Powered wearable devices, such as exoskeletons and soft exosuits, push on the body at specific moments in each stride. That push always arrives late. Compressed air takes about 0.2 s to fill a bladder. Signal processing in muscle-driven devices can add close to half a second. Wearers notice timing errors as small as 2.8% of a stride. Designers increasingly tune assistance timing by asking the wearer which of two settings felt better, but what wearers say about timing does not match what their muscles do. People report feeling relaxed while their leg muscles brace against the device.
This note proposes that the mismatch has a predictable structure, and that the structure comes from the latency itself. We split delivered timing error into two parts. The first is a steady offset: assistance that is always late or always early by the same amount. The second is jitter: stride-to-stride variability in when the push arrives. Our hypothesis is that wearers' self-reports mostly track jitter, because unpredictability is what they consciously notice. Their muscle co-contraction, meaning opposing muscles tightening together to stiffen the joint, mostly tracks the steady offset, because a consistently mistimed push must be physically resisted even after the wearer has stopped noticing it. If this holds, then an optimizer driven only by preferences will leave steady lag uncorrected, and an optimizer driven only by muscle signals will ignore the jitter that erodes trust. We propose to test the hypothesis directly. We then build a timing optimizer that uses each signal only for the part of the error it actually reports on.
Core Claim
Let $t_c$ be the commanded onset of assistance within a stride, and let $t_d = t_c + \delta(s) + \epsilon$ be the onset actually delivered to the body. Here $\delta(s)$ is an actuator delay that depends on the device state $s$ (supply pressure, cadence, load), and $\epsilon$ is stride-to-stride noise. The delivered timing error relative to a person's best onset $t^*$ then has a mean offset $m = \mathbb{E}[t_d] - t^*$ and a jitter $\sigma = \mathrm{sd}(t_d)$. We model pairwise reports and muscle bracing as two different readouts of the same pair $(m, \sigma)$:
$$P(A \succ B) = \Phi\big(-\alpha_m(|m_A| - |m_B|) - \alpha_\sigma(\sigma_A - \sigma_B)\big), \qquad C = \gamma_m |m| + \gamma_\sigma \sigma + \eta .$$
In these equations, $\Phi$ is the standard normal cumulative distribution and $C$ is a co-contraction index computed from surface electromyography (EMG), which is the electrical activity of muscles recorded through skin electrodes. The coefficients $\alpha$ and $\gamma$ are per-person sensitivities, and $\eta$ is measurement noise.
The claim is that $\alpha_\sigma/\alpha_m \gg \gamma_\sigma/\gamma_m$. In words, reports weight jitter far more heavily than muscles do, relative to how each weights steady offset. The prediction rests on a familiar intuition about motor adaptation: people adjust to a consistent delay and stop perceiving it, but they cannot adjust to an unpredictable one. That intuition comes from outside the file and is exactly what the experiment tests. The claim also explains the "illusion of relaxation." Wearers of a device with low jitter and a steady lag should rate it as comfortable while bracing against it.
How This Sits Against Prior Work
The motivating evidence comes from Measuring Trust for Exoskeleton Systems. That work compares trust surveys against EMG co-contraction, joint kinematics and gaze. It shows that wearers feel timing errors of 2.8% of a stride, that ratings of predictability fall as the controller's error rate rises, and that wearers report relaxation while their muscles resist. It treats timing error as a single quantity and proposes no controller. We split timing error into offset and jitter and predict that the two measurement channels diverge along that split.
Two adjacent findings point the same way without examining timing. Hybrid Robotic and Electrical Stimulation Assistance Can Enhance Performance and Reduce Mental Demand combines a wrist robot with electrical muscle stimulation. It finds that the delay between a stimulation pulse and the resulting muscle contraction ruins tracking when stimulation is used alone. It also finds that participants report lower mental demand in the hybrid condition despite higher muscle activation. Our reading of that result is that reports followed smoothness rather than effort, which is what our model predicts if smoothness is a stand-in for low jitter. Evaluation of a machine-learning-driven active–passive upper-limb exoskeleton robot documents about 465 ms of end-to-end latency and a subjective "table effect," but only in a static holding task where timing hardly matters.
The method side draws on two lines of work. Preference-based personalization supplies the query machinery. Preference-Based Learning for Exoskeleton Gait Optimization (COSPAR, which combines pairwise comparisons with user-suggested parameter changes) and Preferential Multi-Attribute Bayesian Optimization with Application to Exoskeleton Personalization (SDTS, which asks attribute-by-attribute comparisons instead of one overall judgment) both assume that the commanded parameters are what the body experiences. We drop that assumption. Latency compensation by learning supplies the controller. Reinforcement Learning Control of a Physical Robot Device for Assisted Human Walking without a Simulator learns to fire a pneumatic exosuit about 0.2 s early and cuts muscle effort by 5% to 48.2%. It never asks whether wearers perceive the early firing correctly. Look-ahead inverse models such as Deep Neural Networks for Improved, Impromptu Trajectory Tracking of Quadrotors and A Neural Network-Based Approach to Robot Motion Control cut tracking error by 43% and more than 50%. Both were developed without a human attached, and the second was tested with the robot's limbs in the air.
Two papers constrain the design. Electromyography (EMG)-Based Feature Selection for Detecting Movement Effort in Human-in-the-Loop Optimization of Lower Limb Exoskeletons shows that even the best EMG features detect changes in assistance timing only about 71% of the time, so EMG cannot serve as perfect ground truth. Continual Human-in-the-Loop Optimization shows that people tune their own settings worse than an optimizer does, which warns against trusting self-adjustment.
Research Program
The work has three parts.
The first part is an instrumented testbed that measures delivered timing rather than assuming it. We will use a soft pneumatic hip exosuit of the kind described by Zhong et al. A load cell and a bladder pressure sensor timestamp the actual force onset $t_d$ on every stride. This lets us estimate $\delta(s)$ as a function of supply pressure and cadence. Software can then inject controlled steady offsets and controlled jitter on top of the device's native delay. EMG is recorded from antagonist pairs at the hip and ankle. Calf channels are chosen because the EMG feature benchmark found gastrocnemius and soleus to be the most responsive muscles. Co-contraction is computed from the features that benchmark found consistent across tasks, specifically the log difference absolute mean value and the root mean square.
The second part is a psychophysics study that estimates $\alpha$ and $\gamma$ for each participant. The third part is a dual-channel timing optimizer. It places a Gaussian-process utility over delivered timing $(m, \sigma)$ rather than commanded timing. It assigns the steady-offset dimension mainly to the EMG channel and the jitter dimension mainly to pairwise reports, with weights set from each person's fitted sensitivities. It chooses the commanded onset and a smoothing gain, the gain being the device's lever on jitter, using a state-dependent delay model learned online. Before running people, we will reproduce the failure mode in simulation. We extend the Atalante simulation benchmark used by SDTS with injected actuator delay and with simulated reporters whose $\alpha_\sigma/\alpha_m$ is varied. This checks whether a preference-only posterior converges to a lagged optimum, and by how much, as the reporter's bias grows.
Baselines
The naive reference point is the device's default timing with no personalization, compared against walking without the suit and with the suit powered off. Every arm reports results against these.
The strongest preference-based competitor is COSPAR from Preference-Based Learning for Exoskeleton Gait Optimization, run over commanded onset. Its paper reports that subjects' blind rankings agreed with its posterior, but no reported result was found on Lacuna for a quantitative utility or timing accuracy on physical hardware. We will also run SDTS from Preferential Multi-Attribute Bayesian Optimization with separate queries for "felt on time" and "felt predictable." This tests whether simply asking about the two dimensions separately already fixes the bias. That paper reports larger Pareto hypervolume than overall-preference querying, with no numbers given on Lacuna.
The strongest objective competitor is an EMG-reward optimizer modeled on Reinforcement Learning Control of a Physical Robot Device for Assisted Human Walking without a Simulator. It uses that paper's muscle-effort reward with a kinematic-normalcy penalty. The paper reports muscle effort reductions of 5% to 48.2% relative to walking without the suit, from 5 participants. We will reimplement it on our device rather than compare against those numbers across devices.
The key ablation is a calibrated-preference baseline. It is COSPAR over delivered timing, where the mean delay $\mathbb{E}[\delta]$ is measured once at the start of the session and subtracted from the commanded onset. This isolates whether the gain comes from the dual-channel split or merely from knowing the latency.
All arms receive the same participants in counterbalanced order, the same 150-stride walking blocks and the same budget of 25 queries or evaluations per session. Every arm is followed by the same held-out evaluation block. We will count the proposal as beating a baseline if it achieves lower held-out co-contraction and higher blind-ranked preference at the same time. Beating a baseline on one measure while losing on the other does not count. The comparison uses within-subject paired tests and a minimum effect set in advance from pilot variance.
Why not just…?
Why not just optimize on EMG and ignore reports? EMG detects changes in assistance timing only about 71% of the time, according to the EMG feature benchmark. Zhong et al. found that rewarding effort reduction alone distorted gait. Tucker et al. found that energetically optimal gaits could be intolerable to wearers. EMG alone also has no channel for the trust loss caused by jitter. The EMG-only arm will show whether that loss matters: if its held-out preference rankings and trust survey scores are no worse than the dual-channel arm's, reports add nothing and we will say so.
Why not just measure the mean delay and subtract it? This is the most serious threat. If the delay is nearly constant, and if reports and muscles respond to timing in the same proportions, then calibrated COSPAR captures essentially all of the benefit. Our proposal only earns its complexity if the psychophysics study finds $\alpha_\sigma/\alpha_m$ substantially larger than $\gamma_\sigma/\gamma_m$, and if pneumatic delay varies with pressure and cadence. The calibrated-preference ablation exists to settle this. If it matches the dual-channel arm, the conclusion is "measure your latency," which is still a useful and publishable result.
Why not just ask attribute-wise questions with SDTS? Asking separately about "on time" and "predictable" might remove the confusion without any physiology. However, if wearers truly adapt to a steady lag, they will answer "on time" even when it is not, so attribute-wise reports would inherit the same blind spot. The SDTS arm tests this directly.
Why not just remove jitter in hardware? Faster valves help, but the biological and processing delays documented by Cazenave et al. and Nasr et al. do not disappear. The pressure-dependent part of pneumatic delay also remains. We treat jitter as an experimental factor precisely because it cannot be engineered away in many devices.
Proposed Experiments
Study 1 tests the core claim with about 16 able-bodied participants walking on a treadmill at level grade and at a 7-degree incline. In each block, delivered steady offset relative to the person's EMG-estimated best onset is set at one of five levels from −150 to +150 ms. Jitter is set at a standard deviation of 0, 30 or 60 ms. Each block yields pairwise comparisons against a reference block, a Stirling-style trust survey and a co-contraction index. The decisive check is an interaction. It would show up as a steady-offset condition with low jitter that is preferred over a centered condition with high jitter, while producing more co-contraction. We will estimate the ratio $(\alpha_\sigma/\alpha_m)/(\gamma_\sigma/\gamma_m)$ with a hierarchical model. A credible interval excluding 1 supports the claim. A ratio near 1 falsifies it.
Study 2 compares the dual-channel optimizer against the baselines on the same participants, with delivered timing measured on every stride. Native delay varies through two supply pressures, and native jitter is reduced through the smoothing gain. The decisive checks are three. The preference-only arms should converge to a measurably lagged delivered onset. The EMG-only arm should converge to an onset whose jitter is higher than the dual-channel arm's. The dual-channel arm should beat calibrated COSPAR on the joint criterion defined in the Baselines section. A shorter out-of-lab validation will reuse WearGait-PD insole recordings, whose Android timestamps show non-deterministic latency, to check whether realistic sensor jitter falls in the range where Study 1 found reports to be sensitive.
References
- Measuring Trust for Exoskeleton Systems
- Hybrid Robotic and Electrical Stimulation Assistance Can Enhance Performance and Reduce Mental Demand
- Evaluation of a machine-learning-driven active–passive upper-limb exoskeleton robot: Experimental human-in-the-loop study
- Preference-Based Learning for Exoskeleton Gait Optimization
- Preferential Multi-Attribute Bayesian Optimization with Application to Exoskeleton Personalization
- Reinforcement Learning Control of a Physical Robot Device for Assisted Human Walking without a Simulator
- Deep neural networks for improved, impromptu trajectory tracking of quadrotors
- A Neural Network-Based Approach to Robot Motion Control
- Electromyography (EMG)-Based Feature Selection for Detecting Movement Effort in Human-in-the-Loop Optimization of Lower Limb Exoskeletons
- Continual Human-in-the-Loop Optimization