Learning Pick to Place Objects using Self-supervised Learning with Minimal Training Resources

Grasping objects is a critical but challenging aspect of robotic manipulation. Recent studies have concentrated on complex architectures and large, well-labeled data sets that need extensive computing resources and time to achieve generalization capability. This paper proposes an effective grasp-to-place strategy for manipulating objects in sparse and chaotic environments. A deep Q-network, a model-free deep reinforcement learning method for robotic grasping, is employed in this paper. The proposed approach is remarkable in that it executes both fundamental object pickup and placement actions by utilizing raw RGB-D images through an explicit architecture. Therefore, it needs fewer computing processes, takes less time to complete simulation training, and generalizes effectively across different object types and scenarios. Our approach learns the policies to experience the optimal grasp point via trial-and-error. The fully conventional network is utilized to map the visual input into pixel-wise Q-value, a motion agnostic representation that reflects the grasp's orientation and pose. In a simulation experiment, a UR5 robotic arm equipped with a Parallel-jaw gripper is used to assess the proposed approach by demonstrating its effectiveness. The experimental outcomes indicate that our approach successfully grasps objects with consuming minimal time and computer resources