Learning Stable State-Dependent Variable Impedance Control for Compliant Manipulation

State-dependent variable impedance control (VIC) has been shown as the necessary skill for compliant and dexterous robot operations. However, introducing a variable impedance law will raise the stability issue of the robot system, which introduces potential safety risks. Unfortunately, there is not yet a generally effective framework to handle this issue, i.e., existing approaches achieve stability by sacrificing the VIC performance. This article presents a learning-based VIC approach to maintain stability and control performance simultaneously. Specifically, we design a neural energy function called SNEUM, which has a globally unique minimum. Then, the SNEUM is used to encode the variable stiffness behaviors to result in stable and learnable VIC structures. Due to the special form of the SNEUM, we show that the control modes focusing on control smoothness or accuracy could be mathematically determined by tuning a hyper-parameter of the SNEUM. Various comparative experiments are conducted to show the effectiveness of the proposed approach.