New AI framework personalizes prosthetic legs with 42–59% better biomimicry

A team led by Duong Le at the University of Michigan has introduced a replay-constrained simulation framework that uses deep reinforcement learning to personalize impedance controllers for powered knee-ankle prostheses. The framework leverages a MuJoCo-based simulator that replays recorded hip kinematics and feedback-based ground reaction forces from individual walking data — avoiding the need to model the complex neuromuscular system of the human user. This design allows a deep RL policy to simultaneously tune phase-dependent stiffness, damping, and equilibrium angles for both knee and ankle joints, maximizing a biomimicry-based reward computed solely from onboard prosthesis sensors.

In experiments with three participants with transfemoral amputation walking on level ground at 0.8 m/s, the framework demonstrated strong predictive validity: simulation results correlated with hardware performance at Pearson r values between 0.96 and 0.997. The best-performing policy on hardware was consistently among the top five simulation policies for all participants. Learned controllers improved overall biomimicry rewards by 42–59% relative to an unpersonalized baseline. The approach supports high-dimensional personalization and can be extended to neural-network-based controllers, offering a scalable path toward individualized prosthetic leg tuning without time-intensive human-in-the-loop iterations.