Impedance control is a very classic control method in robot and human interaction. Humans use their ability to adaptively adjust arm impedance parameters to perform well in a variety of tasks. This ability allows us to successfully complete interactive tasks even in uncertain disturbance environments. By analyzing the results of many force field experiments, We got a conclusion that humans usually use two methods to change their impedance in external perturbations: 1) In the situation of unpredictable perturbations, humans adapt their impedance through muscle contraction; 2) In the situation of predictable perturbations, humans adds a constant term to offset the known perturbations. In this paper, We show how 3-DOFs simulated robot use the reinforcement learning algorithm to perform similar behavior in both situations. We apply our model-free reinforcement learning algorithm PI2 (policy improvement with path integrals) to the robot to learn the end-effector variable impedance schedules and trajectories. Our findings provide a approach to automatically learn the impedance empirically without requiring to build a physical or environmental dynamics model.
Paper
Full text
Impedance Control in Uncertain Environment using Reinforcement Learning
Semantic Scholar · Engineering · 2019
Abstract
Impedance control is a very classic control method in robot and human interaction. Humans use their ability to adaptively adjust arm impedance parameters to perform well in a variety of tasks. This ability allows us to successfully complete interactive tasks even in uncertain disturbance environments. By analyzing the results of many force field experiments, We got a conclusion that humans usually use two methods to change their impedance in external perturbations: 1) In the situation of unpredictable perturbations, humans adapt their impedance through muscle contraction; 2) In the situation of predictable perturbations, humans adds a constant term to offset the known perturbations. In this paper, We show how 3-DOFs simulated robot use the reinforcement learning algorithm to perform similar behavior in both situations. We apply our model-free reinforcement learning algorithm PI2 (policy improvement with path integrals) to the robot to learn the end-effector variable impedance schedules and trajectories. Our findings provide a approach to automatically learn the impedance empirically without requiring to build a physical or environmental dynamics model.