Biomechanical Reinforcement Learning

Florian Fischer, Arthur Fleig, Miroslav Bachinski

Abstract: Biomechanical Reinforcement Learning (RL) describes the idea of training an RL policy that turns a static musculoskeletal model of the human body into a “simulated user” capable of interacting with a given task environment. By simulating interaction at the level of muscles and body dynamics, biomechanical RL can predict motion trajectories and physical effort for complex tasks, including keyboard typing, physical object manipulation, smartphone use, and VR interaction, offering enormous potential for automated testing of early-stage prototypes. In this course, we will explore the latest developments in biomechanical RL for HCI tasks. We will examine a biomechanical model of the shoulder, arm and hand in detail, as well as how it can be actuated. We will learn how to set up interactive task environments for various interaction tasks and contexts using a simple GUI, and take a brief look at the underlying Python code base. Finally, we will train simulated users, monitor the training process, and discuss how to best analyse, evaluate and debug biomechanical user simulations

 All speakers  Program