Grasping preprint reports over 90% success after 50 online attempts

A robotics preprint reports a grasping system that reaches over 90% success on several challenging object categories after only 50 online grasp attempts. The researchers evaluated it in simulation and real-world experiments comprising over 1500 grasp trials. The system adapts through stored experience rather than fine-tuning a large model. Grasp outcomes update future scores, while optional user demonstrations supply candidate grasps that can transfer to new scenes. It matches existing baselines before adaptation and improves on unfamiliar objects. The demonstrated setup uses a parallel-jaw gripper, a single view of the scene and cluttered environments.