Stabilization of a Two-Wheeled Self-Balancing Robot Using Reinforcement Learning

This work deals with the stabilization of a two-wheeled self-balancing robot (TWSBR) and its displacement from one reference point to another by implementing a reinforcement learning (RL) scheme in conjunction with a proportional-derivative (PD) control. Conventional control schemes require either precise or partial knowledge of the system, or numerous computations. The proposed RL-based framework does not depend on any a priori knowledge of the system model. This RL-based control strategy is contrasted with a proportional-integral-derivative (PID) controller and an adaptive neural network (ANN) controller. Simulation results from a virtual prototype for a TWSBR, which captures the mechanical and dynamic characteristics from the real system, demonstrate that the RL-based framework exhibits an excellent performance concerning to the displacement of the TWSBR, even when an external disturbance affects to the system.