The Quiet Shift in 'Personalized' Prosthetics: Replay the Patient's Walk, Train a Controller Against It

The hard part of a powered prosthetic leg isn't the hardware. It's that every above-knee amputee walks differently, so each controller has to be tuned to that one person by hand in a clinic — hours of trial and error.

A new University of Michigan preprint sidesteps that bottleneck with a counterintuitive move. Instead of teaching a simulator how a human body works, the simulator replays what this specific amputee's hip already did on a recorded walk. A learned controller is then trained against that replay.

The mechanism matters. Rather than modeling the unmodelable — the human neuromuscular system — the framework treats the patient's own recorded motion as a constraint. The simulator reproduces the prosthetic's knee and ankle dynamics around it. A deep reinforcement learning policy personalizes phase-dependent stiffness, damping, and equilibrium angle at both joints simultaneously, maximizing a biomimicry reward computed solely from onboard measurements.

On three participants with transfemoral amputation walking on level ground at 0.8 m/s, the result was a 42–59% improvement in biomimicry rewards over the unpersonalized baseline. Sim-to-hardware predictive validity held: Pearson r = 0.96–0.997 across participants, and the best-performing policy on hardware was consistently predicted within the top five simulation policies.

The scope is narrow. Three amputees, level ground, one fixed cadence. Slopes, stairs, and variable cadence remain unaddressed. The paper is a preprint, not peer-reviewed, and its authors flag the framework as amenable to higher-dimensional controllers such as neural networks — a stepping stone rather than a finished clinic workflow.

The structural claim still holds: when modeling the body is too hard, replay the patient's recorded motion and let the controller personalize to that. It's a smaller, quieter shift than "AI walks for you" — and a more honest one.