Under Review

Safer Personalization of Prosthesis Controllers Through Replay-Constrained Simulation

A replay-constrained simulation framework for personalizing powered knee–ankle prosthesis controllers before testing them on real users. Instead of blindly tuning controller parameters on hardware, this workflow replays a user's recorded hip motion and ground-reaction interaction in MuJoCo, then searches for impedance strategies that are likely to transfer safely to the physical prosthesis.

Replay-Constrained Simulation MuJoCo Impedance Personalization Sim-to-Real Transfer Reinforcement Learning Powered Prosthesis

Replay-constrained sim-to-real impedance personalization concept diagram — Replay-constrained simulation ranks personalized impedance controllers before they reach the physical prosthesis.

Problem & Contribution

A prosthesis never walks alone — it is coupled to a human whose neuromuscular response is hard to model. Rather than simulate the person, this work replays recorded user motion while simulating the device, so risky high-dimensional impedance strategies can be screened and ranked before any real-user trial.

My Role

I contributed to a replay-constrained simulation framework for powered knee–ankle prosthesis controller personalization, connecting user-recorded gait data, MuJoCo prosthesis dynamics, and hardware validation.

Challenge

Full-dimensional impedance personalization involves many coupled parameters across the knee and ankle joints, making direct hardware exploration slow, risky, and difficult to scale.

Contribution

A replay-constrained MuJoCo framework that ranks personalized knee–ankle impedance controllers in simulation before real-user testing.

Ph.D. research conducted in the LocoLab at the University of Michigan, advised by Prof. Robert D. Gregg.

How It Works

1 Recorded User Data Hip motion and ground-reaction interaction captured while the user walks.
2 Replay-Constrained Simulation The recording drives a MuJoCo prosthesis model, so only the device is simulated.
3 Controller Optimization Phase-dependent knee–ankle impedance is searched for biomimetic behavior.
4 Hardware Validation Top-ranked candidates are tested on the physical prosthesis.
5 Personalized Controller A per-user controller that transfers safely from simulation to hardware.

Evidence

Simulation Ranking Transferred

Sim-to-Real

42–59% Reward Gain (Preliminary)

vs. Baseline

Full Knee–Ankle Personalization

Both Joints

Technical Stack

Replay-Constrained Simulation MuJoCo Impedance Personalization Sim-to-Real Transfer Reinforcement Learning Human-in-the-Loop Control

More Related Work

View all projects →

Control Continuous Impedance Control for Wearable Locomotion Smooth swing motion isn't enough — during stance the robot must also provide the right support, so this extends continuous transition control into impedance behavior during weight-bearing. View Project →

Personalization Low-Data Personalization for Wearable Robot Intent Prediction Personalization is a major barrier for intent recognition; transfer learning cuts the new-user data burden while preserving intent-prediction performance. View Project →

← Back to All Projects Featured Projects