Exo-Plore: Exploring Exoskeleton Control Space through Human-aligned Simulation

Geonho Leem, Jaedong Lee, Jehee Lee, Seungmoon Song, Jungdam Won

January 2026

Simulated normal, crouch and equinus gaits walking with the hip exoskeleton.

TL;DR Exo-plore tunes a hip exoskeleton's gain and delay entirely in a muscle-driven simulation of a walker trained with deep RL, with no human walking trials. Its assisted gaits follow human trends (0.96 correlation on RMS assistive moment), and the optimal gain varies linearly with severity in four of five simulated gait pathologies.

What they do

Tuning an exoskeleton controller with human-in-the-loop optimization takes hours of walking, which people with mobility impairments often cannot do. Exo-plore moves the search into simulation: a musculoskeletal walker (23 bones, 164 Hill-type muscles) wearing a simulated Samsung GEMS hip exoskeleton is trained with deep RL to react to assistance as people do, and the controller’s two parameters, gain and delay, are then optimized to minimize metabolic cost of transport (CoT).

How it works

Gait data generator. A pose network trained with RL outputs PD targets; the exoskeleton torque is subtracted from the PD torque and a muscle coordination network turns the remainder into muscle activations. A regularizer keeps the line segments of one anatomical muscle activating together.
Sim-to-real matching. The exponents of the metabolic energy model are chosen so the simulated preferred walking speed matches the human 1.25 m/s. A human-exoskeleton interaction (HEI) reward penalizes moments when the exoskeleton resists the wearer.
Exoskeleton optimizer. An MLP surrogate is fit to simulated CoT from Latin hypercube samples with a Huber loss and gradient penalty, giving a smooth landscape for gradient-based search.
Pathologies. Equinus, waddling, crouch, calcaneus and foot drop gaits are produced by weakening or shortening specific muscles, with one model per condition.

Results

Unassisted knee and hip flexion match human data with correlation 0.97; hip internal rotation does not (-0.44).
With the HEI reward, RMS assistive moment correlates 0.964 with human experiments, against 0.745 for assistance maximization and 0.726 with no HEI reward. At 4 km/h, raising the delay from 0.05 s to 0.25 s increases assistive power 1.73-fold in simulation and 1.88-fold in humans.
For able-bodied gait the optimal delay decreases monotonically as walking speed increases.
Optimal gain is linear in severity for four pathologies: it rises for crouch and calcaneus and falls for equinus and waddling. Foot drop shows no clear trend.

Limitations

Nothing is validated on human subjects or patients; the optimized controllers are simulation predictions only. The reward models are simplified and not personalized, and the rigid-tendon Hill-type muscles miss history-dependent effects. The simulated rectus femoris is overactivated, and the rigid box-shaped foot distorts short steps and vertical ground reaction forces.

Videos

The five simulated pathological gaits: equinus, waddling, crouch, calcaneus and foot drop.

Close-up of foot contact in normal, equinus, calcaneus and foot drop gaits.

Figures

From Figure 1: The Exo-plore framework. The generator produces gait trajectories under hip exoskeleton assistance usi... — From Figure 1: The Exo-plore framework. The generator produces gait trajectories under hip exoskeleton assistance using a musculoskeletal character with 164 muscles. The generated data are then used to train a surrogate network, which enables efficient and customizable optimization of exoskeleton control parameters.

From Figure 5: Comparison of gait kinematics and dynamics under exoskeleton assistance between human experimental res... — From Figure 5: Comparison of gait kinematics and dynamics under exoskeleton assistance between human experimental results (top) and simulated results (bottom).

From Figure 7: Comparison of maximum metabolic reduction rate for the HEI reward, assistance maximization and no HEI ... — From Figure 7: Comparison of maximum metabolic reduction rate for the HEI reward, assistance maximization and no HEI reward, against the human experiment.

From Figure 9: Optimal exoskeleton gain over five pathological gait types, with values reported for different levels ... — From Figure 9: Optimal exoskeleton gain over five pathological gait types, with values reported for different levels of pathology severity in each gait type. Each color corresponds to a different random seed.

Abstract

Exoskeletons show great promise for enhancing mobility, but providing appropriate assistance remains challenging due to the complexity of human adaptation to external forces. Current state-of-the-art approaches for optimizing exoskeleton controllers require extensive human experiments in which participants must walk for hours, creating a paradox: those who could benefit most from exoskeleton assistance, such as individuals with mobility impairments, are rarely able to participate in such demanding procedures. We present Exo-plore, a simulation framework that combines neuromechanical simulation with deep reinforcement learning to optimize hip exoskeleton assistance without requiring real human experiments. Exo-plore can (1) generate realistic gait data that captures human adaptation to assistive forces, (2) produce reliable optimization results despite the stochastic nature of human gait, and (3) generalize to pathological gaits, showing strong linear relationships between pathology severity and optimal assistance.

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