm generates initial grasp candidates. Testing involved 300 objects from the Dex-Net dataset, utilizing a UR5 manipulator. The approach has been further validated in real-world applications through a Delta parallel robot. The detailed findings are published in a paper available on arXiv under the identifier 2608.17628.