A data-driven reinforcement learning framework for optimal balance control and stabilization of a rotary inverted pendulum

The rotary inverted pendulum is a nonlinear, underactuated and inherently unstable benchmark system widely used for evaluating advanced control strategies. This work presents a data-driven reinforcement-learning framework for upright balance control and stabilization of the pendulum. The system is represented using four state variables: the rotary-arm angle, pendulum angle, rotary-arm angular velocity and pendulum angular velocity. A twin delayed deep deterministic policy gradient (TD3) agent is employed to search the continuous state-feedback gain space while preserving the conventional control structure. Candidate gain sets are evaluated using a combined criterion based on cumulative reward and reconstruction accuracy of the pendulum response. The selected gains are assessed under nominal operating conditions, external disturbances and measurement noise. The results demonstrate effective upright stabilization, improved damping and satisfactory disturbance-rejection capability. The study shows that TD3-based gain selection can provide a transparent and data-driven alternative to manual gain tuning for nonlinear and underactuated systems. The proposed framework is therefore relevant to applications such as self-balancing platforms, robotic manipulators and other autonomous control systems.