Approaching Collaborative Manipulation by Pushing-Grasping Fusion

Smart manipulation is always the desired performance for the interesting researches in the field of robotic control. The complicated fusion among motion primitives could offer the advanced adaptions in presence of highly success rate or unknown objects. In this paper, a hierarchical framework for pushing-grasping fusion in the cluttered environment is demonstrated. Our method involves three-phase training process, integration of masks and the reinforcement learning scheme. Both simulation and experimentation are undertaken to validate the efficacy and feasibility of the proposed methodology. Our contributions in this work are (i) to propose both grasp mask and push mask for encouraging the model to focus on exploiting, adjusting the proper grasping posture in the desired target area as well as avoiding the phenomenon of the gripper slipping on the surface of an object, (ii) to recommend the reinforcement learning scheme without object model, and (iii) to release a hierarchical training method to enhance the interactive efficiency between grasping and pushing actions.