Musculoskeletal Motion Imitation for Learning Personalized Exoskeleton Control Policy in Impaired Gait

Itak Choi, Ilseung Park, Eni Halilaj, Inseung Kang

April 2026

TL;DR An RL policy drives a 90-muscle musculoskeletal model to imitate human walking and running, then is fine-tuned to also command hip or ankle exoskeleton torques, including for simulated muscle weakness. In simulation, ankle assistance cuts metabolic cost by 13.8% in running, and deficit-specific assistance cuts it by 8.1-13.1% in impaired gait.

What they do

Tuning exoskeleton assistance usually needs human-in-the-loop optimization or large biomechanics datasets, both hard to do with patients. This paper trains a muscle-driven model with RL to imitate recorded gait, then fine-tunes it to command hip or ankle exoskeleton torques, for a healthy model and for models with one weakened muscle group. Everything is in simulation.

How it works

Model and data. A 21-DoF, 90-muscle Hyfydy model with sagittal torque motors at hip, knee and ankle. It imitates a two-minute treadmill clip of one subject walking and running at 0.6-2.2 m/s, mirrored for symmetry.
Base policy. SAC outputs muscle excitations at 25 Hz, rewarded for tracking joint angles, velocities, body positions and inverse-dynamics torques, and penalized for metabolic cost. 600M steps, about 48 h on one RTX 5090.
Fine-tuning. 150M more steps per device with exoskeleton outputs enabled, device mass added, looser tracking and larger energy penalties. For impaired gait, one side’s plantarflexors or hip flexors are capped at 5% excitation and the model still imitates healthy gait.

Results

Joint-angle RMSE is 2.85-4.03° and joint-moment RMSE 0.10-0.11 Nm/kg; ground reaction forces match with R² 0.95 although no reference forces were used in training.
Learned torque profiles resemble experimentally optimized ones. Hip assistance lowers simulated metabolic cost by 4.57% (walking) and 6.95% (running), ankle assistance by 6.15% and 13.81%.
Weakness alone yields compensations such as hip hiking. With assistance at 1.2 m/s, torque is 59% (ankle) and 49% (hip) higher on the affected side, metabolic cost drops 8.1% and 13.1%, and left-right joint-angle RMSE falls from 6.31° to 3.56° and from 6.15° to 2.59°.

Limitations

Nothing has run on hardware, and simulated weakness may not capture real pathological gait. Peak torques for the impaired models may exceed what lightweight portable devices deliver. Only level-ground gait from one subject is covered.

Figures

From Figure 1: Framework overview. (a) A muscle-driven policy imitates reference kinematics and joint moments extract... — From Figure 1: Framework overview. (a) A muscle-driven policy imitates reference kinematics and joint moments extracted from motion capture and force plates, with a metabolic cost probe in the reward. (b) Fine-tuning adds exoskeleton torques to the action space. (c) An activation mask weakens targeted muscles to produce impaired gait.

From Figure 3: Hip and ankle assistive torque profiles (Nm/kg) learned for walking and running at 0.6 to 2.2 m/s (blu... — From Figure 3: Hip and ankle assistive torque profiles (Nm/kg) learned for walking and running at 0.6 to 2.2 m/s (blue), compared with experimentally optimized profiles from prior work (red).

From Figure 4: Simulated gross metabolic cost (W/kg) of the able-bodied model with no exoskeleton (gray), hip assista... — From Figure 4: Simulated gross metabolic cost (W/kg) of the able-bodied model with no exoskeleton (gray), hip assistance (blue) and ankle assistance (green), (a) across speeds and (b) averaged over walking and running.

From Figure 7: Impaired gait at 1.2 m/s. (a) Joint angles of the affected (orange) and non-affected (blue) sides with... — From Figure 7: Impaired gait at 1.2 m/s. (a) Joint angles of the affected (orange) and non-affected (blue) sides without and with the exoskeleton, for ankle plantarflexor weakness (top) and hip flexor weakness (bottom). (b) Gross metabolic cost without (gray) and with (green) assistance.

Abstract

Designing generalizable control policies for lower-limb exoskeletons remains fundamentally constrained by exhaustive data collection or iterative optimization procedures, which limit accessibility to clinical populations. To address this challenge, we introduce a device-agnostic framework that combines physiologically plausible musculoskeletal simulation with reinforcement learning to enable scalable personalized exoskeleton assistance for both able-bodied and clinical populations. Our control policies not only generate physiologically plausible locomotion dynamics but also capture clinically observed compensatory strategies under targeted muscular deficits, providing a unified computational model of both healthy and pathological gait. Without task-specific tuning, the resulting exoskeleton control policies produce assistive torque profiles at the hip and ankle that align with state-of-the-art profiles validated in human experiments, while consistently reducing metabolic cost across walking speeds. For simulated impaired-gait models, the learned control policies yield asymmetric, deficit-specific exoskeleton assistance that improves both energetic efficiency and bilateral kinematic symmetry without explicit prescription of the target gait pattern. These results demonstrate that physiologically plausible musculoskeletal simulation via reinforcement learning can serve as a scalable foundation for personalized exoskeleton control across both able-bodied and clinical populations, eliminating the need for extensive physical trials.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.

Matt Strong