Simulation of Adaptive Running with Flexible Sports Prosthesis using Reinforcement Learning of Hybrid-link System

Researchers Yuta Shimane and Ko Yamamoto have published a novel simulation framework that uses reinforcement learning (RL) to model how athletes with transtibial amputations run with flexible, leaf-spring-type sports prostheses. The core innovation is a 'hybrid-link system' that integrates a Piece-wise Constant Strain model. This allows the simulation to accurately represent the complex, flexible deformation of the prosthesis during high-impact activity, which is a significant challenge for traditional rigid-body dynamics models. The methodology combines imitation learning—trained on real human motion capture data—with precise computational dynamics to generate realistic, whole-body running motions.

The study's application demonstrates its practical value. The team simulated running under various virtual prosthetic stiffness conditions and analyzed the resulting metabolic cost of transport—a key performance metric. Their findings suggest that optimal stiffness is not one-size-fits-all and significantly influences running economy. This moves prosthesis design beyond physical trial-and-error, enabling engineers and clinicians to virtually test and tailor prosthetic properties for individual users before manufacturing. The approach promises to accelerate the development of high-performance, user-specific athletic prosthetics by providing data-driven insights that were previously difficult or impossible to obtain.