Multimodal large language models evaluate robot sense of safety in cities

Keonhee Jang and other authors released a study examining how the presence and type of robots on urban streets affect pedestrians' perceived safety. The study used 18 images of Seoul streets, distinguishing daytime versus nighttime, road class, and three conditions: no robot, delivery robot, and humanoid robot. Two multimodal large language models, Qwen3.8-27B and Gemma4-26B, were each conditioned on 500 synthetic personas and asked to choose the safer image in each pair, and the results were compared with those of 25 human respondents who completed the same task. The authors state that the results show the impact of robots on perceived safety depends on the time period in which they appear, and that persona-conditioned multimodal large language models captured only part of human judgments. This conclusion is based on the comparison using the above Seoul streetscape images and 25 human respondents, and does not state applicability to other cities or larger samples.

AI 根据报道生成 · 6 days ago更新