Combining data and first principles in the control of muscle-redundant systems in Biomechanics and Robotics - exploring the contribution of prior knowledge through morphology variation

Muscle-driven biological systems move with grace and ease in amazingly different types and forms. The motion is robust and stable in different conditions interfacing with the environment. Simulating musculoskeletal system models reveals that the mapping from muscle stimulations to movement dynamics is highly nonlinear and complex. This makes it difficult to control those systems with classical techniques. In previous studies we were able to synthesise controllers for arm and leg motion in simulation and biomimetic robots using appropriate heuristic and optimisation techniques as well as MPC. Additionally, we showed that biology exhibits properties, which are favourable for learning the controller, autonomously. In this continuation project, our aim is to integrate different levels of prior controller model knowledge into the biological control hierarchy to mimic human motion for various simple movements both with arm and leg models in simulation and with muscle-driven robots in real world experiments. The leading research question is as to whether and how the design of the biological motor system facilitates the control task as well as the learning task to control the movement.