Effective hybrid force/position control of robots using reinforcement learning

Abstract

This research proposes a novel hybrid force/position control (RL-HC) approach for robotic systems, utilising reinforcement learning. The proposed controller demonstrates superior robustness and performance, particularly in tracking the end-effector's position and contact force, compared to the hybrid sliding mode force/position control (HSMC) method, even in the presence of structural and non-structural uncertainties. The RL-HC controller is designed around a fixed-time sliding mode model, integrating force and position control components. The stability of this controller is established using the Lyapunov theory. Additionally, the reward function within the reinforcement learning network is carefully crafted to align with key objectives, including minimising chattering, force error, position error, and control effort. A simulation performed using a 3-DOF Delta Robot illustrates the effectiveness of the RL-HC approach. Results indicate that RL-HC outperforms traditional methods, showcasing better performance and robustness when facing various external disturbances and uncertainties. Specifically, the findings highlight a significant reduction in position error, force error, total control effort, and chattering. The study also illustrates how different reward function designs impact the achievement of the desired objectives. © 2025 Informa UK Limited, trading as Taylor & Francis Group.