When do simulation-trained controllers beat traditional methods?
Simulation-trained controllers can match or exceed the performance of controllers tuned through lengthy human experiments, but only when the simulation is rich enough to model how a person adapts to the device. The most dramatic example is a 2024 Nature study that trained a hip exoskeleton controller entirely in simulation, with no human tests, and then deployed it on hardware—it reduced metabolic energy use by 24.3% during walking, 13.1% during running, and 15.4% during stair climbing [1]. That means a person wearing the exoskeleton burned roughly a quarter less energy walking than without it, a benefit comparable to what traditional human-in-the-loop tuning achieves, but without the hours of lab time.
Another study used a staged training approach that mimics how people naturally adapt to wearing a device: first the human learns to walk unassisted, then adapts to the device's mass, then the exoskeleton learns to assist, and finally both co-adapt. This produced a 10.1% reduction in hip muscle activation in simulation, and when deployed on a physical exoskeleton with five subjects, it delivered consistent assistance without any subject-specific retraining [3]. The key takeaway: if the simulation includes a musculoskeletal model and a curriculum that mirrors human adaptation, the learned controller can be as good as—or better than—one tuned by hand, and it generalizes across users without extra tuning.
Where does sim-to-real transfer still fall short?
Simulation-trained controllers are not perfect—they can struggle at the extremes of human movement, and the gap between simulation and reality grows as tasks become more demanding. One study that trained controllers to reduce biological joint moments found strong agreement at the hip (correlation 0.94–0.98) but discrepancies increased at higher speeds and steeper inclines, especially at the knee [4]. In plain terms, the controller predicted assistance well for normal walking, but its accuracy dropped when people walked faster or on steeper slopes, meaning the assistance might be mistimed or mis-scaled in those situations.
Another study that trained a controller in simulation and deployed it on hardware found that the assistance profiles transferred well (correlation 0.82, with a small error of 0.03 Nm/kg), but the benefits were modest—only a 3.4% reduction in muscle activation and 7.0% reduction in joint power [2]. That is a much smaller effect than the 24% metabolic reduction seen in the Nature study, likely because the simulation was less comprehensive or the task was different. The lesson: simulation-trained controllers are reliable for common activities like level walking, but their performance degrades at the edges of the movement envelope, and the magnitude of benefit can vary widely depending on the model's fidelity.
What about people with mobility impairments?
The biggest unanswered question is whether simulation-trained controllers work for people who actually need exoskeletons—those with mobility impairments. The studies here focus on able-bodied users, and only one explicitly addresses pathological gaits. That study, called Exo-plore, used a neuromechanical simulation to optimize hip exoskeleton assistance and found it could generalize to pathological gaits, showing a strong linear relationship between the severity of the impairment and the optimal assistance [5]. That means the simulation could predict that someone with a more severe gait impairment would need more assistance, which is promising.
But none of these studies tested the controller on actual patients. The Exo-plore authors note that current state-of-the-art methods require hours of walking in human experiments, which is impractical for people with mobility impairments [5]. Simulation-trained controllers could solve that problem, but until they are validated in clinical populations, we cannot be sure they will work as well as they do for able-bodied users. So, for now, the answer is: simulation-trained controllers are a powerful tool for able-bodied users and common activities, but for clinical populations, they remain an unproven promise.
About These Sources
This answer is built on 5 studies (1 peer-reviewed, 4 preprints) — published from 2024 to 2026, 5 from 2024 or later, 1 in Q1 journals, collectively cited 105 times — selected as the most relevant from 5 studies that passed quality screening, drawn from 74 papers retrieved from a database of over 500 million.
Sources used in this answer
Experiment-free exoskeleton assistance via learning in simulation
In a 2024 Nature study, a hip exoskeleton controller trained entirely in simulation, without human experiments, reduced metabolic energy use by 24.3% (walking), 13.1% (running), and 15.4% (stair climbing) when deployed on hardware.
Learning Hip Exoskeleton Control Policy via Predictive Neuromusculoskeletal Simulation
A physics-based neuromusculoskeletal learning framework trained a hip exoskeleton policy in simulation and deployed it on hardware, achieving a 0.82 correlation between simulated and real assistance profiles, with a small error of 0.03 Nm/kg, and modest reductions in muscle activation (3.4%) and joint power (7.0%).
SMAT: Staged Multi-Agent Training for Co-Adaptive Exoskeleton Control
The SMAT (Staged Multi-Agent Training) approach used a four-stage curriculum to co-train a human and exoskeleton in simulation, resulting in a 10.1% reduction in hip muscle activation in simulation and consistent assistance on a physical exoskeleton with five subjects, without subject-specific retraining.
Exoskeleton Control through Learning to Reduce Biological Joint Moments in Simulations
Simulation-trained MLP controllers for reducing biological joint moments showed strong agreement at the hip (correlation 0.94–0.98) but discrepancies increased at higher speeds and steeper inclines, especially at the knee, and delay tuning could bias assistance toward greater positive power injection.
Exo-Plore: Exploring Exoskeleton Control Space through Human-aligned Simulation
The Exo-plore framework combined neuromechanical simulation with deep reinforcement learning to optimize hip exoskeleton assistance without human experiments, and showed it could generalize to pathological gaits with a strong linear relationship between pathology severity and optimal assistance.