A rehabilitation robot that helps a stroke patient move has to know how hard the patient is actually trying. If it helps too much, the patient's nervous system learns to let the robot do the work. This is called motor slacking, and it quietly turns therapy into passive exercise. Current "assist-as-needed" controllers (controllers that reduce help when the patient seems capable) judge effort from tracking error, measured interaction force, or muscle-electrical signals (surface EMG). Each of these signals has the same flaw: it is many-to-one. A drop in the force the patient pushes against the robot could mean the patient stopped trying. It could equally mean their joint got looser, a spastic reflex faded, or an electrically stimulated muscle tired. Physicists call this situation parameter degeneracy: different internal causes produce identical observations. No amount of filtering fixes it, because the ambiguity is present even without noise.
Physics has a standard answer to degeneracy. Add a second measurement whose ambiguity runs in a different direction, or probe the system at a different scale where the competing causes behave differently. We propose to apply this directly inside the therapy loop. The robot would inject small, designed mechanical perturbations whose frequency, velocity, and timing are chosen to maximize information about the patient's voluntary torque, after accounting for everything that could mimic it. The approach works because the competing contributions respond on different timescales. Passive tissue answers instantly and in proportion to the displacement. Stretch reflexes answer after a short fixed delay and depend on velocity. Voluntary corrections come later and are not phase-locked to the perturbation. The output is an engagement estimate with an explicit identifiability score, and an assistance controller that withdraws help only when that estimate is trustworthy.
Core Claim
The measured interaction torque at an assisted joint can be written as
$$\tau_{\text{int}}(t) = \tau_{\text{pas}}(q, \dot q;\, \theta_p) + \tau_{\text{ref}}(\dot q(t-d);\, \theta_r) + u(t) + \epsilon(t),$$
where $q$ is the joint angle, $\theta_p$ are passive stiffness and damping parameters, $\tau_{\text{ref}}$ is a reflex torque driven by joint velocity at delay $d$ with gains $\theta_r$, $u$ is the voluntary torque we care about, and $\epsilon$ is sensor noise. Under ordinary assisted motion along a smooth trajectory, $q$ and $\dot q$ are nearly deterministic functions of time. Changes in $\theta_p$, $\theta_r$, and $u$ therefore trade off almost perfectly.
The right measure of this is the efficient Fisher information about $u$: the information left after the nuisance parameters $\theta = (\theta_p, \theta_r)$ have been estimated. It is
$$\tilde I_{u} = I_{uu} - I_{u\theta}\, I_{\theta\theta}^{-1}\, I_{\theta u},$$
where the blocks of the Fisher information matrix $I$ are indexed by the voluntary and nuisance parameters. Our claim has three parts. First, $\tilde I_u$ is close to zero under the passive observation used by existing assist-as-needed controllers. Second, a perturbation schedule chosen to maximize $\tilde I_u$ raises it by a large factor at a disruption cost the patient barely notices. Third, gating assistance reduction on the resulting estimate prevents the "crutch" behavior in which patients lean on the controller's boundary.
This is the orthogonal-complementarity logic of Cosmic Complementarity, which combined two datasets with crossing degeneracy directions to shrink Hubble-constant uncertainty from about 20% to about 3%. Here the second "dataset" is not a second sensor but a second probing regime, closer to Taming the off-shell Higgs boson, where off-shell, high-energy events separate two couplings that on-shell data cannot distinguish because the two couplings scale differently with energy. The analogue of energy here is perturbation frequency and latency.
Why This Gap Is Real
Several rehabilitation papers hit the degeneracy without naming or resolving it. Each motivates a different part of the proposal.
Some papers show the ambiguity directly. AR3n trains a Soft Actor-Critic reinforcement learning agent to modulate assistance gain from tracking-error history alone. It therefore cannot tell improved tracking due to learning from improved tracking due to a stiffer limb. Force/position-based velocity control treats all tangential pedal force as voluntary intent, so gravitational sag or extensor tone is read as participation. CARR normalizes interaction torque by a static passive-ankle model whose parameters were hand-tuned. Towards Variable Assistance for Lower Body Exoskeletons uses whole-body metabolic cost, which cannot separate pushing with the robot from fighting it.
Other papers show that adding physiology alone does not solve the problem. Voluntary control via neuromechanical modeling drives an exoskeleton from an EMG-driven Hill-type muscle model. It works in real time and smooths spastic EMG, but its per-muscle parameters are calibrated offline. Its net-moment output also collapses agonist and antagonist co-contraction into one number. Assistance level quantification fuses EMG and torque with an LSTM plus Kalman filter. Its offline mode classification exceeds 95%, but online accuracy drops to 61–73%. This pattern fits a structural ambiguity that a Gaussian filter cannot remove. Triadic Human-Robot-FES control makes the degeneracy explicit as a sum of voluntary, stimulated, and robot torque, and handles it with fixed dead bands and an open-loop fatigue model.
No paper in the material designs the robot's own motion to make voluntary effort identifiable. That is the specific contribution here. One caveat applies. Perturbation-based joint-impedance identification has a history in motor-control research outside this file, which we cannot cite here. What is new is embedding identifiability-optimal probing inside the assistance loop, quantifying degeneracy with the efficient information, and testing whether it changes slacking outcomes.
Research Program
The work has three stages.
Stage 1: simulated patient with known ground truth. We extend the OpenSim gait1018 musculoskeletal setting used by Durandau et al. and Christou et al. with a velocity-dependent, delayed stretch-reflex model of spasticity and time-varying passive stiffness. The simulated patient chooses its voluntary torque by minimizing a weighted sum of effort and tracking error, so slacking emerges on its own when assistance rises. The simulator provides the true split of $\tau_{\text{int}}$ into its terms, which is impossible to obtain in humans. We compute $\tilde I_u$ under passive observation to test the first part of the claim before any design work.
Stage 2: perturbation design and estimation. We restrict probes to small superimposed position or torque perturbations: multisine and pseudo-random sequences, amplitude-capped, and kept inside the existing safety tube. Their spectrum and timing are chosen online to maximize $\tilde I_u$ subject to a cap on added tracking deviation. We use two-sided finite differences for all Fisher derivatives. Cosmic Complementarity warns that one-sided derivatives produce falsely tight constraints, which is a directly relevant constraint on our method. The estimator is a moving-window Bayesian fit of $(u, \theta_p, \theta_r)$ with an EMG channel as an optional extra likelihood term, reporting both the estimate and $\tilde I_u$.
Stage 3: closed-loop control. An assist-as-needed controller lowers assistance only when the estimated voluntary torque is high and $\tilde I_u$ exceeds a threshold. Otherwise it holds assistance and schedules a probe.
Baselines
The naive reference is a tracking-error assist-as-needed controller that reduces gain as error falls, with slacking inferred from interaction-force magnitude. AR3n reports that threshold-style error-reduction control produces boundary-riding and poor retention, with its own advantage reported only as "p < 0.05" on its Unity3D virtual patient. No numeric attribution accuracy is reported on Lacuna for this baseline, so we measure it ourselves.
The strongest existing estimator of voluntary effort in the file is the LSTM plus Kalman filter of Assistance level quantification. It reports a perceived-torque mean absolute error of 0.3890 Nm and online mode classification of 61–73%. We reimplement it on the same EMG and torque streams, window lengths, and hyperparameter search budget as our estimator. Beating it means lower voluntary-torque error than that model achieves on our data, specifically in the conditions where passive stiffness or reflex gain changes mid-session.
The second strong competitor is the EMG-driven Hill-type model of Durandau et al.. We calibrate it with the same calibration time our method is allowed. The model-based versus proportional comparison in Model-Based Mid-Level Regulation (actuation RMSE 0.311 Nm model-based, 0.268 Nm fuzzy, 0.817 Nm proportional) sets the scale of what model-based control buys in a 1-DoF setting.
The third competitor is catch trials: briefly withdrawing all assistance and measuring what the patient does alone.
Three ablations isolate the key ingredient. Random perturbations of equal energy isolate the Fisher-optimal design. A single-frequency probe with no latency separation isolates the multi-timescale component. Removing the EMG term shows whether mechanics alone suffice.
Why not just…?
Why not just add EMG and fuse it with torque? EMG helps, but it has its own degeneracy. Co-contraction and spastic bursts produce large EMG with little net voluntary torque, and electrode drift shifts the mapping during a session. The 61–73% online accuracy in Assistance level quantification is the evidence. This is a real threat, so the no-perturbation, EMG-fusion condition is an explicit arm. If it matches our method on voluntary-torque error in the co-contraction and stiffness-drift conditions, the probing idea fails.
Why not just use a calibrated Hill-type musculoskeletal model? Its parameters are themselves non-identifiable from net moment. Tendon slack length, fiber length, and maximum force trade off against each other, as the calibration difficulties in Durandau et al. and CARR show. We compute $\tilde I_u$ for this model under its normal calibration protocol. If that information is already high, the model is sufficient and our contribution shrinks to the probe schedule.
Why not just remove assistance briefly and see what the patient does? This is the most serious simple alternative. It is cheap and directly measures unassisted capability. However, it disrupts the movement, patients can anticipate it, and a single step response still mixes reflex and voluntary torque in its first few hundred milliseconds. We compare the two methods at matched disruption, measured as added tracking deviation. Catch trials win if they give equal attribution error for equal disruption.
Why not just train a model-free RL policy on richer inputs? Richer inputs do not create information that the observation lacks. A policy fed passively observed signals faces the same flat likelihood. AR3n also needed a hand-built virtual patient for training. We give AR3n-style Soft Actor-Critic the same sensor streams, including the probe responses, to test whether learning can exploit probes as well as explicit estimation.
Proposed Experiments
The first decisive check is in simulation. We measure $\tilde I_u$ under passive observation and under optimized probing, across simulated patients who vary in spasticity and in the rate of stiffness drift. The prediction is at least an order-of-magnitude gain for sub-degree perturbations. If the gain is small, the physics analogy does not transfer and we stop.
The second check is voluntary-torque attribution error against simulator ground truth for every baseline and ablation. Scenarios include slacking, stiffness increase, reflex fade, and, in a triadic variant following Christou et al., stimulated-muscle fatigue. We also report detection accuracy (AUC) for "slacking versus passive change."
The third check is closed loop. The effort-minimizing simulated patient is run under each controller, and we measure sustained voluntary effort and boundary-riding time. The expectation is that the gated controller keeps effort up where the tracking-error controller lets it decay.
The fourth check is a healthy-participant elbow study on a 6-DoF arm with an 8-channel EMG setup, modeled on the setup in Assistance level quantification. Participants are instructed to push, relax, or co-contract, and passive stiffness is changed with an elastic attachment. The instructions and the known attachment stiffness supply approximate ground truth. We report attribution error, probe perceptibility ratings, and online slacking-detection accuracy against the 61–73% reference.
Clinical testing with patients is out of scope until these checks pass.
References
- Cosmic Complementarity: Joint Parameter Estimation from Cosmic Microwave Background Experiments and Redshift Surveys
- Taming the off-shell Higgs boson
- AR3n: A Reinforcement Learning-Based Assist-as-Needed Controller for Robotic Rehabilitation
- Force/position-based velocity control strategy for the lower limb rehabilitation robot during active training: design and validation
- Adaptive Patient-Cooperative Control of a Compliant Ankle Rehabilitation Robot (CARR) With Enhanced Training Safety
- Towards Variable Assistance for Lower Body Exoskeletons
- Voluntary control of wearable robotic exoskeletons by patients with paresis via neuromechanical modeling
- Assistance level quantification-based human-robot interaction space reshaping for rehabilitation training
- Adaptive Control for Triadic Human-Robot-FES Collaboration in Gait Rehabilitation: A Pilot Study
- Model-Based Mid-Level Regulation for Assist-As-Needed Hierarchical Control of Wearable Robots: A Computational Study of Human-Robot Adaptation