Energy-Based Model Predictive Control and Reinforcement Learning for Double Inverted Pendulum Stabilization and Obstacle Avoidance

This study presents a comparative investigation of Model Predictive Control (MPC) and Reinforcement Learning (RL) strategies for the swing-up, stabilization, and obstacle avoidance of a double inverted pendulum (DIP) mounted on a cart. The MPC framework employs a direct orthogonal collocation method, incorporating both state-tracking and energy-based cost functions to ensure efficient convergence under strict input constraints. An energy-based formulation significantly improves the feasibility of swing-up maneuvers while maintaining stability during set-point transitions and satisfying obstacle avoidance constraints. The RL approach implements the Soft Actor-Critic (SAC) algorithm within a modified OpenAI Gym environment, leveraging entropy regularization for improved exploration. While MPC consistently achieves swing-up and obstacle avoidance with high accuracy, the RL method effectively stabilizes the pendulum but exhibits limited performance in swing-up tasks, indicating the need for task-specific reward design and parameter tuning. Simulation results highlight the trade-offs between model-based and model-free control in terms of computational cost, robustness, and learning efficiency, offering valuable insights for advanced nonlinear control in robotics and automation.

کلیدواژه ها:

Double inverted pendulum ، Model Predictive Control ، Reinforcement Learning ، Soft Actor-Critic