New hybrid policy PerSim personalizes robot object search for variable placements

A new paper accepted to IROS 2026 tackles the question of when robots should personalize their search for household objects. The authors introduce PerSim, a rigidity-gated hybrid policy that dynamically chooses between a trait-conditioned personalization prior and a population-frequency baseline. The key insight: personalization is only beneficial for objects whose placement varies widely across individuals (low rigidity), while universally placed items are better handled by generic spatial priors.

To scale resident-conditioned dynamics, the team built a human-calibrated simulation pipeline that generates synthetic object-placement transitions across diverse home layouts. They trained a predictor to inject continuous Big Five personality vectors, outputting room-level priors and within-room co-occurrence cues. In a unified human study (N=200), synthetic transitions were rated behaviorally plausible (mean 3.85/5, p < 1e-6), and a blinded A/B comparison showed personalization favored only for low-rigidity objects (p=0.005). Offline tests showed small but significant gains (p=0.035) over nearest discrete trait matching. In a digital twin environment, PerSim reduced expected search cost end-to-end.