Robots in the City: Persona-Based Perceptions of Urban Robots

Keonhee Jang, Kanghoon Lee, Seokwoo Yun, Namwoo Kim, Yoonjin YoonarXiv 2026 · arXiv

This study, titled "Robots in the City: Persona-Based Perceptions of Urban Robots," explores a critical yet overlooked aspect of the future city: how the mere presence of autonomous agents changes our psychological relationship with public spaces. As delivery bots and humanoid patrollers move from labs to sidewalks, understanding their impact on "perceived safety" is essential for urban design and robotics alike.

TL;DR

Researchers used persona-conditioned Multimodal Large Language Models (MLLMs) and human surveys to evaluate how delivery and humanoid robots affect perceived street safety. While robots generally make streets feel less safe than empty ones, they provide a surprising "safety boost" at night, acting as perceived sources of surveillance or protection.

What Problem Does This Paper Address?

Urban perception research has long studied how street features—like lighting, greenery, or road width—affect how safe pedestrians feel. However, robots are a new variable. Does a delivery bot feel like a helpful companion or a tripping hazard? Does a humanoid guard feel like a protector or a dystopian threat?

Traditional human-subject studies are difficult to scale across diverse demographics and urban contexts. This paper investigates whether persona-conditioned MLLMs can serve as a scalable proxy for human perception, while simultaneously uncovering how robot type, time of day, and road class interact to shape our feelings of safety.

Methods

The authors employed a controlled visual comparison framework:

  1. Stimulus Generation: Using generative AI (specifically GPT-5.6 Sol), they created 18 variants of Seoul street scenes. These varied by road class (primary, secondary, local), time of day (day vs. midnight), and robot condition (no robot, delivery robot, or humanoid).
  2. Persona Conditioning: They sampled 500 synthetic personas from the Nemotron-Personas-Korea dataset, balancing for gender, age, income, and region.
  3. MLLM Evaluation: Two models—Qwen3.8-27B and Gemma4-26B—were prompted to adopt these personas and perform pairwise safety comparisons (e.g., "Which street looks safer: the one with the delivery robot or the one with no robot?").
  4. Rationale Analysis: The models provided written justifications for their choices, which were then coded into five thematic topics (e.g., collision risk, surveillance) using Claude Opus.
  5. Human Validation: A parallel survey with 25 human respondents provided a baseline to check model alignment.

Datasets

  • Mapillary: Used to source the original Seoul street view images.
  • Nemotron-Personas-Korea: A dataset of synthetic personas aligned to real-world South Korean demographic distributions.
  • Generated Stimuli: 18 controlled images (3 scenes × 2 times × 3 robot states).

Key Insights and Ideas

  • The "Night Effect": Robots are viewed much more favorably at night. In the dark, the perceived risk of a robot (like a collision) is outweighed by its perceived benefit as a source of light or "eyes on the street."
  • Functional vs. Social Risk: The models (and humans) perceive the two robot types fundamentally differently. Delivery robots are seen as physical obstacles (functional risk), whereas humanoids are seen as social agents that are either threatening or protective.
  • Age-Related Waryness: Older personas were significantly less likely to choose the robot-present condition as "safer," suggesting a demographic divide in technology acceptance that persists even in simulated models.
  • Model Divergence: While MLLMs captured the general "night effect," they struggled to replicate the specific human preference for humanoids. Humans were more willing to accept humanoids than the models predicted.

Key Results

  • Overall Preference: Both models and humans ranked the conditions in the same order: No Robot > Delivery Robot > Humanoid.
  • Contextual Shifts: As shown in Figure 1 (below), the "win rate" (the frequency a condition is judged safer) for delivery robots against no robots jumped from ~32% during the day to ~58% at night for the Qwen model.
Fig. 1: Win rate by urban contexts. Rows show the two conditions: (a) time of day and (b) road class, and columns show robot pairs.
  • Demographic Trends: Figure 2 illustrates that as persona age increases, the win rate for robots steadily declines, dropping from over 50% for those under 20 to roughly 27% for those over 70.
Win rate by persona attribute, highlighting the decline in robot acceptance with age.
  • Rationale Coding: For humanoids, "surveillance/protection" accounted for over 90% of the rationale topics when the robot was chosen as the safer option, confirming the "protector" persona.

Limitations and Future Work

The authors acknowledge several constraints:

  • Static Images: The study uses still photos; the actual motion and "body language" of a robot would likely change safety perceptions significantly.
  • Small Human Sample: With only 25 human participants, the human baseline has wide confidence intervals, making precise model-to-human alignment difficult to claim.
  • Model Bias: The MLLMs showed some sensitivity to image positioning (left vs. right), indicating that "LLM-as-a-judge" still requires careful calibration.
  • Future Directions: The researchers suggest moving toward video-based stimuli and exploring a wider variety of robot tasks (e.g., a robot actively helping a pedestrian) to see if functional context overrides physical form.

Summary updated

Abstract

Pedestrians increasingly encounter robots in urban public spaces without interacting with them, but how robots change the perceived safety of a street remains unclear. We examine how robot presence and type affect perceived safety across urban contexts. We generate 18 images of Seoul streets that vary in time of day, road class, and robot condition (no robot, a delivery robot, or a humanoid). Two multimodal large language models (MLLMs), Qwen3.8-27B and Gemma4-26B, each conditioned on 500 synthetic personas, choose the safer image in each pair, and 25 human respondents complete the same task. Both models prefer streets without robots and favor the delivery robot over the humanoid. Robots are chosen more often at night and less often by older personas, whereas road class has smaller effects. In the model rationales, the delivery robot appears as a physical obstacle, whereas the humanoid appears as either a threat or a source of protection. Like humans, both models chose robots more often at night, but chose the humanoid over the delivery robot far less often, and Gemma showed a night effect for the humanoid about three times that of humans. These results show that the effect of a robot on perceived safety depends on when it appears, and that persona-conditioned MLLMs capture only part of human judgments.
BibTeX
@misc{jang2026robots,
  author = {Keonhee Jang and Kanghoon Lee and Seokwoo Yun and Namwoo Kim and Yoonjin Yoon},
  title = {{Robots in the City: Persona-Based Perceptions of Urban Robots}},
  year = {2026},
  eprint = {2611.05954},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.05954},
}
Figure