A Data-Driven Learning-From-Demonstration Framework for Robotic Grasping

Autonomous grasp planning remains a key chal- lenge in robotic manipulation, particularly in unstructured environments where object types, poses, and arrangements vary. This work presents a data-driven grasp planning method for a robotic manipulator tasked with clearing a table containing diverse objects. The method encodes human-demonstrated grasping strategies by representing Cartesian trajectories with Dynamic Movement Primitives (DMPs), whose parameters are predicted by a neural network from grasp-specific inputs. A second neural network estimates feasible grasp poses based on the object pose estimate data, which is used as the new goal parameter for the grasp trajectory generation. To reduce demonstration effort, synthetic datasets are gen- erated via data augmentation of the recorded trajectories. The approach is implemented on a real system and evaluated on four objects with varying shapes and sizes. The experiments show a high success rate for grasping as well as the ability to incorporate new objects into the system through minimal additional effort.