Subspace Hierarchies for Musculoskeletal Control

Musculoskeletal systems are responsible for the rich and diverse ranges of motion and behaviors exhibited by all sorts of animals, including humans. Recreating these motions is key for applications in medical simulation, robotics, and virtual reality. Unfortunately, modeling even simple control tasks such as grasping or locomotion with such musculoskeletal systems is extremely difficult. The key problem is that faithfully modeling the dynamics of the musculoskeletal anatomy requires extremely high resolution detailed deformable geometry. This not only makes simulating such systems slow, but it also makes the optimal control task ill-determined. This in turn makes the numerical method used to solve any optimal control problem at hand struggle to converge reliably to a viable solution.
To solve this problem, we draw inspiration from the model reduction community, which simplifies the dynamics of a system by only modeling its most dominant modes. By only modeling a small number of degrees of freedom, and then gradually increasing that number, we aim to construct a reliable optimal control solver for arbitrarily complex musculoskeletal control problems.

Faculty Supervisor:

Eitan Grinspun

Student:

Partner:

ETH Zurich

Discipline:

Computer science

Sector:

Education

University:

University of Toronto

Program: