Abstract
Separating world-relative object motion from retinal motion requires an estimate of self-motion. We examined how experimentally degrading one vestibular observation channel alters this attribution in a recurrent state-space observer. We reduced effective otolith reliability by increasing generative otolith noise and held canal and visual noise fixed. The observer jointly estimated self-motion and object motion and used otolith prediction errors to form an approximate belief over two internal reliability contexts. Across 30 paired realizations, reduced otolith reliability increased false-positive world-motion RMS by 47.61% and object-motion RMSE by 48.77%, compared with a 4.37% increase in angular self-motion RMSE. An instantaneous residual-only model captured part of the object-event error but not the object-free false-positive increase. Supplying oracle self-motion largely eliminated the latter effect. Two independent nonrecurrent baselines also produced more false-positive motion under noisier otolith input. Both had higher object-motion RMSE than the recurrent observer and lacked a mechanism for history-dependent recovery. The principal recurrent effect persisted with reliability inference fixed, adaptation removed, and the two adaptation rates replaced by one. Degraded otolith evidence can therefore shift ambiguous retinal evidence toward external-motion estimates. Recurrent estimation improves object-motion recovery and makes this attribution dependent on predictive history.