Variable Impedance Control for Force Tracking Based on PILCO in Uncertain Environment

Traditional impedance control is a simple and valid way for robot force tracking, but the uncertainty of the contact environment can seriously interfere with tracking accuracy. In this paper, we present a novel reinforcement learning variable impedance scheme based on PILCO algorithm, which trains a RBF policy network that dynamically adjusts the damping coefficient to compensate for environment uncertainty. Considering the randomness of environment and learning efficiency, a contact state transition model is established by Gaussian process regression, which can be used for state prediction and policy evaluation. The policy is then updated by a gradient-based approach. The simulation study indicates that our robot only takes 18 interactions with an unknown environment to learn an optimal variable impedance policy, which can be applied to various unknown contact environments and has better control accuracy than traditional methods.