Cartesian Tasks Oriented Friction Compensation Through a Reinforcement Learning Approach

The paper describes an algorithm to compensate
for the friction in the robot joints, while executing a target
impedance controlled Cartesian task. The proposed method
relies on the reinforcement learning procedure: given a target
task in the Cartesian space and a joint space friction model,
the algorithm is capable to adapt the friction model parameters
based on a specified error function. The proposed error function
correlates the Cartesian position tracking error to the joint
space friction torques, allowing to independently learn the
friction parameters for each joint. In such a way, the friction
model parameters can be updated in subsequent iterations,
compensating for the friction effects. The proposed algorithm
has been validated through experiments. A target Cartesian
motion has been specified (such as in a pick and place operation)
and the proposed method has allowed to learn the friction
model parameters. A Universal Robot UR10 has been used as
a test platform, developing the impedance control with robot
dynamics compensation.