Current Research · GIST · Advisor: Prof. Pilwon Hur

Trust & State Anxiety in Physically Coupled Human–Robot Interaction

Real-time state-anxiety estimation and a Trust-POMDP for adaptive wearable-exoskeleton control. Mahta Akhyani, Pilwon Hur.

Download the ICROS 2026 paper (PDF)

Akhyani & Hur, Proc. 41st ICROS Annual Conference (ICROS 2026), Daegu, Korea, pp. 199–200, Jul. 2026. The proceedings are not yet indexed online, so this is the only accessible copy.

The problem

Trust in a wearable robot fails differently from trust in detached automation — and it fails before any trust measure can register the change.

An exoskeleton applies force directly to the wearer’s body, inside the peripersonal space where threat sensitivity is highest; timing errors as small as 2.8% of a stride are perceptually detectable. Existing HRI trust models operate at the outcome layer: they quantify the resulting trust state from behaviour or self-report, without representing the physiological cascade that produces it.

Sensorimotor prediction error → State anxiety → Higher sensory precision weighting → Amplified next mismatch

This loop closes before any trust outcome is observable. Modelling anxiety — the mechanism — rather than trust — the outcome — is therefore the prerequisite for intervening in real time.

A closed-form state-anxiety estimate

State anxiety at time t is a single scalar, computable online from signals a wearable HRI system already collects:

A(t) = A_base + Σ_i w_i · U_i · [ HRS(t − t_i) − HRS_∞ ] / [ HRS₀ − HRS_∞ ] Sum runs over stressor events up to time t; each event is discounted by cardiac-recovery progress since its onset.

A_base — the individual’s baseline trait score (STAI). A perfectly anticipated event contributes near zero.
w_i · U_i — a negativity-bias weight (2.5 for error events, 1.0 for neutral ones) times the event’s unexpectedness U_i = |E_iR_i − μ_i| / (σ_i · η_i^1/4), built from kinematic error, an impact grade, and a running error distribution.
HRS recovery term — a heart-rate-response function with exponential sympathetic decay and sigmoidal saturation; normalised, it equals 1 right after a stressor and approaches 0 as the wearer recovers.

Anxiety responds to unpredictability, not magnitude — a large but fully expected perturbation barely moves A(t).

Line plot of the state-anxiety estimate A(t) over a 10-minute session. A(t) sits at the baseline of 35, steps up slightly at a neutral perturbation at 1 minute, rises and decays after a mild error at 3 minutes staying below the threshold, then a larger error at 6 minutes pushes A(t) across the intervention threshold of 43 before it decays back. — Simulated A(t) for a 10-minute session with three events (A_base = 35, A_thresh = A_base + 8, M = 0.40, T_s = 6 min). The neutral perturbation at 1 min contributes ~0; the first error (3 min) stays sub-threshold; the second (6 min) crosses A_thresh and triggers the controller. Exponential recovery follows.

Closed-loop architecture

Kinematic data (IMU / MoCap) yield the error, unexpectedness and impact terms; the cardiac trace yields the normalised HRS recovery. The scalar A(t) is compared against a per-individual threshold A_thresh:

Below threshold — the controller holds; therapeutic load is unchanged.
At or above — increase actuation predictability (slower ramps, pre-cues, reduced timing jitter), targeting the prediction error without withdrawing assistance.
Sustained ≥ T_s/2 ≈ 3 min — escalate to assistance reduction.

Block diagram of the closed-loop anxiety-mediated control architecture: a kinematic stream and a cardiac stream feed compute and normalise blocks, which feed the anxiety model A(t); A(t) is compared against a threshold whose output adapts the exoskeleton controller; the controller's forces feed back to the kinematic stream, and an optional EDA channel feeds the anxiety model.

Closed-loop anxiety-mediated control. Kinematic and cardiac streams feed the anxiety model; A(t) drives controller adaptation; the controller’s forces return to the kinematic stream. EDA (dashed) is an optional logging channel.

Parameter grounding

Individual parameters need per-subject fitting, but their ranges are verified against published data: M ≈ 0.4 (max fractional heart-rate elevation under acute psychosocial stress, from WESAD TSST sessions) and T_s ≈ 6 min (cardiac recovery). A(t) updates at the cardiac sampling rate with roughly 3–5 s latency from a kinematic error — well below the minute-scale dynamics of trust failure.

Validity threat & next step

Because therapeutic exertion is the normal operating condition for a rehabilitation exoskeleton, exercise-induced cardiac arousal is the central threat to the estimate. Per-subject HRS normalisation and an optional EDA channel mitigate it; exertion-matched controls are needed for formal validation. The immediate next step is per-subject fitting against VAS-rated anxiety, targeting a moderate-to-strong correlation between predicted and reported values.

This model is the perception layer for a broader Trust-POMDP for human–exoskeleton interaction (GIST, 2024–), where the anxiety estimate becomes an observation the controller uses to reason about the wearer’s latent trust state.

Cite

M. Akhyani and P. Hur, “Real-Time State Anxiety Estimation for Adaptive Wearable Exoskeleton Control,” in Proc. 41st ICROS Annual Conference (ICROS 2026), Daegu, Korea, pp. 199–200, Jul. 2026.

All publications →