Wearable robots such as hip exoskeletons need to know where the leg will be a fraction of a second from now, not where it was. Sensing, radio links, computation and motors all add delay, so a command computed for the current pose arrives late. Sparse pressure sensors are a promising source of that look-ahead. Examples are yarn sensors woven into a knee sleeve that feel the muscle bulge under the fabric, or shoe insoles that feel foot loading. Muscle shape and foot loading change as movement is being prepared, which is before the joint visibly rotates. The catch is that the same muscle contraction produces a different pressure pattern on a thick thigh than on a thin one, and the same gait produces a different insole signal at a different body mass. A model trained on one narrow body type may therefore fail on others. Biomechanics datasets come from small, homogeneous convenience cohorts, so this failure has never been measured.
We propose to build and test a pressure-driven forecaster of future lower-limb joint rotations, along with the evaluation protocol that tells whether it transfers across human body morphologies. The two-sentence version is this. Sparse pressure signals carry information about motion about to happen that inertial sensors alone do not, but the mapping from pressure to motion depends on soft-tissue and segment geometry. A forecaster is only useful in the field if that dependence is measured and modelled with continuous body measurements, rather than ignored or replaced by coarse ancestry labels.
Core Claim
We make three linked, falsifiable claims. First, adding knee-sleeve pressure and insole channels to inertial measurement units (IMUs, small accelerometer–gyroscope packages) reduces the error of forecasting joint rotations 50–200 ms ahead. The reduction should be larger than the reduction for estimating the current pose, because pressure leads kinematics. Second, forecasters that use pressure lose more accuracy than IMU-only forecasters when tested on body types far from the training cohort, because pressure transduction depends on limb circumference, adiposity and mass. Third, conditioning the forecaster on a handful of tape-measure body measurements closes most of that gap, and self-reported ancestry explains little residual error once those measurements are included. If any of the three fails, the result still says something concrete about where wearable pressure sensing should and should not be used.
The population-generalizability direction enters as a lens, not as a genetics problem. The lesson we borrow comes from Precision diagnosis for monogenic diabetes requires ethnicity specific criteria for genetic testing. That study showed a European-derived BMI cutoff misdiagnosed non-White patients and that group-specific cutoffs doubled detection from 8.8% to 16%. It also acknowledged that self-reported categories are a coarse proxy. We take the motivating half, that one calibration population silently fails others, and we treat the proxy limitation as a design constraint. We therefore model continuous morphometry and test ancestry only as a residual explanatory variable.
Prior Work and What It Leaves Open
Two sparse-sensing lines supply the inputs. Intelligent Knee Sleeves (with Zhang et al.) fuses 14 yarn pressure channels and 9 IMU channels in a stacked long short-term memory network (LSTM, a recurrent sequence model). The network reports a mean angular error of 7.21 degrees across eight lower-body joints, including uninstrumented hips and ankles. It estimates the current frame only. It also flags soft-tissue artifact and starting-posture calibration as weaknesses, and both of these are plausibly morphology-dependent. Digital wearable insole-based identification of knee arthropathies shows that insole vertical ground reaction force curves carry person-specific "gait fingerprints". Its stride embeddings are much closer within a subject than between subjects. This motivates our concern, because a model that can identify people from insoles can also exploit identity shortcuts. Both works are reactive or offline, and we add a forecast horizon to them.
The predictive-estimation line supplies the target and the reason for it. Data Efficient Reinforcement Learning for Legged Robots forecasts the robot state at the moment a planner finishes, and Event-Triggered Stabilization of Nonlinear Systems with Time-Varying Sensing and Actuation Delay proves stability for predictor feedback under known delays. Neither involves human wearables. Forward dynamic models in human motor control is the biological precedent: the nervous system combines a forward model with delayed sensing. QuaMo forecasts acceleration from recent poses on the quaternion sphere, but it uses video and is blind to contact. The hip exoskeleton of Adaptive hip exoskeleton control using heart rate feedback times its torque reactively from an insole "Foot Load Ratio". Its authors call for predictive modelling, and the system was tested on 12 participants. None of these works forecasts human motion from pressure, and none tests transfer across body types.
The physics-based alternative of Comparing sparse inertial sensor setups for sagittal-plane walking and running reconstructions constrains our design. Its musculoskeletal optimal control takes 48 ± 3 min per simulation, which rules it out online. It was run on 10 healthy males and showed large between-subject variability in kinetics. We use it as evidence that body parameters matter, not as a runtime component.
Research Program
The forecaster takes a sliding window of sleeve pressure, insole pressure and IMU signals. It outputs unit-quaternion rotations for eight lower-body joints at horizons τ ∈ {0, 50, 100, 200} ms, together with probabilities of heel-strike and toe-off within the horizon. The τ = 0 output reproduces the existing estimation task. The backbone matches the stacked LSTM of Zhang et al. so that any gain is attributable to the targets and conditioning, not to architecture. As a second variant, we feed the network's outputs through an error-state Kalman filter in the style of Wouda et al. to test whether physical smoothing helps at long horizons.
Morphology enters in two separable ways. The first is a fixed physical normalization: insole channels are divided by body weight, and each sleeve channel is referenced to its value in a standing calibration pose and scaled by measured local limb circumference. The second is a learned conditioning layer. A small network maps a vector $m$ of measurements (thigh and calf circumference, thigh and shank length, their ratio as a crural-index proxy, height, mass, and sleeve size) to per-channel gains and offsets applied to the pressure features. Keeping the two separate lets us tell whether simple scaling is enough.
The central evaluation quantity is the morphology gap. We fit the forecast error at horizon τ against the distance of each test subject from the training cohort in standardized morphometric space:
$$ e_\tau(s) = a_\tau + b_\tau \, d(m_s, \mathcal{M}_{\text{train}}) + \varepsilon_s , $$
Here $e_\tau(s)$ is subject $s$'s mean angular error at horizon τ, $d$ is the Mahalanobis distance of that subject's measurement vector $m_s$ to the training subjects' measurements $\mathcal{M}_{\text{train}}$, and $b_\tau$ is the slope we call the morphology gap. Claim two predicts larger $b_\tau$ for pressure-using models than for IMU-only models. Claim three predicts that conditioning on $m$ drives $b_\tau$ toward zero. Adding self-reported ancestry as a covariate tests whether anything beyond measured morphology is left.
Data come in two tiers. The existing tier is the Intelligent Knee Sleeves dataset, with over 140,000 synchronized frames, optical motion capture ground truth, multiple subjects and multiple sleeve sizes. On it we establish forecasting under leave-one-subject-out splits, following the identity-leakage warnings of EMGBench. Its small, homogeneous cohort cannot answer the population question. The new tier is a collection of roughly 40 participants with the same sleeve hardware, digital insoles and optical capture, performing squats, curls, leg raises and treadmill walking and running. Recruitment is stratified to span thigh circumference, BMI and segment ratio, and self-reported ancestry is recorded with consent. Splits for this tier are leave-morphology-cluster-out: we cluster participants in measurement space and hold out whole clusters, including the extremes. GaitRec force-plate vertical force curves (N = 836) serve as a large auxiliary set for pretraining the insole branch on forecasting within the stride. Whether GaitRec carries body-mass metadata is not stated in our sources, so its use for the morphology analysis depends on checking that.
Baselines
The naive reference point is a zero-order hold, which predicts that the pose at t + τ equals the current estimate. A constant-angular-velocity extrapolation on the quaternion sphere is the next step up. Any learned forecaster that fails to beat these at 100 ms has no case.
The strongest existing method is the stacked LSTM of Intelligent Knee Sleeves, which reports 7.21 degrees mean joint error at the current frame. We found no reported result on Lacuna for this or any model forecasting future joint rotations on this dataset. We will therefore retrain that exact architecture with shifted targets, a configuration we call LSTM-shift. It gets the same windows, the same splits, and the same hyperparameter search budget, about 30 configurations per model, as our variants. It is the strongest competitor because it already fuses pressure and IMU in real time. If shifting its targets is all that is needed, our contribution reduces to the evaluation. We will also run a QuaMo-style acceleration-enhanced extrapolator on top of the τ = 0 LSTM output, following QuaMo, to represent the kinematic-forecasting school.
Three ablations isolate the key ingredients. An IMU-only forecaster with identical capacity isolates the value of pressure. A forecaster with physical normalization but no learned conditioning isolates whether learned conditioning is needed. A forecaster conditioned on ancestry category instead of measurements tests the categorical approach the diabetes study used. Beating the strongest baseline means two things on held-out morphology clusters. The first is a lower error at τ = 100 ms whose confidence interval, from a subject-level bootstrap, excludes LSTM-shift's. The second is a morphology-gap slope $b_\tau$ significantly smaller than LSTM-shift's.
Why not just…?
Why not just extrapolate the current-pose estimate forward? If pressure carries no lead information, a good current-pose estimator plus kinematic extrapolation will match a pressure-driven forecaster, and claim one fails. This is a real threat. The prior work supports only that pressure helps current-pose estimation of uninstrumented joints, not that it leads motion. The comparison of the QuaMo-style extrapolator and the IMU-only forecaster against the full model, plotted across τ, settles it. If the pressure advantage does not grow with horizon, we report that and drop claim one.
Why not just normalize pressure by body weight and limb size and train one pooled model? This may well be sufficient. Body-weight normalization of insole force is standard, and if circumference scaling also absorbs sleeve differences, learned conditioning is unnecessary. That is exactly why the normalization-only ablation exists. If it achieves the same $b_\tau$ as the conditioned model, the practical recommendation becomes "measure two circumferences and divide," and that is a useful, publishable result. We will not argue around it.
Why not just calibrate each new user with a short recording? Per-user fine-tuning would erase most morphology effects. However, it needs ground-truth motion, so in practice it can be done only in a lab with motion capture. The knee-sleeve authors already report that their posture calibration is fragile. We will test calibration budgets of 0, 10 and 60 seconds using only the standing pose and sensor data, with no optical ground truth. We will report how much of the gap each budget closes compared with conditioning on body measurements, which needs no recording at all.
Why not just build ancestry-specific models, as the diabetes work did? Group-specific BMI cutoffs worked in Precision diagnosis for monogenic diabetes, but that paper acknowledges that self-reported categories are crude. In biomechanics the causal pathway runs through measurable geometry and tissue. Category models also fragment an already small dataset. The ancestry-conditioned ablation tests this directly. If it matches the measurement-conditioned model, ancestry is acting as a proxy for morphology. If ancestry still explains residual error after measurements are included, we have found an unmeasured factor worth investigating.
Proposed Experiments and Decisive Checks
The first experiment runs on the existing knee-sleeve data with leave-one-subject-out splits. We compare the zero-order hold, constant-velocity extrapolation, the QuaMo-style extrapolator, LSTM-shift, the IMU-only forecaster and the full model at each horizon. The metrics are mean geodesic angular error per joint, the F1 score for timing of contact events within ±20 ms, and jitter. This establishes whether pressure leads motion before any population question is asked.
The second experiment runs on the new stratified cohort with leave-morphology-cluster-out splits. We estimate $b_\tau$ for every model with subject-level bootstrap intervals, then refit with ancestry as a covariate. The decisive check for claim two is whether pressure-using models show a steeper slope than IMU-only models. The check for claim three is whether conditioning on measurements reduces the slope beyond normalization alone.
A third, smaller pilot tests whether the forecast matters for control. In a bench replay of recorded gait through a hip-exoskeleton timing law modelled on the Foot Load Ratio controller of Adaptive hip exoskeleton control, we compare torque-onset error relative to the true gait event under an injected 100 ms delay. Reactive timing is tested against forecast-driven timing, separately for in-distribution and held-out body types. We do not claim any metabolic benefit, since that would require human-in-the-loop trials beyond this proposal.
The study falsifies its own premise if the full model fails to beat LSTM-shift at τ = 100 ms on held-out subjects, or if the IMU-only forecaster shows the same morphology slope as the pressure models. Either outcome is informative for anyone deciding whether to put pressure sensors in a wearable robot.
References
- Precision diagnosis for monogenic diabetes requires ethnicity specific criteria for genetic testing
- Intelligent Knee Sleeves: A Real-time Multimodal Dataset for 3D Lower Body Motion Estimation Using Smart Textile
- Digital wearable insole-based identification of knee arthropathies and gait signatures using machine learning
- Data Efficient Reinforcement Learning for Legged Robots
- Event-Triggered Stabilization of Nonlinear Systems with Time-Varying Sensing and Actuation Delay
- Forward dynamic models in human motor control: Psychophysical evidence
- QuaMo: Quaternion Motions for Vision-based 3D Human Kinematics Capture
- Adaptive hip exoskeleton control using heart rate feedback reduces oxygen cost during ecological locomotion
- Comparing sparse inertial sensor setups for sagittal-plane walking and running reconstructions
- Improving Full-Body Pose Estimation from a Small Sensor Set Using Artificial Neural Networks and a Kalman Filter
- EMGBench: Benchmarking Out-of-Distribution Generalization and Adaptation for Electromyography