ID 2365: Deep Reinforcement Learning for Musculoskeletal Models
The goal of this thesis is to find a controller that can reproduce realistic and robust standing / walking behaviour of a neuromusculoskeletal model based on physiological feedback. In a first step, this should be done using imitation learning from data that we recorded in a previous study. If that works, the next idea would be to incorporate a higher level controller that computes the best action from multiple sub-policies which capture different movement strategies of humans.
Implementation of an imitation-based reward function for standing and walking for deep reinforcement learning.
Implementation of a multi-level control structure for human movements using deep reinforcement learning
Clear and structured code documentation
Experience in Reinforcement Learning / Deep Learning
Independent working style