Learning control for task specific industrial robots

Iterative learning control (ILC) is a strategy that allows a control system to improve its performance by making use of the error signals collected from previous iterations. A prerequisite of using ILC is that the output reference has to be repetitive from trial to trial. A full run of ILC training (taking non-negligible time) is needed when there exist small changes in the reference signal. This paper introduces a new approach to extrapolate the converged ILC policies to previously unseen tracking problems. A time-frequency domain mapping is constructed to approximate the ILC policy for a group of trajectories used in a particular task, say spot welding. We also introduce the idea of feature-frequency space, where the ILC policies from different trajectories can be encoded into a single model. This model can generate a control policy that performs comparably to the ILC policy while having the advantage of not requiring a full training for a new trajectory. The proposed method implemented on a FANUC R-2000iC robot achieved 31.6% of vibration reduction whereas the standard ILC (i.e., with a full training for each particular trajectory) achieved 34.6% of vibration reduction.