Advanced Input Shaping Control Based on Reinforcement Learning for SCARA Robot

초록

In this study, deep reinforcement learning-based input shaping technique (DRL-IST) is proposed to minimize the residual vibration in industrial robot manipulators with flexible joints. The proposed approach automatically tunes the input shaper parameters to enhance the vibration suppression during rapid point-to-point motions under varying operating conditions. The input-shaping design problem is formulated as a Markov Decision Process, in which a deep deterministic policy gradient agent learns an optimal policy to adapt the shaper parameters based on the measured end-effector acceleration. Unlike the conventional fixed-parameter input-shaping method, the proposed DRL-IST framework enables online adaptation without requiring explicit model identification or parameter estimation. A dynamic model of the SCARA robot was employed to define physically meaningful parameter bounds and to facilitate training in a simulation environment. The learned policy was then deployed for real-time control, providing fast and efficient parameter tuning. The effectiveness of the proposed method was validated through numerical simulations and experiments using a SCARA robot. Comparative results with conventional zero-vibration, zero-vibration and derivative, and extra-insensitive shapers demonstrate that the DRL-IST proposal significantly reduces residual vibrations, shortens the settling time, and improves robustness against modeling uncertainties and external disturbances. These results highlight the potential of reinforcement learning for adaptive vibration control in industrial robotic systems.

키워드