Compliant robot arms are difficult to control because their elastic joints introduce nonlinear dynamics and model uncertainty. Traditional iterative learning control (ILC) can achieve high precision while preserving passive compliance, but it must be re-run for every new trajectory. Reinforcement learning (RL) generalizes across trajectories but typically needs many samples and delivers limited tracking accuracy.
The proposed reinforced iterative learning control (RILC) runs both controllers concurrently. ILC supplies high-quality samples, while RL acts as a best-effort policy that generalizes across iterations. The authors provide a proof that the RL component does not destabilize ILC and, under certain conditions, can asymptotically reduce the error bound.
Experiments on an underactuated four-degree-of-freedom serial elastic robot, using different trajectories and modeling mismatches, showed that RILC overcomes the slow adaptation and high initial errors of standard ILC. The paper is accepted for publication in Control Engineering Practice.