Precision-Controlled Active Inference Accounts for Sensory Reweighting in Quiet Standing

Abstract

Human quiet standing relies on vestibular, proprioceptive, and visual information whose contributions change with sensory context. Dynamic posturography characterizes this reweighting through responses to visual-scene and support-surface perturbations, but a compact mathematical account of how sensory reliability propagates from state estimation to postural action remains incomplete. We develop a minimal quiet-standing model in which sensory reliability is encoded by channel-specific precision parameters that weight sensory prediction errors. A one-link inverted pendulum receives vestibular, proprioceptive, and visual observations, estimates posture using an active-inference variational free-energy objective over temporally embedded states, and selects ankle torque by minimizing the same free-energy form under an upright sensory goal. Changes in sensory conditions enter the model only through the relative channel precisions; body dynamics, the action optimizer, and the upright goal prior are held fixed. A fixed-point analysis yields closed-form predictions: the perturbation-induced belief bias, each channels state-update contribution, and the resulting posture shifts are set by relative channel precisions, with a stability condition on the upright goal. Closed-loop simulations confirmed these predictions: reducing an unreliable channels precision reduced perturbation-driven postural shifts by approximately 82%, as predicted, with a matching decrease in that channels state-update contribution, identifying belief updating as the mechanism of reweighting. A graded reliability-to-precision mapping monotonically controlled sensory contribution, and reweighting required relative, channel-selective precision changes rather than a global reduction. These results provide a compact mathematical account of postural sensory reweighting as relative, context-selective precision control in a closed-loop multisensory system.