Adaptive Critic Optimal Control of an Uncertain Robot Manipulator With Applications

Abstract

Realistic manipulation tasks involve a prolonged sequence of motor skills in varying control environments consisting of uncertain robot dynamic models and end-effector payloads. To address these challenges, this article proposes an adaptive critic (AC)-based basis function neural network (BFNN) optimal controller. Using a single neural network (NN) with a basis function, the proposed optimal controller simultaneously learns task-related optimal cost function, robot internal dynamics, and optimal control law. This is achieved through the development of a novel BFNN tuning law using closed-loop system stability. Therefore, the proposed optimal controller provides real-time, implementable, cost-effective control solutions for practical robotic tasks. The stability and performance of the proposed control scheme are verified theoretically via the Lyapunov stability theory and experimentally using a 7-DoF Barrett WAM robot manipulator with uncertain dynamics. The proposed controller is then integrated with learning from demonstration (LfD) to handle the temporal and spatial robustness of a real-world task. The validations for various realistic robotic tasks, e.g., cleaning the table, serving water, and packing items in a box, highlight the efficacy of the proposed approach in addressing the challenges of real-world robotic manipulation tasks.