Reassessing Personalization after Embodiment Shift

Personalization lets a robot reuse personal knowledge. A model learned with one physical embodiment may become unreliable when the embodiment changes. After a shift, we ask whether that model remains useful. Transfer Utility Trust uses one or two labeled target interactions to assess whether the personalized expert will outperform the population expert. It mixes their predictions using this probability. We test 652 interactions involving 33 participants and four mobility devices while excluding target participants and target devices from fitting. Across three fixed seeds, expected excess predictive loss falls by 46.3% relative to blind personalization. The method retains 95.0% of the blind personalization gain over population in negative log likelihood, with no reliable change in mean prediction.