This paper aims to design a deep reinforcement learning (DRL)-based variable impedance control policy that supports stability analysis for robot force tracking in complex geometric environments.
The DRL-based variable impedance controller explores and pre-learns the optimal policy for impedance parameter tuning in simulation scenarios with randomly generated workpieces. The trained results are then used as feedforward inputs to improve the force-tracking performance of the robot during contact. Based on Lyapunov’s theory, the stability of the proposed control policy is analysed to illustrate the interpretability of the results.
Simulations and experiments are performed on different types of complex environments. The results show that the proposed method is not only theoretically feasible but also has better force-tracking effects in practice.
Compared with most other DRL-based control policies, the proposed method possesses stability and interpretability, effectively avoids the overfitting phenomenon and thus has better simulation-to-real deployment results.
