Farbod Fahimi and Susheel Praneeth
Control of mobile robots, trajectory tracking, reinforcement learning,adaptive critic design, optimal control
A universal trajectory tracking optimal controller is introduced that can be used without any modifications for all differential drive robots that have wheel speed controllers. To eliminate the need for dynamic model formulation, and especially system identification, a model- free online reinforcement learning (MFORL) dynamic controller is introduced in this paper. The implementation of the MFORL uses the Standard Lyapunov Extension Theorem on a specially defined Lyapunov function, which not only tests the stability of the tracking error but also includes the convergence error of the neural networks (NNs) that estimate the actor and the critic. In this way, the universal trajectory tracking controller guarantees that (1) the optimal control law will be learned in real-time (i.e., the actor and critic NNs converge in real-time) and (2) the learned control law stabilises the robot (i.e., tracking errors are bounded and can be made arbitrarily small).
Important Links:
Go Back
