Reinforcement Learning-based Control for Swing-up and Stabilization of an Underactuated Double Pendulum System

Reinforcement learning-based control for robots has been a hot research topic in recent years, as  it  has  demonstrated  potential  advantages  over  traditional  control  methods  for  nonlinear systems, including reduced reliance on model information and improved adaptability to model uncertainty and changes in control objectives. The double pendulum system, with its highly nonlinear dynamical model and chaotic behavior, serves as an ideal platform for developing new controllers and algorithms. This thesis aims to applya novel reinforcement learning-based control policyon the double pendulum systemthat achieves both swing-up and stabilization tasks. The proposed control  policywill first  be validated in  a simulation environment,  with subsequent  attempts  to  deploy  it  on the robot  by  directly  training  the  agent  on  hardware  to address the sim2real gap problem.