Stiffness Control for a Soft Robotic Finger based on Reinforcement Learning for Robust Grasping

Abstract

Soft robotic grippers have attracted a great attention for industry application. However, due to the characteristics of non-linearity, it is challenging to control the stiffness of soft robotic grippers. As some of the controlling methods of the soft gripper’s stiffness are developed by physical models which are based on force analysis with strong assumptions, it can be difficult to apply to real-life scenarios. Recent efforts to resolve the problem of soft grippers’ stiffness control based on reinforcement learning might become a potential resolution to this challenge. In this paper, a model-free reinforcement learning based on Deep Deterministic Policy Gradient (DDPG) is trained for controlling the stiffness of the soft gripper with rotation jamming layers. The control strategy is validated in the simulation, the best results of Error <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">i</inf> (i=1, 2, 3, 4) reported in this paper are 11.69%,0.57%,0.18%,0.30%. Compared with the classic PID, DDPG has smaller errors and higher stability, which improves the grasping robustness. The results show that the proposed controlling method is effective and superior, which can ensure grasping robustness during high-speed pick-and-place manipulation.