Imitation Learning based Soft Robotic Grasping Control without Precise Estimation of Target Posture

In this paper, we proposed the implementation of an imitation learning algorithm to support a simplified control scheme of a grasping task for a soft gripper. For this purpose, we combined an instance segmentation algorithm, such as state of the art Mask RCNN, for the object localization in the neural network architecture. The proposed scheme is based on a combination of scene features mapping and object localization with deep learning, which supports grasping objects regardless of target object pose. As a result, the proposed system exploits soft gripper's advantages; such as compliance with the target object shape. We compare the performance of the model to the expert demonstrations use to train the imitation learning algorithm. To this end, we changed the configuration of the environment position, pose, and shape of three different target objects, in which the system shows high performance following expert trajectories. Additionally, we tested the grasping success rate in random environment configurations.