Reliability-weighted target-position estimation in a musculoskeletal arm model: adaptive priors and learned source weighting under violations of fixed-precision assumptions

Reliability-weighted integration is a normative account of cue combination, but its scope in musculoskeletal models and conditions requiring adaptive or learned extensions remains unclear. We studied target-position estimation in a MyoSuite arm within a field-wise architecture representing vision, proprioception, forward prediction, and a task prior. Target-position experiments combined vision, forward prediction, and, where applicable, the task prior by precision weighting; proprioception informed other state fields. At the open-loop state-estimate level, a trusted but wrong prior biased the target estimate more as visual reliability decreased, and a false visual cue was followed under low noise and discounted under high noise; both effects followed the bias ≈ w · offset pattern. Across trials, an adaptive prior tracked shifts in the target mean, while variance tracking reduced the prior weight as the target spread increased. A learned softmax integrator met the prespecified parity criterion relative to precision weighting for calibrated Gaussian synthetic observation channels and improved estimation when reported variance was miscalibrated, or channel errors were biased, heavy-tailed, or correlated. On target-position observations from two held-out MyoSuite rollout seeds, improvements under prediction miscalibration and injected visual outliers were directionally consistent across seeds during offline replay. The MSE benefit in the outlier regime reflected regime-level source weighting rather than consistent per-trial outlier detection. These results define a computational boundary: precision weighting was sufficient in the calibrated conditions tested here, whereas adaptive statistics or learned source weighting became useful when assumptions changed or failed. Evidence from the arm model is limited to target-position estimates; endpoint propagation remains unresolved.