Reinforcement Learning for Musculoskeletal Control with an OpenSim Model

Abstract: Simulations of the human Musculoskeletal system can help in treatment of injuries, planning surgeries and prosthesis design. OpenSim provides a freely available open source software for the development of Musculoskeletal  models and creating dynamic simulations of movement. This enables the learning of control and activations of the Musculoskeletal  system with  modern optimization  methods. The use ofReinforcement Learning allows for direct control of activations via communicated actions. This thesis aims at demonstrating an implementation of a Deep Reinforcement Learning approach called Proximal Policy Optimization (PPO) to control muscle activation of an OpenSim model with one active muscle. Muscle activations arelearned given current position and velocity as well as target position and velocity. The results show a PPO-approach to muscle control of an OpenSim model that can be built upon for more advanced use with several active muscles and training withparallel environments .