Accurate models begin with accurate measurement, so we capture human movement at two levels. Inside the body, our high-speed biplane fluoroscopy system tracks the 3D motion of the knee, ankle and foot bones at 100 Hz during treadmill walking. It shows the small abnormal motions that follow ligament injury, deformity and surgery. On the body surface, we use deformable human body models to account for skin and soft-tissue motion, which conventional methods treat as error. This improves both marker-based and markerless motion capture. Together, these two approaches give us the precise movement data our models are built on.

Many of the forces that injure joints cannot be measured in a living person. These include ligament tension, cartilage contact force and individual muscle forces. We build detailed musculoskeletal models of the lower limb, with anatomically realistic knee and foot joints driven by Hill-type muscles, and validate them against our in vivo measurements. With these models we estimate joint loads during walking and other daily activities. We also show how injury and deformity change those loads, and test surgical options such as ligament reconstruction, osteotomy and joint fusion in simulation. The same models give our movement controllers a body to learn in.

Highlight: Winner, ASME Grand Challenge to Predict In Vivo Knee Loads (2016)

A body model cannot move without a nervous system. We use deep reinforcement learning to train neural-network controllers so that musculoskeletal models with up to 150 muscles learn to walk and run. We then test them in conditions they have never seen, such as slippery or uneven ground. By building controllers with separate supraspinal and spinal reflex pathways and removing one at a time, we show how reflexes keep walking robust against sudden disturbances. The same simulations help explain why we move the way we do: why we tense opposing muscles together, and why we swing our arms when running.

Highlight: Winner, NeurIPS 2024 MyoChallenge (Locomotion)

An assistive device helps only when the body responds to it as intended. We study how devices that support movement, from powered exoskeletons to soft wearable suits, affect the way people move. This includes changes in gait and the forces passing through the soft tissue between the body and the device. Our experiments show that hip assistance changes gait in some common ways, but each person responds to assistance timing with their own motor strategy. Using our simulated humans, we develop and test control strategies that work with the body's own neuromechanics instead of against it. Our goal is assistance that truly benefits people in daily life.