The strongest proof: a simulation-trained controller cut metabolic cost by 19.7% in real humans
The most direct answer to your question comes from a 2026 study that took a controller trained entirely in simulation and tested it on eight healthy adults wearing a hip exoskeleton. The active assistance lowered net metabolic rate by 19.7% compared to the passive device (p < 0.001) — that's a statistically significant, substantial reduction in energy cost, the gold-standard measure for assistive walking [1]. This is the cleanest single experiment because it used whole-body metabolic measurement, not just biomechanical proxies, and the controller was deployed without any subject-specific retraining, meaning it worked 'out of the box' on real people.
The same study also reported that the controller produced predominantly positive hip mechanical power (positive-power ratio 0.98) — meaning the device was consistently helping, not hindering, the wearer's movement. It also generalized across walking speeds and terrains, which is a key sign of robustness beyond the exact conditions it was trained on [1]. This is the kind of evidence that gives medical robotics teams confidence: a simulation-trained policy can deliver a real, measurable benefit on real users.
Supporting evidence: controllers can handle varied users and adapt to them, but reliability has limits
A 2023 study from the same research group reinforced the idea that simulation-trained controllers can be robust across different user conditions. They trained a deep reinforcement learning controller for a lower-limb rehabilitation exoskeleton using a decoupled offline simulation with a musculoskeletal model, and virtually tested it on simulated patients with passive muscles (quadriplegic), muscle weakness, and hemiplegic conditions — all without any control parameter tuning [2]. The controller maintained stable walking (measured by joint tracking error, center-of-pressure stability, and gait symmetry), showing that a single policy can handle a range of disabilities, not just one specific case.
However, a 2024 study on ankle exoskeletons highlights a critical caveat: even in controlled lab settings, exoskeletons can experience errors in torque assistance due to misalignments between measured and actual states. When errors occur repeatedly, users may anticipate them and resist the device, which undermines the collaboration [4]. That study developed a co-adaptive controller that adjusts torque based on muscle activity and joint kinematics, successfully modulating assistance in response to both fluent and non-fluent behaviors [4]. This suggests that while simulation-trained controllers are reliable for many tasks, achieving true 'dependability' in unpredictable real-world environments may require additional co-adaptive mechanisms to maintain human-robot fluency.
So, when can you depend on it?
Across the evidence, the answer is nuanced. For well-defined tasks like walking at various speeds and terrains, a simulation-trained controller can be dependable enough to deliver significant metabolic benefits, as shown by the 19.7% reduction in the 2026 study [1] and the robust performance across simulated disabilities in the 2023 study [2]. These two studies converge on the same conclusion from different angles — one with real humans, one with simulated patients — which strengthens the case that simulation-trained controllers are not just theoretical.
But the 2024 study [4] reminds us that reliability isn't just about average performance; it's about how the system handles errors and maintains user trust. If a controller mis-times assistance, users may adapt in ways that reduce the benefit or even cause resistance. So, for medical robotics teams, the practical answer is: you can depend on a simulation-trained controller for tasks where the dynamics are well-understood and the training includes realistic variability (like muscle strength randomization [2]), but you should plan for co-adaptive mechanisms to handle the inevitable errors in real-world settings. The field is close, but not yet at the point where you can 'set and forget' a controller in all conditions.
About These Sources
This answer is built on 4 studies (2 peer-reviewed, 2 preprints) — published from 2023 to 2026, 3 from 2024 or later, 2 in Q1 journals, collectively cited 74 times — selected as the most relevant from 4 studies that passed quality screening, drawn from 47 papers retrieved from a database of over 500 million.
Sources used in this answer
Staged Multi-Agent Training (SMAT) for Hip Exoskeletons: Metabolic and Biomechanical Validation of a Simulation-Trained Co-Adaptive Controller
In a study with eight healthy adults, a simulation-trained hip exoskeleton controller (SMAT) reduced net metabolic rate by 19.7% compared to passive mode (p < 0.001), with a positive-power ratio of 0.98, and generalized across speeds and terrains without subject-specific retraining.
Robust walking control of a lower limb rehabilitation exoskeleton coupled with a musculoskeletal model via deep reinforcement learning
A deep reinforcement learning controller for a lower-limb rehabilitation exoskeleton, trained in simulation with domain randomization of muscle strength, was virtually tested on simulated patients with quadriplegic, weak, and hemiplegic conditions, maintaining stable walking without parameter tuning.
Robotic exoskeleton adapts to its wearer through simulated training
A 2024 Nature commentary highlights that simulation-based training can lead to versatile and adaptable exoskeleton controllers, reducing the need for extensive human data, and suggests this approach paves the way for everyday integration.
Characterizing Human-Exoskeleton Fluency for Co-Adaptive Control of Ankle Exoskeletons
A 2024 thesis on ankle exoskeletons found that torque errors can cause users to resist the device, and developed a co-adaptive controller that adjusts assistance based on muscle activity and joint kinematics to improve human-exoskeleton fluency.