Robotic Deep Rolling with Iterative Learning Motion and Force Control

Large industrial robots offer an attractive optionfor deep rolling in terms of cost and flexibility. These robotsare typically designed for fast and precise motion, but maybe commanded to perform force control by adjusting theposition setpoint based on the measurements from a wrist-mounted force/torque sensor. Contact force during deep rollingmay be as high as 2000 N. The force control performance isaffected by robot dynamics, robot joint servo controllers, andmotion-induced inertial force. In this paper, we compare threedeep rolling force control strategies: position-based rolling withopen-loop force control, impedance control, and gradient-basediterative learning control (ILC). Open loop force control iseasy to implement but does not correct for any force deviation.Impedance control uses force feedback, but does not track wellnon-constant force profiles. The ILC augments the impedancecontrol by updating the commanded motion and force profilesbased on the motion and force error trajectories in the previousiteration. The update is based on the gradient of the motionand force trajectories with respect to the commanded motionand force. We show that this gradient may be generatedexperimentally without the need of an explicit model. Thisis possible because the mapping from the commanded jointmotion to the actual joint motion is nearly identical for alljoints in industrial robots. We have evaluated the approachon the physical testbed using an ABB robot and demonstratedthe convergence of the ILC scheme. The final ILC trackingperformance of a trapezoidal force profile improves by over70 % in terms of the RMS error compared with the impedance controller.