NUST Institutions Library Catalogue

Balancing an inverted pendulum is a classic control problem that traditionally requires precise system modeling for effective controller design. Reinforcement Learning (RL) offers a model-free alternative but requires extensive training, which is impractical and risky when performed directly on physical hardware. Existing methods
typically rely on simulation environments built on accurate models, which are often
difficult to obtain. In this work, we use RL to balance flywheel inverted pendulum
by constructing an approximate model of the system through parameter estimation.
Despite its inaccuracies, the model proved sufficient for training RL agents in simulation. We developed a simulation environment based on the estimated model and
trained agents using Deep Q-Network (DQN), Proximal Policy Optimization (PPO),
and Discrete Soft Actor-Critic (SAC) algorithms. The trained policies were deployed
on real hardware without any additional fine-tuning. All agents achieved successful swing-up and stabilization, with SAC achieving the fastest swing-up time (1.65
s) and lowest steady-state error (0.0220 rad), demonstrating that RL can tolerate
model imperfections and still perform effectively on real systems.