Subject-Specific Predictive Musculoskeletal Simulations of Lower-Limb Exoskeleton Assistance: Metabolic and Biomechanical Effects of Joint Assistance Strategies

What are the energy-optimal single- and multi-joint hip, knee, and ankle assistance strategies across subject-specific musculoskeletal models, and how do individual anthropometric differences influence predicted metabolic and biomechanical outcomes?

Designing individualized lower-limb exoskeleton assistance is difficult because users vary in anthropometry, strength, and biomechanics, and experimental tuning is expensive. Generic predictive models have also shown poor agreement with individual-level measured metabolic savings, so makers lack reliable pre-hardware guidance on which joints to assist.

Earlier work used expert-guided or heuristic torque tuning, human-in-the-loop optimization, computed muscle control with prescribed kinematics, and single-shooting predictive simulations in SCONE. Fixed-kinematics methods cannot predict how gait changes in response to assistance, shooting methods are computationally expensive, and generic models have shown weak subject-level prediction of experimental metabolic savings.

The authors built six subject-specific OpenSim Moco direct-collocation models scaled from an open gait dataset, with BMI-based muscle strength scaling. They added idealized bilateral sagittal hip, knee, and ankle torque actuators, simulated unassisted walking and all seven joint combinations at 25 and 50 Nm, and optimized motion with a cost function combining metabolic rate, muscle effort, joint and center-of-mass acceleration, and tracking of reference gait.

At 50 Nm, H+K+A reduced predicted cost of transport by 48.50 ± 4.55% relative to an unassisted baseline of 3.07 W/kg, with individual reductions from 42.22% to 54.65%; at 25 Nm the mean reduction was 41.20%. Hip alone was the best single joint at 50 Nm with 31.77 ± 5.21%, followed by knee at 19.17% and ankle at 17.02%. H+A produced 44.60 ± 7.08%, nearly matching H+K+A. Mean positive power at 50 Nm for single-joint cases was 0.375 W/kg at the hip, 0.243 at the knee, and 0.220 at the ankle; positive power ratio was 0.861 at the ankle, 0.820 at the hip, and 0.672 at the knee. Raising torque from 25 to 50 Nm increased hip mean positive power from 0.267 to 0.359 W/kg and ankle from 0.124 to 0.205 W/kg, but knee only from 0.163 to 0.168 W/kg. Peak assistive torque increased from 0.372 to 0.729 Nm/kg at the hip and 0.372 to 0.745 at the ankle, versus 0.310 to 0.339 at the knee. In the H+K condition at 50 Nm, knee positive power ratio fell to 0.442 with mean positive power of only 0.073 W/kg. One subject, AB14, reached 0.708 W/kg hip mean positive power in H+K, roughly double the six-subject mean of 0.388 W/kg.

The authors state that the model is simplified: only nine muscles per limb, rigid tendons, no co-contraction beyond the effort term, simplified foot-ground contact, and a tracking term that constrains kinematic adaptation, so absolute metabolic savings are likely inflated and require experimental validation. This is a preprint with no hardware trials. The subject sample is small, consisting of six young able-bodied adults; the model is 2D and bilaterally symmetric; actuators are ideal with no device mass or dynamics; and deep muscle activations such as BFSH could not be validated.