This paper proposes an enhanced version of the integral sliding mode (ISM) control, where a deep neural network (DNN) is first trained as a deep reinforcement learning (DRL) agent. Then, such a DNN is fine-tuned relying on a Lyapunov-based weight adaptation law, with the aim of compensating the lack of knowledge of the full dynamics in the case of robot manipulators. Specifically, a DRL agent is trained off-line on a reward depending on the sliding variable to estimate the unknown drift term of the robot dynamics. Such an estimate is then exploited to initialize and perform the fine tuning of the online adaptation mechanism based on the DNN. The proposal is theoretically analysed and assessed in simulation relying on the planar configuration of a Franka Emika Panda robot manipulator.
Integral Sliding Modes Generation via DRL-Assisted Lyapunov-Based Control for Robot Manipulators
Sacchi, Nikolas;Ferrara, Antonella
2023-01-01
Abstract
This paper proposes an enhanced version of the integral sliding mode (ISM) control, where a deep neural network (DNN) is first trained as a deep reinforcement learning (DRL) agent. Then, such a DNN is fine-tuned relying on a Lyapunov-based weight adaptation law, with the aim of compensating the lack of knowledge of the full dynamics in the case of robot manipulators. Specifically, a DRL agent is trained off-line on a reward depending on the sliding variable to estimate the unknown drift term of the robot dynamics. Such an estimate is then exploited to initialize and perform the fine tuning of the online adaptation mechanism based on the DNN. The proposal is theoretically analysed and assessed in simulation relying on the planar configuration of a Franka Emika Panda robot manipulator.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
