FOPID-Based Self-Balancing Robot Using Advanced Machine Learning

Abstract

PID precision is crucial for self-balancing robots, but conventional tuning methods struggle with uncertainties. This work employs Reinforcement Learning (RL) and Deep Neural Networks (DNNs) to adapt PID gains in real time. Using the Deep Deterministic Policy Gradient (DDPG) algorithm, the system continuously fine-tunes dip angle and angular velocity, enhancing stability and responsiveness. Simulations show that our RL-tuned PID outperforms fixed-gain controllers in stabilization, overshoot, and disturbance rejection. Additionally, Fractional Order PID (FOPID) is explored for greater control flexibility. Future research includes Proximal Policy Optimization (PPO) and real-world validation to strengthen the proposed strategy for adaptive and efficient robot control.