CSIR planning framework lets robots find target people under incomplete information, beating distance-only graph baselines while semantics alone destabilizes search

A planning framework for PersonNav that combines distance and semantic information about the recipient (habits and intent) weighted by the trust of user-provided information. Existing solutions are rigid person-finding systems requiring exact knowledge of the individual, or learned methods relying on public or sparse data unavailable due to human privacy; this framework fuses distance and semantics via trust weighting and keeps an explicit belief representation. Method and design as described at the abstract level; the explicit belief representation is stated to naturally support future Bayesian filtering.

A synthetic benchmark of scenarios including actors, items, and requests, used to evaluate performance before real-world deployment and to simulate natural-language human-robot interaction. Addresses the unavailability of public or sparse data caused by human privacy by providing reproducible synthetic evaluation scenarios. Benchmark composition as described at the abstract level; no scenario counts or statistical details are given.

The informed search outperforms classical distance-based graph baselines, while semantics alone leads to ungrounded, sporadic search; an LLM-based variant performs comparably. Separates the effect of fusing semantics with distance from two controls (distance-only baseline and semantics-only), showing semantics alone is insufficient for grounded search. Comparative results reported at the abstract level; no specific metric values or effect sizes are given.

Hardware tests show the method supports real-world embodiment able to leverage between semantic and distance information in a real setting. Moves evaluation from the synthetic benchmark toward embodied deployment validation in a real setting. Hardware tests as described at the abstract level; no platform, scenario, or sample details are given.

The work targets PersonNav in indoor item-delivery settings where user-provided information varies in trust, language is uncertain, and exact knowledge of the individual is unavailable. The synthetic benchmark includes actors, items, and requests to evaluate performance before real-world deployment and to simulate natural-language human-robot interaction; hardware tests show the method supports real-world embodiment. The explicit belief representation is stated to naturally support future Bayesian filtering, so the framework can be extended by later probabilistic inference methods.

The abstract gives no specific performance values, effect sizes, scenario counts, hardware platform, or participant scale, so the magnitude of gains and statistical robustness cannot be judged. The gap between the synthetic benchmark and real environments, how language uncertainty is modeled, and how the trust of user-provided information is estimated are open questions for readers to confirm. The LLM-based variant performing comparably also raises questions about its applicable conditions and computational cost.