Smart city planning gets new decision-support tool powered by XR and AI

Researchers have developed and tested a decision support system that uses extended reality (XR) and conversational artificial intelligence to collect citizen input for urban planning, offering cities a new way to move public participation beyond meetings, paper surveys and online platforms.

The study, titled A Decision Support System Integrating Extended Reality and Conversational AI for Participatory Urban Planning, presents XRCity, a public-space system built around Olivia, a life-sized virtual assistant displayed on outdoor interactive screens, and a planner-side backend that helps urban planners design engagement campaigns, manage conversational scripts, process transcripts and analyze citizen opinions.

The research was published in the journal Virtual Worlds.

Cities seek better ways to capture public opinion

Urban planning increasingly depends on citizen input, but the tools used to gather public views often fail to reach broad groups or produce evidence that planners can use directly. Public meetings tend to attract self-selected participants, while paper questionnaires and online platforms can exclude people with limited time, weak digital access, low motivation or lower digital literacy. This results in a long-running gap between open participation and structured decision-making.

The study argues that XR and conversational AI can help close that gap, provided they are embedded in a complete decision-support workflow. XR can make engagement visible and accessible in public places, while conversational AI can reduce reliance on forms, menus and formal consultation channels by allowing people to speak naturally. However, technology alone is not enough. A virtual assistant that attracts attention does not automatically produce planning evidence. For that, planners need a system that can define the decision context, ask targeted questions, classify responses and return usable analytics.

XRCity was designed to meet that need. The system places a virtual assistant, Olivia, on a weatherproof outdoor digital display installed in public space. Citizens can approach the screen and talk to Olivia without using their own phone, downloading an app or logging into a platform. This public installation model is central to the project because it aims to reduce barriers to participation.

The system's public-facing interface is connected to a planner-side backend. Urban planners can upload municipal or project-specific knowledge, prepare campaign scripts, define the planning issue under discussion, create questions and set predefined answer options. The backend also includes tools to test scripts before public deployment, process interaction transcripts and analyze structured responses. This turns the installation from a public engagement novelty into a decision-support system.

XRCity, as the authors claim, is different from many XR-based urban planning tools, which often focus on visualization, immersion or design collaboration. It also differs from common digital participation platforms that depend on personal devices and user initiative. It combines public-space XR, natural-language interaction and planner-controlled analytics in one participatory loop.

Olivia links open conversation to structured planning input

The XRCity system uses two connected interaction modes. In the first, Olivia can answer questions about local events, services, cultural heritage or other municipal information. This gives the interaction a public-service role and makes it less like a formal questionnaire. In the second mode, Olivia can shift into structured elicitation, introducing a planning issue and asking citizens for their view on predefined alternatives.

Citizens are allowed to respond in ordinary language, but the system maps their responses to options that planners defined in advance. That allows the conversation to remain natural while still producing comparable data. The original wording is preserved for review, and the matched response is stored as structured evidence.

The authors describe this as a decision-support loop. Planners define the issue and the alternatives, Olivia interacts with citizens, the system maps responses, and the backend returns aggregated findings for planners. This gives municipalities a way to connect spontaneous public interaction with decision-relevant analysis.

The backend is organized around five main modules. The knowledge provision module prepares the information Olivia can use in conversation. The interaction script designer helps planners turn an urban issue into a conversational campaign. The script validator lets planners test how the AI behaves before the system is deployed. The transcript processor anonymizes and structures the recorded conversation data. The opinion analytics module aggregates responses so planners can inspect patterns without manually reading every transcript.

The system's design also includes safeguards. Planning alternatives are not invented by the AI during conversation. They are set by planners before deployment. The AI operates within that decision frame, using configured knowledge resources and prompts intended to limit off-topic or unsupported responses. When a citizen's answer does not clearly match one of the available options, the system can flag the response for closer review rather than treating it as a clean classification.

The study gives examples of ambiguous responses. A person who comments on when they prefer to visit a public event may not be answering whether they are satisfied with the event itself. Another response may mix criticism and approval in a way that does not fit one option. These cases show why conversational participation still requires planner oversight and careful script design.

The system also reflects privacy limits. The analyzed data consisted of anonymized transcripts and session metadata from conversations where users gave consent at the start. The study did not retain original audio, images or biometric data. The authors state that the deployment followed GDPR and Portuguese data protection requirements, with guidance from the municipality's legal department.

Real-world test shows feasibility and limits

XRCity was tested in Torres Vedras, Portugal, during St. Peter's Fair from June 26 to July 6, 2025, followed by a final public street installation. The fair offered a high-footfall environment, while the later deployment provided a more ordinary urban setting. The system was tested across more than 40 hours of activity over 11 days.

The deployment generated more than 250 citizen interactions and more than 740 minutes of conversation. Of these, 191 sessions were considered usable for analysis. The average usable session included 6.7 messages per user and lasted 2.8 minutes. Some exchanges were brief, while others developed into multi-turn conversations.

The key planning result was that 14.7 percent of usable sessions produced at least one structured response to an urban planning question. That amounted to 28 of the 191 usable sessions. The result exceeded the project's 10 percent feasibility target. The authors caution that this target was an operational benchmark for the demonstrator, not a standard drawn from the literature. They also make clear that the percentage applies only to the usable sessions, not to every observed interaction in the field.

The result is modest compared with a formal survey, but the study notes the comparison is not direct. XRCity operated in open public space, where people approached voluntarily and often began with curiosity or information-seeking rather than a prior intention to answer planning questions. Under those conditions, the ability to convert a measurable share of spontaneous conversations into structured input is presented as evidence of operational feasibility.

The authors do not claim that XRCity outperforms traditional participation tools. The study did not include a controlled comparison with public meetings, QR-code questionnaires, formal surveys or other consultation methods. Its claim is narrower: public-space conversational XR can produce analyzable planning input under real deployment conditions.

There are several limitations as well. Outdoor noise affected speech capture during the fair, especially because of loudspeakers and music nearby. Longer conversations sometimes drifted away from the intended script. The system's opinion matching was not formally validated against independently coded human judgments. Therefore, the study documents field feasibility and structured-response yield, not classification accuracy.

The authors also note that the system cannot yet prove whether it reaches a representative public. It did not assess the demographics of all participants against the local population. It also did not measure cost per usable response or compare sustained engagement over time against conventional participation methods.

Future versions need stronger multilingual support, better state management to reduce conversational drift, more robust handling of noisy public environments and closer integration with GIS and municipal systems. The authors also call for comparative studies that test XRCity against conventional participation tools using the same questions and alternatives.

Participatory urban planning is moving toward more interactive and data-ready models. Cities need citizen views, but they also need methods that turn those views into evidence that planners can interpret. XRCity shows one possible path: a public-facing AI avatar connected to a planner-controlled decision pipeline, designed not merely to attract attention but to make participation more accessible and analytically useful.