Implementation of 56 Transition Control of a Triple Inverted Pendulum Using Sim-to-Real Reinforcement Learning

This paper proposes the implementation of equilibrium-to-equilibrium transition control for a triple inverted pendulum system using Sim-to-Real reinforcement learning. Recently, multi-link inverted pendulum systems have introduced the new control challenge, extending beyond conventional swing-up and balancing controls toward equilibrium-to-equilibrium transition control. Transition control, which involves continuous transitions between multiple unstable equilibrium points, is particularly sensitive to disturbances. To address this, we apply the Sim-to-Real reinforcement learning technique, transferring control policies learned in simulation to the physical system. Furthermore a triple inverted pendulum system with high model consistency was designed and constructed to minimize the reality gap between simulation and physical environments. The proposed controller successfully achieved all 56 possible transitions among the eight defined equilibrium points. The results demonstrate that transition control based on Sim-to-Real reinforcement learning effectively resolves complex nonlinear control problems.